WO1999067728A1 - Methods and apparatus for classifying text and for building a text classifier - Google Patents

Methods and apparatus for classifying text and for building a text classifier Download PDF

Info

Publication number
WO1999067728A1
WO1999067728A1 PCT/US1999/014086 US9914086W WO9967728A1 WO 1999067728 A1 WO1999067728 A1 WO 1999067728A1 US 9914086 W US9914086 W US 9914086W WO 9967728 A1 WO9967728 A1 WO 9967728A1
Authority
WO
WIPO (PCT)
Prior art keywords
textual information
category
training set
information objects
monotonic function
Prior art date
Application number
PCT/US1999/014086
Other languages
French (fr)
Inventor
Susan T. Dumais
David Heckerman
Eric Horvitz
John C. Platt
Mehran Sahami
Original Assignee
Microsoft Corporation
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Microsoft Corporation filed Critical Microsoft Corporation
Priority to EP99933534.2A priority Critical patent/EP1090365B1/en
Priority to JP2000556321A priority patent/JP4514082B2/en
Publication of WO1999067728A1 publication Critical patent/WO1999067728A1/en

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/35Clustering; Classification
    • G06F16/353Clustering; Classification into predefined classes
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/24Classification techniques
    • G06F18/241Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • YGENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y10TECHNICAL SUBJECTS COVERED BY FORMER USPC
    • Y10STECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y10S707/00Data processing: database and file management or data structures
    • Y10S707/99931Database or file accessing
    • Y10S707/99933Query processing, i.e. searching
    • Y10S707/99936Pattern matching access

Definitions

  • the present invention concerns determining whether an object, such as a textual information object for example, belongs to a particular category or categories. This determination is made by a classifier, such as a text classifier for example.
  • a classifier such as a text classifier for example.
  • the present invention also concerns building a (text) classifier by determining appropriate parameters for the (text) classifier.
  • computers are called upon to classify (or recognize) objects to an ever increasing extent.
  • computers may use optical character recognition to classify handwritten or scanned numbers and letters, pattern recognition to classify an image, such as a face, a fingerprint, a fighter plane, etc., or speech recognition to classify a sound, a voice, etc.
  • Text classification may be used to organize textual information objects into a hierarchy of predetermined classes or categories for example. In this way, finding (or navigating to) textual information objects related to a particular subject matter is simplified. Text classification may be used to route appropriate textual information objects to appropriate people or locations. In this way, an information service can route textual information objects covering diverse subject matters (e.g., business, sports, the stock market, football, a particular company, a particular football team) to people having diverse interests.
  • diverse subject matters e.g., business, sports, the stock market, football, a particular company, a particular football team
  • Text classification may be used to filter textual information objects so that a person is not annoyed by unwanted textual content (such as unwanted and unsolicited e-mail, also referred to as junk e-mail, or "spam") .
  • unwanted textual content such as unwanted and unsolicited e-mail, also referred to as junk e-mail, or "spam”
  • spam unwanted e-mail
  • rule-based system may be used to effect such types of classification. Basically, rule-based systems use production rules of the form:
  • the conditions may include whether the textual information includes certain words or phrases, has a certain syntax, or has certain attributes. For example, if the textual content has the word “close”, the phrase “nasdaq” and a number, then it is classified as "stock market” text.
  • rule-based systems become unwieldy, particularly in instances where the number of measured or input values (or features or characteristics) becomes large, logic for combining conditions or rules becomes complex, and/or the number of possible classes becomes large. Since textual information may have many features and complex semantics, these limitations of rule-based systems make them inappropriate for classifying text in all but the simplest applications.
  • classifiers Over the last decade or so, other types of classifiers have been used increasingly. Although these classifiers do not use static, predefined logic, as do rule-based classifiers, they have outperformed rule-based classifiers in many applications. Such classifiers are discussed in ⁇ 1.2.2.2 below and typically include a learning element and a performance element. Such classifiers may include neural networks, Bayesian networks, and support vector machines. Although each of these classifiers is known, each is briefly introduced below for the reader's convenience.
  • classifiers having learning and performance elements outperform rule-based classifiers, in many applications.
  • these classifiers may include neural networks (introduced in ⁇ 1.2.2.2.1 below for the reader's convenience), Bayesian networks (introduced in ⁇ 1.2.2.2.2 below for the reader's convenience) , and support vector machines (introduced in ⁇ 1.2.2.2.3 below for the reader's convenience).
  • a neural network is basically a multilayered, hierarchical arrangement of identical processing elements, also referred to as neurons.
  • Each neuron can have one or more inputs but only one output.
  • Each neuron input is weighted by a coefficient.
  • the output of a neuron is typically a function of the sum of its weighted inputs and a bias value.
  • This function also referred to as an activation function, is typically a sigmoid function. That is, the activation function may be S-shaped, monotonically increasing and asymptotically approaching fixed values (e.g., +1, 0, -1) as its input (s) respectively approaches positive or negative infinity.
  • the sigmoid function and the individual neural weight and bias values determine the response or "excitability" of the neuron to input signals .
  • the output of a neuron in one layer may be distributed as an input to one or more neurons in a next layer.
  • a typical neural network may include an input layer and two (2) distinct layers; namely, an input layer, an intermediate neuron layer, and an output neuron layer. Note that the nodes of the input layer are not neurons. Rather, the nodes of the input layer have only one input and basically provide the input, unprocessed, to the inputs of the next layer.
  • the input layer could have 300 neurons (i.e., one for each pixel of the input) and the output array could have 10 neurons (i.e., one for each of the ten digits).
  • the use of neural networks generally involves two (2) successive steps. First, the neural network is initialized and trained on known inputs having known output values (or classifications) . Once the neural network is trained, it can then be used to classify unknown inputs. The neural network may be initialized by setting the weights and biases of the neurons to random values, typically generated from a Gaussian distribution. The neural network is then trained using a succession of inputs having known outputs (or classes) .
  • the values of the neural weights and biases are adjusted (e.g., in accordance with the known back-propagation technique) such that the output of the neural network of each individual training pattern approaches or matches the known output.
  • a gradient descent in weight space is used to minimize the output error. In this way, learning using successive training inputs converges towards a locally optimal solution for the weights and biases. That is, the weights and biases are adjusted to minimize an error.
  • the system would be "over trained” such that it would be too specialized to the training data and might not be good at classifying inputs which differ, in some way, from those in the training set.
  • the system is tested on a set of validation data. Training is halted when the system's performance on the validation set no longer improves.
  • the neural network can be used to classify unknown inputs in accordance with the weights and biases determined during training. If the neural network can classify the unknown input with confidence, one of the outputs of the neurons in- the output layer will be much higher than the others.
  • the algorithm uses small steps. Furthermore, the back propagation (gradient descent) technique used to train neural networks is relatively slow. (See, e.g., the article: Sch ⁇ tze, et al., "A Comparison of Classifiers and Document Representations for the Routing Problem", International ACM SIGIR Conference on Research and Development in Information Retrieval, Section 5 (1995) (Hereafter referred to as "the Sch ⁇ tze article”). Consequently, convergence is slow. Also, the number of neurons in the hidden layer cannot easily be determined a priori . Consequently, multiple time-consuming experiments are often run to determine the optimal number of hidden neurons.
  • Bayesian networks use hypotheses as intermediaries between data (e.g., input feature vectors) and predictions (e.g., classifications).
  • the probability of each hypothesis, given the data (“P (hypo I data) ") , may be estimated.
  • a prediction is made from the hypotheses using posterior probabilities of the hypotheses to weight the individual predictions of each of the hypotheses.
  • the probability of a prediction X given data D may be expressed as:
  • H x is the i th hypothesis.
  • D)) is referred to as a maximum a posterior hypothesis (or and may be expressed as follows :
  • the probability of the data D remains fixed. Therefore, to find H MAP , the numerator must be maximized.
  • the first term of the numerator represents the probability that the data would have been observed given the hypothesis i.
  • the second term represents the prior probability assigned to the given hypothesis i.
  • a Bayesian network includes variables and directed edges between the variables, thereby defining a directed acyclic graph (or "DAG"). Each variable can assume any of a finite number of mutually exclusive states. For each variable A, having parent variables Bi, . . . B n , there is an attached probability table (P(A
  • the structure of the Bayesian network encodes the assumptions that each variable is conditionally independent of its non-descendants, given its parent variables.
  • a variable “MML” may represent a "moisture of my lawn” and may have states “wet” and “dry”.
  • the MML variable may have "rain” and “my sprinkler on” parent variables each having "Yes” and “No” states.
  • Another variable, “MNL” may represent a "moisture of my neighbor's lawn” and may have states “wet” and “dry”.
  • the MNL variable may share the "rain” parent variable.
  • a prediction may be whether my lawn is "wet” or “dry”. This prediction may depend of the hypotheses (i) if it rains, my lawn will be wet with probability (xj .
  • conditional probability tables in Bayesian networks may be trained, as was the case with neural networks.
  • the learning process may be shortened.
  • prior probabilities for the conditional probabilities are usually unknown, in which case a uniform prior is used.
  • Support vector machines are another type of trainable classifier. SVMs are reportedly more accurate at classification than naive Bayesian networks in certain applications, such as text classification for example. (See, e.g., the article, Joachims, "Text Categorization with Support Vector Machines: Learning with Many Relevant Features", LS-8 , Report 23, University of Dortmund Computer Science Department (Nov. 1997).) They are also reportedly more accurate than neural networks in certain applications, such as reading handwritten characters for example. (See, e.g., the article, LeCun et al .
  • An object to be classified may be represented by a number of features. If, for example, the object to be classified is represented by two (2) features, it may be represented by a point in two (2) dimensional space. Similarly, if the object to be classified is represented by n features, also referred to as a "feature vector", it may be represented by a point in n-dimensional space.
  • the simplest form of an SVM defines a plane in the n-dimensional space (also referred to as a hyperplane) which separates feature vector points associated with objects "in a class" and feature vector points associated with objects "not in the class".
  • a number of classes can be defined by defining a number of hyperplanes.
  • the hyperplane defined by a trained SVM maximizes a distance (also referred to as an Euclidean distance) from it to the closest points (also referred to as "support vectors") "in the class” and "not in the class".
  • a hyperplane is sought which maximizes the distances between the support vectors and the hyperplane, so that the SVM defined by the hyperplane is robust to input noise.
  • the hyperplane (or hypersurface) is defined by a training process, some of which are discussed in ⁇ 4.2.1.4.1 below.
  • rule-based classifiers are feasible in only the simplest text classification applications, some significant challenges exist when using systems having learning and performance ELEMENTS (ALSO REFERRED TO AS "LEARNING MACHINES") FOR text classification. Some of the more significant challenges will be introduced in ⁇ 1.2.2.2.4.1 through 1.2.2.2.4.3 below.
  • Each learning example includes a vector of features associated with a textual information object.
  • feature vectors may have on the order of 10 8 features.
  • a large number of features can easily be generated by considering the presence or absence of a word in a document to be a feature. If all of the words in a corpus are considered as possible features, then there can be millions of unique features. For example, web pages have many unique strings and can generate millions of features. An even larger number of features are possible if pairs or more general combinations of words or phrases are considered, or if the frequency of occurrence of words is considered.
  • the number of features in a feature vector may be reduced by so-called “feature reduction” or “feature selection” methods such that a reduced feature vector, having a subset of the features of the original feature vector, is produced. Indeed, some believe that for learning machine text classifiers to be feasible, feature selection is needed.
  • DF-Thresholding See, e.g., the paper: Yang and Peterson, "A Comparative Study on Feature Selection in Text Categorization, " International Conference on Machine Learning (1997); hereafter referred to as: the Yang-Peterson article"
  • Chi-Squared Test See, e.g., the Sch ⁇ tze article
  • Term Strength Criterion See, e.g., the article: Yang and Wilbur, "Using Corpus Statistics to Remove Redundant Words in Text Categorization, " Journal of the American Society for Information Science, Vol. 47, No. 5, pp.
  • the Yang-Wilbur article Information Gain Criteria (See, e.g., the Yang-Peterson article) , correlation coefficient which is the square root of the Chi-Squared measure (See, e.g., the article: Ng, et al, "Feature Selection,
  • the Joachims article which espouses the use of SVMs for text classification, concludes that feature selection is likely to hurt performance due to a loss of information. (See, e.g., the Joachims article, page 3.)
  • the Joachims article further concludes that since SVMs can generalize well in high dimensional feature spaces, they eliminate the need for feature selection thereby making text categorization easier. (See, e.g., the Joachims article, page 11.)
  • feature selection has been avoided in some instances.
  • a learning machine When a learning machine is trained, it is trained based on training examples from a set of feature vectors. In general, performance of a learning machine will depend, to some extent, on the number of training examples used to train it. Even if there are a large number of training examples, there may be a relatively low number of training examples which belong to certain categories.
  • Many learning machines utilize sigmoid functions (that is, s-shaped, monotonic functions) to determine an output (e.g., a determination as to whether an object is in a category or not) based on an input (e.g., a feature vector of an unknown object).
  • sigmoid functions that is, s-shaped, monotonic functions
  • Training time and response (e.g., classification) time are two (2) important characteristics of a learning machine. For example, even those that espouse the use of learning machines, and SVMs in particular, for classifying text concede that the training time of SVMs is longer than other methods.
  • This challenge is addressed in the U.S. Patent Application Serial No. 09/055,477, by John Platt, entitled “Methods and Apparatus for Building a Support Vector Classified", filed on April 6, 1998 and assigned to a common assignee (incorporated by reference) and will not be addressed in detail here.
  • response (or classification) time There remains the issue of response (or classification) time.
  • the present inventors believe that the SVM classifier discussed in the Joachims paper classifies unknown objects in accordance with the following expression:
  • 0 the classification output
  • nte the number of training examples
  • Gi Lagrange multiplier training example i
  • X x feature vector of training example i
  • - j feature vector of unknown object j
  • y x known output of training example i.
  • the Schutze article discusses logistic regression classifier which employs a weight vector ( ⁇ ) derived using a maximum likelihood and the Newton-Raphson method of numerical optimization.
  • the present invention may perform at least one (1) of two (2) basic functions, namely generating parameters for a classifier, and classifying objects, such as textural information objects.
  • parameters are generated for a classifier based on a set of training examples.
  • a set of feature vectors may be generated from a set of training examples.
  • the features of the set of feature vectors may be reduced.
  • the parameters to be generated may include a defined monotonic (e.g., sigmoid) function and a weight vector.
  • the weight vector may be determined by means of SVM training (or by another, known, technique).
  • the monotonic (e.g., sigmoid) function may be defined by means of an optimization method.
  • the text classifier may include a weight vector and a defined monotonic (e.g., sigmoid) function.
  • sigmoid a defined monotonic
  • O c a classification output for category c
  • c a weight vector parameter associated with category c
  • x is a (reduced) feature vector based on the unknown textual information object
  • a and B are adjustable parameters of a monotonic (e.g., sigmoid) function.
  • the calculation of the output from expression (2) is quicker than the calculation of the output from expression (1) .
  • the classifier may (i) convert a textual information object to a feature vector, and (ii) reduce the feature vector to a reduced feature vector having less elements.
  • Figure 1 is a high level diagram of processes that may be performed by a first aspect of the present invention.
  • Figure 2 is a high level diagram of processes that may be performed by a second aspect of the present invention.
  • Figure 3 is an exemplary data structure of a training example of a textual information object.
  • Figures 3B and 3C are alternative exemplary sparse vector data structures of the textual information object .
  • Figure 4 is an exemplary data structure of the training example of Figure 3, in which a feature vector has been reduced.
  • Figures 4B and 4C are alternative exemplary sparse vector data structures of the reduced feature vector.
  • Figure 5 is an exemplary data structure of the training example of Figure 4, in which values of a reduced feature vector have been binarized.
  • Figures 5B and 5C are alternative exemplary sparse vector data structures of the binarized reduced feature vector.
  • Figure 6 is a flow diagram of an exemplary feature extraction process.
  • Figure 7 is a flow diagram of an exemplary feature reduction process.
  • Figure 8 is a flow diagram of an exemplary processes for generating parameters for a classification engine.
  • Figure 9 depicts information used by some feature reduction techniques.
  • Figure 10 depicts various sigmoid functions.
  • Figure 11A is a block diagram of a computer environment which may be used to effect processes carried out by the present invention.
  • Figure 11B is a block diagram of a machine which may be used to effect processes carried out by the present invention.
  • the present invention concerns a novel classifier, such as a text classifier, and building such a (text) classifier.
  • a novel classifier such as a text classifier
  • building such a (text) classifier building such a (text) classifier.
  • a vector may be depicted with an arrow "hat” or by bold typeface .
  • the present invention may perform at least one (1) of two (2) basic functions.
  • the first basic function namely generating parameters for a classifier
  • the second basic function namely classifying text, is discussed in ⁇ 4.1.2 below.
  • parameters are generated for a classifier based on a set of training examples.
  • the present invention may function to generate a set of feature vectors from the set of training examples.
  • the present invention may also function to reduce the number of features of feature vectors of the set.
  • the parameters to be generated include a defined monotonic (e.g., sigmoid) function and a weight vector.
  • the weight vector may be determined by means of SVM training.
  • the monotonic (e.g., sigmoid) function may be defined by means of an optimization method. The optimization method may use a special target function to avoid overfitting.
  • the present invention may classify text.
  • a text classifier may include a weight vector and a defined monotonic (e.g., sigmoid) function. Basically, the output of the text classifier may be expressed as:
  • 0 C a classification output for category c
  • w c a weight vector parameter associated with category c
  • x is a (reduced) feature vector based on the unknown textual information object
  • a and B are adjustable parameters of a monotonic (e.g., sigmoid) function.
  • the classifier may (i) convert a textual information object to a feature vector, and (ii) reduce the feature vector to a reduced feature vector having less elements.
  • Exemplary processes and methods for generating parameters of a classifier are described in ⁇ 4.2.1 below.
  • Exemplary processes and methods for classifying text are described in ⁇ 4.2.2 below.
  • exemplary apparatus (or architecture) for effecting the described processes and methods are described in ⁇ 4.2.3 below.
  • Figure 1 is a diagram which depicts processes and data which may be used by the present invention to generate parameters of a text classifier.
  • a feature extraction process (or more generally, a feature extractor) 110 generates a feature vector corresponding to each of the textual information objects of the training set.
  • the feature vectors may be stored at storage location 125.
  • An exemplary feature extraction process is discussed in ⁇ 4.2.1.1 below.
  • Figure 3 illustrates an exemplary data structure 300 of a training object which includes a feature vector 310 and a category vector 320.
  • the feature vector 310 includes fields 312 corresponding to features (or attributes) which were derived from a textual information object. In this example, there are "n" features in the feature vector 310.
  • the feature (or attribute) fields 312 of the feature vector 310 include corresponding values derived from the textual information object. For example, if it is assumed that the textual object may include words from a 2.5 million word vocabulary, then the feature vector 310 may include 2.5 million fields, each corresponding to a word of the vocabulary. The value stored in a field may correspond to the number of times a particular word of the vocabulary appears in the textual information object.
  • a sparse vector technique may be used such that the feature vector is represented by two (2) smaller arrays rather than one large array.
  • a first array indexes features having non-zero values.
  • a second array includes the values of the features indexed by the first array. For example, referring to Figure 3B, if only the first, 50 th , 57 th , 101 st , and 220 th features of a feature vector have non-zero values, a sparse data structure 310' may be used instead of the feature vector 310 of Figure 3.
  • the sparse data structure 310' includes a first array 350 which includes the indices (e.g., 1, 50, 57, 101, and 220) to the non-zero features, and a second array 360 which includes the values of the indexed features.
  • Figure 3C depicts another sparse data structure 310" which has alternating indexing fields and value fields corresponding to the non-zero features.
  • the category vector 320 includes fields 322 corresponding to "m" different categories.
  • the fields 322 of the category vector 310 may include a binary value which indicates whether or not the textual information object belongs to a particular category.
  • a textual information object may belong to more than one category.
  • the data structures 300 corresponding to each textual information object of the training set may be stored at storage area 125.
  • a feature reduction process (or more generally, a feature reduction facility) 120 accepts the data structures 300, either from the storage area 125 or the feature extraction process 110, and reduces the number of features of the feature vector 310 from "n" to "z", where "z" is less than "n", and may be as much as three (3) orders of magnitude less than "n” .
  • a new data structure 400 is generated by the feature reduction process 120.
  • the data structure 400 includes a category vector 320 which corresponds to that of the data structure 300.
  • the data structure 400 also includes a reduced feature vector 410 having fields 412 corresponding to features (or attributes) of the textual information objects of the training set that are determined to be relatively important features by the feature reduction process 120.
  • the feature reduction process 120 also produces a reduced feature set which defines which "z" of the "n" fields 412 of the feature vector are considered to be relatively important.
  • An exemplary feature reduction process 120 is discussed in ⁇ 4.2.1.2 below. Although each textual information object was described as being a single, relatively large, feature vector array, in practice, sparse array techniques are used such that each textual information object may be represented by two (2) smaller arrays, rather than one (1) large array, as discussed above.
  • the sparse data structure 410' includes an first array 450 which includes the indices (e.g., 50, 101, and 220) of non-zero, selected, features, and a second array 460 which includes the values of the features indexed by the first array 450.
  • Figure 4D depicts another sparse data structure 410" which has alternating indexing fields and value fields corresponding to the non-zero, selected, features.
  • An optional (depicted in dashed lines) binarizing process 150 may be used to convert each value of each field 412 of the reduced feature vector 412 of each of the objects of the training set, to a binary value.
  • This conversion of values to binary values may take place by applying the values to a predetermined (or derived) threshold, in the case of each field 412. Thus, a different threshold may be associated with each field 412. Naturally, a single threshold could be applied to all fields 412.
  • Figure 5 illustrates a data structure which may be used to represent the resulting reduced feature vector containing binary values 51C.
  • the vector 510 has the same number of fields "z" as the vector 410 of the data structure 400 of Figure 4.
  • each of the fields 512 include a binary value (e.g., represented by one (1) bit) rather than a more general value.
  • the category vector 320 of the data structure 500 is the same as that of the data structures 300 and 400. An exemplary binarizing process is discussed in ⁇ 4.2.1.3 below.
  • a sparse data structure 510' includes a first array 550 which includes the indices (e.g., 101 and 220) of the selected , non-zero, threshold passing features, and a second array 560 which includes the values of the indexed features.
  • a sparse data structure 510 including only the index array 550 may be used.
  • the reduced feature vector may be stored at storage unit 135.
  • a general text classifier will include classifiers for each of the "m” different categories. Recall that a weight vector parameter and a monotonic (e.g., sigmoid) function, having adjustable parameters A and B, are to be determined for each of the "m” classifiers. Accordingly, "m" processes 140 (which may include a support vector machine training facility and a function generator) accept the reduced feature vector (or binarized reduced feature vector) from storage unit 135 or from the feature reduction process 120 (or binarizing process 150).
  • m parallel processes 140 are depicted in Figure 1, a single process 140 could be used to process "m" successive categories (or p processes could be run m/p times) .
  • a weight vector parameter and a monotonic (e.g., sigmoid) function are generated for each of the "m" categories.
  • An exemplary process for generating parameters of a text classifier is described in ⁇ 4.2.1.4 below.
  • a feature extraction process 110 generates a set of feature vectors (e.g., stored in data structures 300) from a set of training textual information objects.
  • FIG. 6 is a flow diagram of an exemplary feature extraction process 110'.
  • a COUNT value to be used as an index to textual information objects, is initialized by setting it to 1.
  • a textual information object, as indexed by the COUNT value is accepted as shown in step 620.
  • a FEATURE_COUNT value is initialized to 1 and a VALUE value, as indexed by the FEATURE_COUNT value, is initialized to 0.
  • the VALUE value, as indexed by the FEATURE_COUNT value, of the textual information object, as indexed by the COUNT value is determined.
  • the VALUE (5) of OBJECT (1) is three (3).
  • the values for each feature of a given object are determined.
  • steps 670 and 680 once the values of the features of one object have been determined, the values of the features of the next object are then determined.
  • the process 110' terminates via return node 690.
  • step 640 determines a value of a feature of a textual information object. This determination may be performed as follows.
  • a text analyzer may be used to parse each textual information object into constituent tokens.
  • a token is any textual component, such as a word, letter, internal punctuation mark or the like, that is separated from another such component by a blank (white) space or leading (following) punctuation marks. Syntactic phrases and normalized representations for times and dates are also extracted by the text analysis module. Domain-specific features may also be defined and extracted as appropriate. Each of the features may define a word, phrase or letter grouping. Thus, a feature vector is associated with each of the textual information objects .
  • step 640 if VALUE (FEATURE_COUNT) is non-zero, then (i) the next element of the index array (or next index field) is set to FEATURE_COUNT, and (ii) the next element of the value array (or next value field) is set to VALUE (FEATURE COUNT) .
  • FIG. 7 is a high level flow diagram of an exemplary feature reduction process 120' ("feature reduction” may also be referred to as "feature selection") which may be used in the present invention.
  • feature reduction may also be referred to as "feature selection” which may be used in the present invention.
  • the present invention may use a category independent feature reduction step (such as ZIPF or frequency) (known but discussed in ⁇ 4.2.1.2.1 below for the reader's convenience).
  • a category dependent feature reduction technique (some of which are discussed in ⁇ 4.2.1.2.2 below) may be performed.
  • the process 120' then terminates via return node 730.
  • a matrix of feature vectors of a number of training example textual information objects may be reduced by applying collection-wide frequency statistics.
  • category dependent feature reduction may use known techniques such as the mutual information technique discussed in ⁇ 4.2.1.2.2.1 below, or the information gain technique discussed in ⁇ 4.2.1.2.2.2 below.
  • Other known category dependent feature reduction techniques such as chi-square, correlation coefficient, and SVD (all known and not discussed here) may be used instead.
  • Category dependent feature reduction may also be carried out in accordance with a proprietary technique such as the Bayesian score technique discussed in ⁇ 4.2.1.2.2.3 below. ⁇ 4.2.1.2.2.1 MUTUAL INFORMATION
  • Mutual information is a measure of the extent to which a feature is associated with a class, in this instance to a text category. Features that exhibit high mutual information measures are good discriminators of the class, and are therefore chosen.
  • the features may be words and the classes may be categories.
  • the mutual information "MI" for a given feature (fj . ) such as a word for example, in a given category c ⁇ can be expressed as:
  • P(f + c + ) is the probability that a document has feature f and belongs to category c;
  • P(f " c + ) is the probability that a document does not have feature f, but belongs to category c
  • P(f + c ⁇ ) is the probability that a document has feature f but does not belong to category c
  • P(f ⁇ c " ) is the probability that a document does not have feature f and does not belong to category c
  • Figure 9 depicts a two-way contingency table 900 of a feature "f" 950 and a category "c” 960.
  • Each textual information object exhibits one (1) of the following four (4) characteristics: (i) has the feature "f” 950 and belongs to the category "c” 960 (f + c + 910); (ii) has the feature "f” 950 but does not belong to the category "c” 960 (f + c ⁇ 920); (iii) does not have the feature "f” 950 but belongs to the category "c” 960 (f “ c + 930); or (iv) does not have the feature "f” 950 and does not belong to the category "c” 960 (f ⁇ c ⁇ 940) .
  • MI mutual information
  • mutual information is computed for each unique term. Those terms whose mutual information value is less than some predetermined threshold may be removed from the feature vector. Alternatively, if the number of features is to be reduced to a predetermined number "x", then the features with the x highest mutual information values are kept while other features are not considered.
  • Information gain is a term selection criterion which measures the number of bits of information obtained for category prediction by knowing the presence or absence of a term in a document.
  • a binary classification task i.e., in category (c + ) vs. not in category (c ⁇ )
  • binary feature values i.e., feature present (f + ) vs. feature absent (f ⁇ )
  • the information gain (G(f)) of a feature (f) is defined as:
  • P(c + ) is the probability that a document belongs to category c
  • P(c ⁇ ) is the probability that a document does not belong to category c
  • P(f + ) is the probability that a document contains feature f
  • P(f " ) is the probability that a document does not contain feature f
  • Pff ⁇ c 1 is the probability that a document has feature f and belongs to category c 1 -- for binary classification i takes on values + and - ;
  • P(f ⁇ c 1 ) is the probability that a document does not have feature f and belongs to category c 1 — for binary classification i takes on values + and -;
  • This measure can be generalized to m-ary classification by summing over m possible categories.
  • information gain is computed for each unique feature. Those features whose information gain value is less than some predetermined threshold may be removed from the feature vector. Alternatively, if the number of features is to be reduced to a predetermined number "x", then the features with the x highest information gain values are kept while other features are not considered.
  • the feature vector may be reduced based on a Bayesian Score (BS) .
  • BS Bayesian Score
  • the determination of a Bayesian Score is described in patent application serial no. 08/240,019, entitled “Method for Improving Belief Networks” (incorporated herein by reference) .
  • A is a number of textual information objects exhibiting characteristic f + c + 910
  • B is a number of textual information objects exhibiting characteristic f + c ⁇ 920
  • C is a number of textual information objects exhibiting characteristic f " c + 930
  • D is a number of textual information objects - xhibiting characteristic f ⁇ c ⁇ 940
  • N is a total number of textual information objects
  • Bayesian Score threshold may also be used to reduce the number of features .
  • the values held in a reduced feature vector may be applied to one threshold, or to thresholds corresponding to the features, to binarize the values of the reduced feature vector.
  • the values in the feature vector may be applied directly to the classifier, or may be encoded as q-ary (where q>2) code through quantization, as is well known in the neural network art.
  • the foregoing steps may be thought of as pre-processing steps to permit parameters of a text classifier, such as an SVM text classifier, to be more easily determined.
  • the classifier will classify a textual information object having a reduced feature vector x, based on the following:
  • f(z) a monotonic (e.g., sigmoid) function of z; and is a weight vector.
  • the classifier has a weight vector parameter, the determination of which is described in ⁇ 4.2.1.4.1 below, and a monotonic function having adjustable parameters, the determination of which is described in ⁇ 4.2.1.4.2 below.
  • the weight vector is determined, as shown in step 810, and the monotonic function (and its adjustable parameters) is determined, as shown in step 820.
  • the process 140' is terminated via return node 830.
  • the classifier of each category has a weight vector parameter.
  • the weight vector parameter may be generated by methods used to train a support vector machine.
  • support vector machines or "SVMs” may be trained with known objects having known classifications to define a hyperplane, or hypersurface, which separates points in n-dimensional feature vector space which are in the category and not in the category.
  • a pair of adjustable parameters, (a "weight vector") and b (a "threshold”) can be defined such that all of the training data x, having a known classification y, satisfy the following constraints:
  • i 1, . . ., number of training examples
  • x is the i th input vector
  • w is a weight vector parameter
  • b is a threshold parameter
  • ⁇ ⁇ is a known classification associated with the i th training example and is +1 if the example is "in the class” and -1 if the training example is "not in the class”.
  • the inequality conditions of equations (7) and (8) can be combined into the following inequality condition by multiplying each side of the equations by y and +1 or -1, and subtracting 1 from both sides:
  • training an SVM presents a constrained optimization (e.g., minimization) problem. That is, a function is minimized (referred to as "the objective function") subject to one or more constraints .
  • the objective function a function that is minimized (referred to as "the objective function") subject to one or more constraints .
  • a point that satisfies all constraints is referred to as a "feasible point” and is located in a “feasible region”.
  • a constrained optimization (e.g., minimization) problem is solved by a feasible point having no feasible descent directions.
  • Lagrange multipliers provide a way to transform equality-constrained optimization problems into unconstrained extremization problems.
  • f (w)
  • 2 and g(w) ⁇ ⁇ (Xj. dot + b) - 1
  • the solution to the dual quadratic programming problem can be determined by maximizing the Lagrangian dual problem L D .
  • the support vector machine can therefore be trained by solving the dual Lagrangian quadratic programming problem.
  • the points for which is greater than zero are the "support vectors".
  • These support vectors are the critical elements of the training set since they lie closest to the decision boundary and therefore define the margin from the decision boundary.
  • equation (13) merely represents the optimization problem with constraints in a more manageable form than the problems of equations (7) and (8).
  • Numerical methods such as constrained conjugate gradient ascent, projection methods, Bunch-Kaufman decomposition, and interior points methods, may be used to solve the quadratic problem. Suffice to say that these numerical methods are not trivial, particularly when there are a large number of examples i in the training set. Referring to equation (13), solving the quadratic problem involves an "nte" by "nte” matrix (where "nte” is the number of examples in the training set) .
  • the quadratic programming problem may also be solved with a chunking algorithm.
  • Vladimir Vapnik who is credited with developing SVMs, introduced an iterative training method referred to as "chunking". Though the "chunking" method of training SVMs is understood by those skilled in the art, it will be introduced here for the reader's convenience.
  • a subproblem consisting of a union of non-zero Lagrange multipliers from a previous step and a number of worst training examples that violate Kuhn-Tucker conditions is solved. These steps are repeated until all examples fulfill the Kuhn-Tucker conditions.
  • the matrix at each step is now on the order of the number of non-zero Lagrange multipliers, squared. Since the matrices are smaller, the algorithm can be faster. Although chunking makes training SVMs tractable, it is still too slow for some applications. For example, the number of non-zero Lagrange multipliers can still grow into the thousands in certain applications. The quadratic problem is still solved with a numerical quadratic programming technique and each iteration is time consuming.
  • the Osuna method stores a fixed size matrix, rather than a matrix that grows with the number of non-zero Lagrange multipliers.
  • a fixed number of Lagrange multipliers are optimized at every step. With each new step, one of the old Lagrange multipliers is removed from consideration and a new example whose output violates the Kuhn-Tucker conditions is added.
  • the Osuna paper provides a proof that the quadratic problem will decrease with every step such that the Osuna method converges to a solution.
  • the Osuna method still disadvantageously requires a numerical solution at each step. Although it converges to a solution, the Osuna method uses an entire numerical quadratic programming problem just to "fix" one Lagrange multiplier that violates the Kuhn-Tucker conditions.
  • the threshold b is zero (0) .
  • the Lagrange multipliers corresponding to each of examples of the training set, are initialized. For example, all of the Lagrange multipliers may be initialized by setting them to zero (0) .
  • the solution is optimized for the example.
  • Each example is examined for optimality by determining whether or not a Kuhn-Tucker condition is violated. If the solution is not optimal (e.g., if a Kuhn-Tucker condition is violated) , then a new Lagrange multiplier for the n th training example is determined. More specifically, the new Lagrange multiplier for the n th training example may be determined by the following expression: eeww old + K-. uhn - Tucker Error (14) k( ⁇ réelle, ⁇ vide)
  • the Kuhn-Tucker (KT) Error may be expressed as:
  • the denominator is the kernel function which, in the case of a linear support vector machine, is the dot product of the feature vectors x of the n th training example.
  • the new Lagrange multiplier is then limited (or clipped) as follows:
  • An initial sweep processes all examples of the training set. Assuming that a Lagrange multiplier was changed, the next sweep processes only the non-boundary examples of the training set. Subsequent sweeps also process only the non-boundary Lagrange multipliers until no Lagrange multipliers change. Then, the next sweep processes the entire set of training examples. If no Lagrange multipliers were changed, then processing ends. If, on the other hand, a Lagrange multiplier was changed, the processing continues as discussed above. Naturally, each sweep could process all of the training examples in the alternative .
  • the entire SVM state may be stored as a weight vector rather than a weighted sum of kernels.
  • the weight vector is directly updated when the Lagrange multiplier is changed. That is:
  • the Kuhn-Tucker conditions may be checked using interval arithmetic: namely, if the condition is approximately met within a certain tolerance, then condition is met.
  • the tolerance may be set to le-3. This should be accurate enough for recognition machines, and will prevent the method from spending a large amount of computational resources generating unneeded precision in the Lagrange multipliers.
  • the embodiment shown can be dramatically sped up by using a sparse evaluation of the dot product.
  • the dot product between the weight and the input point is the sum of the weight elements that correspond to the non-zero input dimensions. This can be expressed as:
  • This optimization is useful for training linear SVMs to be used for text classification since most text objects are represented by sparse feature vectors (e.g., text objects have relatively few of all words of a feature vector) .
  • the alpha vector which includes the values of the Lagrange multipliers for all training examples, may be stored as a vector of length equal to the size of the training set. That is:
  • the alpha vector may be stored as two (2) vectors of length equal to the number of non-zero Lagrange multipliers, one storing the identity (index) of the non-zero Lagrange multiplier, the other storing the value of the non-zero Lagrange multiplier .
  • the quadratic programming process is subject to a linear equality constraint (i.e. the threshold b is not necessarily equal to zero (0)).
  • the Lagrange multipliers corresponding to each of examples of the training set, are initialized. Recall that, since most of the Lagrange multipliers may be zero (0), all of the Lagrange multipliers may be initialized by setting them to zero (0) . For each training example from the set of training examples, it is determined whether a
  • Kuhn-Tucker condition is violated. If a Kuhn-Tucker condition is violated, another example is selected from the training set (as will be discussed below) , thereby creating a pair of Lagrange multipliers to be jointly optimized. An attempt is made to determine new
  • Lagrange multipliers such that the examples are jointly optimized. It is determined whether the examples were optimized. If they were not optimized, another feature vector is selected from the training set thereby creating another pair of Lagrange multipliers.
  • the first step in the joint optimization of the two (2) Lagrange multipliers is to determine the bounds on one of the variables. Either the first or the second multiplier can be bounded. Here, the second is chosen .
  • L be the lower bound on the second Lagrange multiplier and H be the higher bound on the second Lagrange multiplier.
  • yi be the desired output of the first example and let y 2 be the desired output of the second example.
  • cti be the current value of the first Lagrange multiplier and ⁇ 2 be the current value of the second Lagrange multiplier. If yi is the same as y 2 , then the following bounds are computed:
  • the new optimized value of ⁇ 2 may be computed via the formula:
  • the new value of the second Lagrange multiplier is less than L, then it is set to L. Conversely, if the new value of the second Lagrange multiplier is greater than H, then it is set to H. If the new clipped (or limited) value of the second Lagrange multiplier is the same as the old value, then no optimization is possible it is determined that the training examples were not optimized. Otherwise, the new value of the first Lagrange multiplier is then derived from the clipped (or limited) value of the second Lagrange multiplier:
  • the weights and thresholds are updated to reflect the new Lagrange multipliers so that other violations of the Kuhn Tucker conditions can be detected.
  • the second Lagrange multiplier pair may be selected based on the following heuristic.
  • the ideal second Lagrange multiplier would change the most upon joint optimization.
  • An easy approximation to the change upon optimization is the absolute value of the numerator in the change in the second Lagrange multiplier:
  • the QP2 process seeks out a non-boundary Lagrange multiplier ( ⁇ 0 or C) that has a true error that is the most opposite of the true error of the first Lagrange multiplier.
  • a hierarchy of heuristics may be used to find a second example to make forward progress on the joint optimization step. If the first heuristic described above fails, the QP2 process will select a non-boundary Lagrange multiplier as the other Lagrange multiplier of the pair to be jointly optimized. If this heuristic fails for all non-boundary Lagrange multipliers, any of the other Lagrange multipliers may be selected as the other Lagrange multiplier of the pair.
  • the QP2 process first sweeps through all training examples of the training set. Assuming that a Lagrange multiplier was changed, the next sweep processes only the non-boundary training examples of the training set. Subsequent sweeps also process only the non-boundary Lagrange multipliers until no Lagrange multipliers change. Then, the next sweep processes the entire set of training examples. If no Lagrange multipliers change, processing ends. If, on the other hand, a Lagrange multiplier changes, the processing continues as discussed above. Naturally, in an alternative methodology, all training examples of the training set could be processed on every sweep.
  • the Lagrange multipliers can be stored as a either one array having a size corresponding to the number of training examples (also referred to as "the full array") or two (2) arrays that represent a larger sparse array.
  • the real error mentioned used when jointly optimizing a pair of Lagrange multipliers is the desired output of the SVM less the actual output of the SVM (yi - Oi) .
  • These real errors may be cached for the examples that have a non-boundary (i.e., non-zero or a non-C) Lagrange multiplier. Caching the real errors allows an second example to be intelligently chosen.
  • the linear SVM may be stored as one weight vector, rather than as a linear superpositions of a set of input points. If the input vector is sparse, the update to the weight vector can be accomplished in a known sparse manner, also.
  • the weight vector w for a given category c can be determined by training an SVM, for example, based on one of the foregoing (e.g., Chunking, Osuna, and Platt) methods.
  • the text classifier employs a monotonic (e.g., sigmoid) function to classify textual information objects.
  • the characteristics of the sigmoid function may be adjusted using constants A and B (also referred to as “adjustable parameters") such that:
  • the constants A and B may be determined by using a maximum likelihood on the set of training examples of textual information objects.
  • the logarithm of the likelihood that the training targets y were generated by the probability distribution f, assuming independence between training examples, is computed.
  • the log likelihood is described by:
  • is a known result (+1 or -1); and nte is the number of training examples.
  • the optimal sigmoid parameters are then derived by maximizing the log likelihood over possible sigmoid parameters . ⁇ 4.2.1.4.2.1.1 KNOWN
  • Expression (28) may be maximized based on known unconstrained optimization techniques such as gradient descent, Levenberg-Marquardt, Newton-Raphson, conjugate gradient, and variable metric methods for example .
  • the Y J. values may be replaced by target values t x , which may be expressed as:
  • N + is the number of textual information objects of the training set in the class
  • N- is the number of textual information objects of the training set not in the class.
  • the target values (tt x ) expressed as:
  • This version of the target function limits the effect of Laplace' s rule of succession to be no farther away than 0.01 from the true target y. This has been empirically found to be desirable.
  • Figure 2 is a diagram of processes that may be used by a text classifier based on the present invention.
  • a textual information object 205 to be classified is applied to a feature extraction process (or more generally, a feature extractor) 210.
  • this feature extraction process 210 may be the same as the feature extraction process 110 discussed in ⁇ 4.2.1.1 above.
  • a feature reduction process or more generally, a feature reduction facility 220, described in ⁇ 4.2.2.2 below, using the set of reduced features generated by the feature reduction process 120 discussed in ⁇ 4.2.1.2 above, reduces the size of the feature vector associated with the textual information object 205.
  • a reduced feature vector associated with the textual information object 205 is generated.
  • a binarizing process 230 may then be applied to the reduced feature vector.
  • the reduced feature vector generated by the feature reduction process 220 (or the binarized reduced feature vector generated by the binarizing process 230) is applied to one or more classifiers 250.
  • These classifiers 250 may use a classification process described in ⁇ 4.2.2.4 below.
  • the weight parameter w and the parameters of the parameterizable monotonic function for each classifier 250 are those determined by the classification engine parameter generation process(es) 140 discussed in ⁇ 4.2.1.4 above.
  • a probability that the textual information object 205 belongs to a particular class is output.
  • the feature extraction process 210 may be analogous to the feature extraction process 110 discussed in ⁇ 4.2.1.1 above with reference to Figures 1 and 6.
  • An exemplary feature reduction process 220 is now described. Recall that the feature reduction process 120 discussed above reduces (e.g., based on category independent and category dependent techniques) the feature vectors corresponding to the textual information objects of the set of training examples 115. This process 120 also produces a reduced feature set, which may be an index to features deemed relevant or important for the text classification. For example, if the original feature vector included 100 elements, the reduced feature set may include an index to the 2 nd , 5 th , 20 th , 21 st , and 89 th elements of the feature of the vector. Thus, the 100 element feature vector is reduced to a five (5) element reduced feature vector. Typically, the number of elements in the feature vector and reduced feature vector will be much larger than those used in this illustration.
  • the exemplary feature reduction process 220 uses the index of relevant or important features to map the values of the features to sequential indices of the reduced feature vector.
  • the value of the 2 nd element would be placed in the first element of the reduced feature vector
  • the value of the 5 th element would be placed in the second element of the reduced feature vector
  • the value of the 20 th element would be placed in the third element of the reduced feature vector
  • the value of the 21 st element would be placed in the fourth element of the reduced feature vector
  • the value of the 89 th element would be placed in the fifth element of the reduced feature vector.
  • a textual information object is not represented by a single feature vector array, but rather, using sparse techniques known to those skilled in the art, it is represented by smaller feature index and value arrays.
  • the reduced feature set may be provided to the feature extraction process 210.
  • the feature extraction process only extracts those features belonging to the reduced feature set. In this way, a feature reduction is carried out while the features are being extracted.
  • the optional binarizing process 230 may be analogous to the binarizing process 150 discussed in ⁇ 4.2.1.3 above with reference to Figure 1. ⁇ 4.2.2.4 CLASSIFICATION
  • the classification processes 250 may be effected as follows.
  • the probability that the textual information object 205 (as represented by reduced feature vector x) is in category i can be expressed as:
  • This text classifier has a number of advantages. First, the probabilities output by the classifiers of the various categories can be compared across categories. Second, if the target technique, described in ⁇ 4.2.1.4.2.1.2 above is used for determining the monotonic function and its adjustable parameters, that function will not be overfit when relatively small amounts of training data are available for a particular category.
  • FIG. 11A and the following discussion provide a brief, general description of an exemplary apparatus in which at least some aspects of the present invention may be implemented.
  • the present invention will be described in the general context of computer-executable instructions, such as program modules, being executed by a personal computer. However, the methods of the present invention may be effected by other apparatus.
  • Program modules may include routines, programs, objects, components, data structures, etc. that perform a task(s) or implement particular abstract data types.
  • those skilled in the art will appreciate that at lease some aspects of the present invention may be practiced with other configurations, including hand-held devices, multiprocessor systems, microprocessor-based or programmable consumer electronics, network computers, minicomputers, set top boxes, mainframe computers, and the like.
  • At least some aspects of the present invention may also be practiced in distributed computing environments where tasks are performed by remote processing devices linked through a communications network.
  • program modules may be located in local and/or remote memory storage devices.
  • an exemplary apparatus 1100 for implementing at least some aspects of the present invention includes a general purpose computing device in the form of a conventional personal computer 1120.
  • the personal computer 1120 may include a processing unit 1121, a system memory 1122, and a system bus 1123 that couples various system components including the system memory to the processing unit 1121.
  • the system bus 1123 may be any of several types of bus structures including a memory bus or memory controller, a peripheral bus, and a local bus using any of a variety of bus architectures.
  • the system memory may include read only memory (ROM) 1124 and/or random access memory (RAM) 1125.
  • a basic input/output system 1126 (BIOS) , containing basic routines that help to transfer information between elements within the personal computer 1120, such as during start-up, may be stored in ROM 1124.
  • the personal computer 1120 may also include a hard disk drive 1127 for reading from and writing to a hard disk, (not shown) , a magnetic disk drive 1128 for reading from or writing to a (e.g., removable) magnetic disk 1129, and an optical disk drive 1130 for reading from or writing to a removable (magneto) optical disk 1131 such as a compact disk or other (magneto) optical media.
  • the hard disk drive 1127, magnetic disk drive 1128, and (magneto) optical disk drive 1130 may be coupled with the system bus 1123 by a hard disk drive interface 1132, a magnetic disk drive interface 1133, and a (magneto) optical drive interface 1134, respectively.
  • the drives and their associated storage media provide nonvolatile storage of machine readable instructions, data structures, program modules and other data for the personal computer 1120.
  • a number of program modules may be stored on the hard disk 1123, magnetic disk 1129, (magneto) optical disk 1131, ROM 1124 or RAM 1125, such as an operating system 1135, one or more application programs 1136, other program modules 1137, and/or program data 1138 for example.
  • a user may enter commands and information into the personal computer 1120 through input devices, such as a keyboard 1140 and pointing device 1142 for example.
  • Other input devices such as a microphone, joystick, game pad, satellite dish, scanner, or the like may also be included.
  • These and other input devices are often connected to the processing unit 1121 through a serial port interface 1146 coupled to the system bus.
  • input devices may be connected by other interfaces, such as a parallel port, a game port or a universal serial bus (USB) .
  • a monitor 1147 or other type of display device may also be connected to the system bus 1123 via an interface, such as a video adapter 1148 for example.
  • the personal computer 1120 may include other peripheral output devices (not shown) , such as speakers and printers for example.
  • the personal computer 1120 may operate in a networked environment which defines logical connections to one or more remote computers, such as a remote computer 1149.
  • the remote computer 1149 may be another personal computer, a server, a router, a network PC, a peer device or other common network node, and may include many or all of the elements described above relative to the personal computer 1120, although only a memory storage device 1150 has been illustrated in Figure 11A.
  • the logical connections depicted in Figure 11A include a local area network (LAN) 1151 and a wide area network (WAN) 1152, an intranet and the Internet.
  • LAN local area network
  • WAN wide area network
  • the personal computer 1120 When used in a LAN, the personal computer 1120 may be connected to the LAN 1151 through a network interface adapter (or "NIC") 1153.
  • NIC network interface adapter
  • the personal computer 1120 When used in a WAN, such as the Internet, the personal computer 1120 may include a modem 1154 or other means for establishing communications over the wide area network 1152.
  • the modem 1154 which may be internal or external, may be connected to the system bus 1123 via the serial port interface 1146.
  • at least some of the program modules depicted relative to the personal computer 1120 may be stored in the remote memory storage device.
  • the network connections shown are exemplary and other means of establishing a communications link between the computers may be used.
  • FIG 11B is a more general machine 1100 which may effect one or more of the processes discussed above.
  • the machine 1100 basically includes a processor(s) 1102, an input/output interface unit(s) 1104, a storage device (s) 1106, and a system bus or network 1108 for facilitating data and control communications among the coupled elements .
  • the processor (s) 1102 may execute machine-executable instructions to effect one or more aspects of the present invention. At least a portion of the machine executable instructions may be stored (temporarily or more permanently) on the storage devices (1106) and/or may be received from an external source via an input interface unit 1104.
  • the methods and apparatus advantageously: (i) reduce the size of a feature vector representing a textual information object while maintaining the features important for classification; (ii) avoid generating a classifier overfit to relatively low amounts of training data; (iii) train relatively fast; and/or (iv) classify relatively fast.

Abstract

A text classifier and building the text classifier by determining appropriate parameters for the text classifier.

Description

METHODS AND APPARATUS FOR CLASSIFYING TEXT AND FOR BUILDING A TEXT CLASSIFIER
§ 1. BACKGROUND OF THE INVENTION
§ 1.1 Field of the Invention
The present invention concerns determining whether an object, such as a textual information object for example, belongs to a particular category or categories. This determination is made by a classifier, such as a text classifier for example. The present invention also concerns building a (text) classifier by determining appropriate parameters for the (text) classifier.
§ 1.2 Related Art
§ 1.2.1 THE NEED FOR TEXT CLASSIFICATION
To increase their utility and intelligence, machines, such as computers for example, are called upon to classify (or recognize) objects to an ever increasing extent. For example, computers may use optical character recognition to classify handwritten or scanned numbers and letters, pattern recognition to classify an image, such as a face, a fingerprint, a fighter plane, etc., or speech recognition to classify a sound, a voice, etc.
Machines have also been called upon to classify textual information objects, such as a textual computer file or document for example. The applications for text classification are diverse and important. For example, text classification may be used to organize textual information objects into a hierarchy of predetermined classes or categories for example. In this way, finding (or navigating to) textual information objects related to a particular subject matter is simplified. Text classification may be used to route appropriate textual information objects to appropriate people or locations. In this way, an information service can route textual information objects covering diverse subject matters (e.g., business, sports, the stock market, football, a particular company, a particular football team) to people having diverse interests. Text classification may be used to filter textual information objects so that a person is not annoyed by unwanted textual content (such as unwanted and unsolicited e-mail, also referred to as junk e-mail, or "spam") . As can be appreciated from these few examples, there are many exciting and important applications for text classification.
§ 1.2.2 KNOWN TEXT CLASSIFICATION METHODS
In this section, some known classification methods are introduced. Further, acknowledged or suspected limitations of these classification methods are introduced. First, rule-based classification is discussed in § 1.2.2.1 below. Then, classification systems which use both learning elements and performance elements in are discussed in § 1.2.2.2 below.
§ 1.2.2.1 RULE BASED CLASSIFICATION
In some instances, textual content must be classified with absolute certainty, based on certain accepted logic. A rule-based system may be used to effect such types of classification. Basically, rule-based systems use production rules of the form:
IF condi tion , THEN fact.
The conditions may include whether the textual information includes certain words or phrases, has a certain syntax, or has certain attributes. For example, if the textual content has the word "close", the phrase "nasdaq" and a number, then it is classified as "stock market" text.
Unfortunately, in many instances, rule-based systems become unwieldy, particularly in instances where the number of measured or input values (or features or characteristics) becomes large, logic for combining conditions or rules becomes complex, and/or the number of possible classes becomes large. Since textual information may have many features and complex semantics, these limitations of rule-based systems make them inappropriate for classifying text in all but the simplest applications. Over the last decade or so, other types of classifiers have been used increasingly. Although these classifiers do not use static, predefined logic, as do rule-based classifiers, they have outperformed rule-based classifiers in many applications. Such classifiers are discussed in § 1.2.2.2 below and typically include a learning element and a performance element. Such classifiers may include neural networks, Bayesian networks, and support vector machines. Although each of these classifiers is known, each is briefly introduced below for the reader's convenience.
§ 1.2.2.2 CLASSIFIERS HAVING LEARNING AND PERFORMANCE ELEMENTS
As just mentioned at the end of the previous section, classifiers having learning and performance elements outperform rule-based classifiers, in many applications. To reiterate, these classifiers may include neural networks (introduced in § 1.2.2.2.1 below for the reader's convenience), Bayesian networks (introduced in § 1.2.2.2.2 below for the reader's convenience) , and support vector machines (introduced in § 1.2.2.2.3 below for the reader's convenience).
§ 1.2.2.2.1 NEURAL NETWORKS
A neural network is basically a multilayered, hierarchical arrangement of identical processing elements, also referred to as neurons. Each neuron can have one or more inputs but only one output. Each neuron input is weighted by a coefficient. The output of a neuron is typically a function of the sum of its weighted inputs and a bias value. This function, also referred to as an activation function, is typically a sigmoid function. That is, the activation function may be S-shaped, monotonically increasing and asymptotically approaching fixed values (e.g., +1, 0, -1) as its input (s) respectively approaches positive or negative infinity. The sigmoid function and the individual neural weight and bias values determine the response or "excitability" of the neuron to input signals .
In the hierarchical arrangement of neurons, the output of a neuron in one layer may be distributed as an input to one or more neurons in a next layer. A typical neural network may include an input layer and two (2) distinct layers; namely, an input layer, an intermediate neuron layer, and an output neuron layer. Note that the nodes of the input layer are not neurons. Rather, the nodes of the input layer have only one input and basically provide the input, unprocessed, to the inputs of the next layer. If, for example, the neural network were to be used for recognizing a numerical digit character in a 20 by 15 pixel array, the input layer could have 300 neurons (i.e., one for each pixel of the input) and the output array could have 10 neurons (i.e., one for each of the ten digits). The use of neural networks generally involves two (2) successive steps. First, the neural network is initialized and trained on known inputs having known output values (or classifications) . Once the neural network is trained, it can then be used to classify unknown inputs. The neural network may be initialized by setting the weights and biases of the neurons to random values, typically generated from a Gaussian distribution. The neural network is then trained using a succession of inputs having known outputs (or classes) . As the training inputs are fed to the neural network, the values of the neural weights and biases are adjusted (e.g., in accordance with the known back-propagation technique) such that the output of the neural network of each individual training pattern approaches or matches the known output. Basically, a gradient descent in weight space is used to minimize the output error. In this way, learning using successive training inputs converges towards a locally optimal solution for the weights and biases. That is, the weights and biases are adjusted to minimize an error.
In practice, the system is not trained to the point where it converges to an optimal solution.
Otherwise, the system would be "over trained" such that it would be too specialized to the training data and might not be good at classifying inputs which differ, in some way, from those in the training set. Thus, at various times during its training, the system is tested on a set of validation data. Training is halted when the system's performance on the validation set no longer improves.
Once training is complete, the neural network can be used to classify unknown inputs in accordance with the weights and biases determined during training. If the neural network can classify the unknown input with confidence, one of the outputs of the neurons in- the output layer will be much higher than the others.
To ensure that the weight and bias terms do not diverge, the algorithm uses small steps. Furthermore, the back propagation (gradient descent) technique used to train neural networks is relatively slow. (See, e.g., the article: Schϋtze, et al., "A Comparison of Classifiers and Document Representations for the Routing Problem", International ACM SIGIR Conference on Research and Development in Information Retrieval, Section 5 (1995) (Hereafter referred to as "the Schϋtze article"). Consequently, convergence is slow. Also, the number of neurons in the hidden layer cannot easily be determined a priori . Consequently, multiple time-consuming experiments are often run to determine the optimal number of hidden neurons.
§ 1.2.2.2.2 BAYESIAN NETWORKS
Having introducing neural networks above, Bayesian networks are now briefly introduced. Typically, Bayesian networks use hypotheses as intermediaries between data (e.g., input feature vectors) and predictions (e.g., classifications). The probability of each hypothesis, given the data ("P (hypo I data) ") , may be estimated. A prediction is made from the hypotheses using posterior probabilities of the hypotheses to weight the individual predictions of each of the hypotheses. The probability of a prediction X given data D may be expressed as:
Figure imgf000010_0001
where Hx, is the ith hypothesis. A most probable hypothesis Hx that maximizes the probability of Hx given D (P(H;L|D)) is referred to as a maximum a posterior hypothesis (or
Figure imgf000010_0002
and may be expressed as follows :
P{X \ D)*P(X \ HMAP)
Using Bayes1 rule, the probability of a hypothesis Hx given data D may be expressed as:
_ P {D \ Hl)P{Hl)
P{H, \ D) P(D)
The probability of the data D remains fixed. Therefore, to find HMAP, the numerator must be maximized. The first term of the numerator represents the probability that the data would have been observed given the hypothesis i. The second term represents the prior probability assigned to the given hypothesis i.
A Bayesian network includes variables and directed edges between the variables, thereby defining a directed acyclic graph (or "DAG"). Each variable can assume any of a finite number of mutually exclusive states. For each variable A, having parent variables Bi, . . . Bn, there is an attached probability table (P(A|Bι, . . . Bn) . The structure of the Bayesian network encodes the assumptions that each variable is conditionally independent of its non-descendants, given its parent variables.
Assuming that the structure of the Bayesian network is known and the variables are observable, only the set of conditional probability tables need be learned. These tables can be estimated directly using statistics from a set of learning examples. If the structure is known but some variables are hidden, learning is analogous to neural network learning discussed above.
An example of a simple Bayesian network is introduced below. A variable "MML" may represent a "moisture of my lawn" and may have states "wet" and "dry". The MML variable may have "rain" and "my sprinkler on" parent variables each having "Yes" and "No" states. Another variable, "MNL" may represent a "moisture of my neighbor's lawn" and may have states "wet" and "dry". The MNL variable may share the "rain" parent variable. In this example, a prediction may be whether my lawn is "wet" or "dry". This prediction may depend of the hypotheses (i) if it rains, my lawn will be wet with probability (xj.) and (ii) if my sprinkler was on, my lawn will be wet with probability (x2) . The probability that it has rained or that my sprinkler was on may depend on other variables. For example, if my neighbor's lawn is wet and they don't have a sprinkler, it is more likely that it has rained. An example of a Bayesian network associated with the "wet grass" problem is presented in the text: Jensen, An Introduction to Bayesian Networks, pp. 22-25, Spinger-Verlag, New York (1997).
As discussed above, the conditional probability tables in Bayesian networks may be trained, as was the case with neural networks. Advantageously, by allowing prior knowledge to be provided for, the learning process may be shortened. Unfortunately, however, prior probabilities for the conditional probabilities are usually unknown, in which case a uniform prior is used. §1.2.2.2.2.3 SUPPORT VECTOR MACHINES
Support vector machines (or "SVMs") are another type of trainable classifier. SVMs are reportedly more accurate at classification than naive Bayesian networks in certain applications, such as text classification for example. (See, e.g., the article, Joachims, "Text Categorization with Support Vector Machines: Learning with Many Relevant Features", LS-8 , Report 23, University of Dortmund Computer Science Department (Nov. 1997).) They are also reportedly more accurate than neural networks in certain applications, such as reading handwritten characters for example. (See, e.g., the article, LeCun et al . , "Learning Algorithms for Classification: A Comparison on Handwritten Digit Recognition, " Neural Networks : The Statistical Mechanics Perspective, Oh et al . (Eds.), pp. 261-276, World Scientific (1995) . Unfortunately, however, SVMs reportedly take longer to train than naive Bayesian classifiers. A new method and apparatus to build (or train) SVMs in an efficient manner is disclosed in U. S. Patent Application Serial No. 09/055,477, by John Platt, entitled "Methods and Apparatus for Building a Support Vector Classifier", filed on April 6, 1998 and incorporated by reference.
Although SVMs are known to those skilled in the art, their theory and operation will be introduced for the reader's convenience. An object to be classified may be represented by a number of features. If, for example, the object to be classified is represented by two (2) features, it may be represented by a point in two (2) dimensional space. Similarly, if the object to be classified is represented by n features, also referred to as a "feature vector", it may be represented by a point in n-dimensional space. The simplest form of an SVM defines a plane in the n-dimensional space (also referred to as a hyperplane) which separates feature vector points associated with objects "in a class" and feature vector points associated with objects "not in the class". A number of classes can be defined by defining a number of hyperplanes. The hyperplane defined by a trained SVM maximizes a distance (also referred to as an Euclidean distance) from it to the closest points (also referred to as "support vectors") "in the class" and "not in the class". A hyperplane is sought which maximizes the distances between the support vectors and the hyperplane, so that the SVM defined by the hyperplane is robust to input noise. The hyperplane (or hypersurface) is defined by a training process, some of which are discussed in § 4.2.1.4.1 below.
§ 1.2.2.2.4 CHALLENGES TO USING CLASSIFIERS HAVING LEARNING AND PERFORMANCE ELEMENTS FOR TEXT
CLASSIFICATION
Although, as discussed above, rule-based classifiers are feasible in only the simplest text classification applications, some significant challenges exist when using systems having learning and performance ELEMENTS (ALSO REFERRED TO AS "LEARNING MACHINES") FOR text classification. Some of the more significant challenges will be introduced in §§ 1.2.2.2.4.1 through 1.2.2.2.4.3 below.
§ 1.2.2.2.4.1 FEATURE VECTOR SIZE
When training learning machines for text classification, a set of learning examples are used. Each learning example includes a vector of features associated with a textual information object. In some applications, such feature vectors may have on the order of 108 features. A large number of features can easily be generated by considering the presence or absence of a word in a document to be a feature. If all of the words in a corpus are considered as possible features, then there can be millions of unique features. For example, web pages have many unique strings and can generate millions of features. An even larger number of features are possible if pairs or more general combinations of words or phrases are considered, or if the frequency of occurrence of words is considered. The number of features in a feature vector may be reduced by so-called "feature reduction" or "feature selection" methods such that a reduced feature vector, having a subset of the features of the original feature vector, is produced. Indeed, some believe that for learning machine text classifiers to be feasible, feature selection is needed.
Known feature selection techniques include
DF-Thresholding (See, e.g., the paper: Yang and Peterson, "A Comparative Study on Feature Selection in Text Categorization, " International Conference on Machine Learning (1997); hereafter referred to as: the Yang-Peterson article"), Chi-Squared Test (See, e.g., the Schϋtze article) , , Term Strength Criterion (See, e.g., the article: Yang and Wilbur, "Using Corpus Statistics to Remove Redundant Words in Text Categorization, " Journal of the American Society for Information Science, Vol. 47, No. 5, pp. 357-369 (1996); hereafter referred to as "the Yang-Wilbur article"), Information Gain Criteria (See, e.g., the Yang-Peterson article) , correlation coefficient which is the square root of the Chi-Squared measure (See, e.g., the article: Ng, et al, "Feature Selection,
Perceptron Learning, and a Usability Case Study for Text Categorization", Proceedings of SIGIR'97, p.67-73 (1997) and Latent Semantic Indexing (or "LSI") using Singular Value Decomposition (or "SVD") which is a technique that represents features by a low-dimensional linear combination of orthogonal indexing variables (See, e.g., Schϋtze, et al).
However, some have concluded that feature selection should be avoided. For example, the Joachims article, which espouses the use of SVMs for text classification, concludes that feature selection is likely to hurt performance due to a loss of information. (See, e.g., the Joachims article, page 3.) The Joachims article further concludes that since SVMs can generalize well in high dimensional feature spaces, they eliminate the need for feature selection thereby making text categorization easier. (See, e.g., the Joachims article, page 11.) Thus, while some in the art have found feature selection useful, when SVMs are used for text classification, feature selection has been avoided in some instances.
§ 1.2.2.2.4.2 OVERFITTING
When a learning machine is trained, it is trained based on training examples from a set of feature vectors. In general, performance of a learning machine will depend, to some extent, on the number of training examples used to train it. Even if there are a large number of training examples, there may be a relatively low number of training examples which belong to certain categories. Many learning machines utilize sigmoid functions (that is, s-shaped, monotonic functions) to determine an output (e.g., a determination as to whether an object is in a category or not) based on an input (e.g., a feature vector of an unknown object). Unfortunately, when there is little training data available for a given category, the parameters of the sigmoid function are not well defined. Consequently, when there is little training data available for a given category, the sigmoid function can be overfit to past data. Such overfitting can lead to unwarranted confidence in the output of the learning machine and should be avoided.
§ 1.2.2.2.4.3 CLASSIFICATION SPEED
Training time and response (e.g., classification) time are two (2) important characteristics of a learning machine. For example, even those that espouse the use of learning machines, and SVMs in particular, for classifying text concede that the training time of SVMs is longer than other methods. This challenge is addressed in the U.S. Patent Application Serial No. 09/055,477, by John Platt, entitled "Methods and Apparatus for Building a Support Vector Classified", filed on April 6, 1998 and assigned to a common assignee (incorporated by reference) and will not be addressed in detail here. There remains the issue of response (or classification) time. The present inventors believe that the SVM classifier discussed in the Joachims paper classifies unknown objects in accordance with the following expression:
Figure imgf000019_0001
where: 0 = the classification output; nte = the number of training examples;
Gi = Lagrange multiplier training example i; Xx = feature vector of training example i; -j = feature vector of unknown object j; and yx = known output of training example i.
Using expression (1) to determine an output is relatively slow. Naturally, a faster classifier would be desirable.
The Schutze article discusses logistic regression classifier which employs a weight vector (β) derived using a maximum likelihood and the Newton-Raphson method of numerical optimization.
Although it is believed that this classifier is able to classify objects faster than the SVM classifier discussed in the Joachims paper, it is believed that the weight vector (β) determined does not provide the best classification results. § 2. SUMMARY OF THE INVENTION
The present invention may perform at least one (1) of two (2) basic functions, namely generating parameters for a classifier, and classifying objects, such as textural information objects.
Basically, parameters are generated for a classifier based on a set of training examples. A set of feature vectors may be generated from a set of training examples. The features of the set of feature vectors may be reduced. The parameters to be generated may include a defined monotonic (e.g., sigmoid) function and a weight vector. The weight vector may be determined by means of SVM training (or by another, known, technique). The monotonic (e.g., sigmoid) function may be defined by means of an optimization method.
The text classifier may include a weight vector and a defined monotonic (e.g., sigmoid) function. Basically, the output of the text classifier of the present invention may be expressed as:
1
0 =- l + e A(w, »x)+B (2)
where: Oc = a classification output for category c; c = a weight vector parameter associated with category c; x = is a (reduced) feature vector based on the unknown textual information object; and
A and B are adjustable parameters of a monotonic (e.g., sigmoid) function.
The calculation of the output from expression (2) is quicker than the calculation of the output from expression (1) .
Depending on the form of the object to be classified, the classifier may (i) convert a textual information object to a feature vector, and (ii) reduce the feature vector to a reduced feature vector having less elements.
§ 3. BRIEF DESCRIPTION OF THE DRAWINGS
Figure 1 is a high level diagram of processes that may be performed by a first aspect of the present invention.
Figure 2 is a high level diagram of processes that may be performed by a second aspect of the present invention.
Figure 3 is an exemplary data structure of a training example of a textual information object. Figures 3B and 3C are alternative exemplary sparse vector data structures of the textual information object . Figure 4 is an exemplary data structure of the training example of Figure 3, in which a feature vector has been reduced. Figures 4B and 4C are alternative exemplary sparse vector data structures of the reduced feature vector.
Figure 5 is an exemplary data structure of the training example of Figure 4, in which values of a reduced feature vector have been binarized. Figures 5B and 5C are alternative exemplary sparse vector data structures of the binarized reduced feature vector.
Figure 6 is a flow diagram of an exemplary feature extraction process.
Figure 7 is a flow diagram of an exemplary feature reduction process.
Figure 8 is a flow diagram of an exemplary processes for generating parameters for a classification engine.
Figure 9 depicts information used by some feature reduction techniques.
Figure 10 depicts various sigmoid functions.
Figure 11A is a block diagram of a computer environment which may be used to effect processes carried out by the present invention. Figure 11B is a block diagram of a machine which may be used to effect processes carried out by the present invention.
§ 4. DETAILED DESCRIPTION
The present invention concerns a novel classifier, such as a text classifier, and building such a (text) classifier. The following description is presented to enable one skilled in the art to make and use the invention, and is provided in the context of particular applications and their requirements. Various modifications to the disclosed embodiments will be apparent to those skilled in the art, and the general principles set forth below may be applied to other embodiments and applications. Thus, the present invention is not intended to be limited to the embodiments shown.
Regarding notation used in the specification, a vector may be depicted with an arrow "hat" or by bold typeface .
Functions which may be performed by the present invention are described in § 4.1 below. Then exemplary methodologies, and structures for carrying out the methodologies, are described in § 4.2 below. § 4 . 1 FUNCTIONS
The present invention may perform at least one (1) of two (2) basic functions. The first basic function, namely generating parameters for a classifier, is discussed in § 4.1.1 below. The second basic function, namely classifying text, is discussed in § 4.1.2 below.
§ 4.1.1 GENERATING PARAMETERS FOR A
CLASSIFIER
Basically, parameters are generated for a classifier based on a set of training examples. The present invention may function to generate a set of feature vectors from the set of training examples. The present invention may also function to reduce the number of features of feature vectors of the set. The parameters to be generated include a defined monotonic (e.g., sigmoid) function and a weight vector. The weight vector may be determined by means of SVM training. The monotonic (e.g., sigmoid) function may be defined by means of an optimization method. The optimization method may use a special target function to avoid overfitting.
§ 4.1.2 CLASSIFYING TEXT
The present invention may classify text. A text classifier may include a weight vector and a defined monotonic (e.g., sigmoid) function. Basically, the output of the text classifier may be expressed as:
0,=- l+e Λ(wc'x)+B (2)
where: 0C = a classification output for category c; wc = a weight vector parameter associated with category c; x = is a (reduced) feature vector based on the unknown textual information object; and A and B are adjustable parameters of a monotonic (e.g., sigmoid) function.
To reiterate, the calculation of the output from expression (2) is quicker than the calculation of the output from expression (1).
Depending on the form of the object to be classified, the classifier may (i) convert a textual information object to a feature vector, and (ii) reduce the feature vector to a reduced feature vector having less elements.
§ 4.2 EXEMPLARY STRUCTURE/METHODOLOGY
Exemplary processes and methods for generating parameters of a classifier are described in § 4.2.1 below. Exemplary processes and methods for classifying text are described in § 4.2.2 below. Finally, exemplary apparatus (or architecture) for effecting the described processes and methods are described in § 4.2.3 below.
§ 4.2.1 PROCESSES/METHODS FOR GENERATING PARAMETERS OF A CLASSIFIER
Figure 1 is a diagram which depicts processes and data which may be used by the present invention to generate parameters of a text classifier. Assuming that a set of training data 115 (e.g., textual information objects and category labels) is provided, a feature extraction process (or more generally, a feature extractor) 110 generates a feature vector corresponding to each of the textual information objects of the training set. The feature vectors may be stored at storage location 125. An exemplary feature extraction process is discussed in § 4.2.1.1 below.
Figure 3 illustrates an exemplary data structure 300 of a training object which includes a feature vector 310 and a category vector 320. The feature vector 310 includes fields 312 corresponding to features (or attributes) which were derived from a textual information object. In this example, there are "n" features in the feature vector 310. The feature (or attribute) fields 312 of the feature vector 310 include corresponding values derived from the textual information object. For example, if it is assumed that the textual object may include words from a 2.5 million word vocabulary, then the feature vector 310 may include 2.5 million fields, each corresponding to a word of the vocabulary. The value stored in a field may correspond to the number of times a particular word of the vocabulary appears in the textual information object.
Since most values in a feature vector of a textual information object are zero (0), sparse vector techniques may be used such that the feature vector is represented by two (2) smaller arrays rather than one large array. A first array indexes features having non-zero values. A second array includes the values of the features indexed by the first array. For example, referring to Figure 3B, if only the first, 50th, 57th, 101st, and 220th features of a feature vector have non-zero values, a sparse data structure 310' may be used instead of the feature vector 310 of Figure 3. The sparse data structure 310' includes a first array 350 which includes the indices (e.g., 1, 50, 57, 101, and 220) to the non-zero features, and a second array 360 which includes the values of the indexed features. Figure 3C depicts another sparse data structure 310" which has alternating indexing fields and value fields corresponding to the non-zero features.
The category vector 320 includes fields 322 corresponding to "m" different categories. The fields 322 of the category vector 310 may include a binary value which indicates whether or not the textual information object belongs to a particular category. A textual information object may belong to more than one category.
The data structures 300 corresponding to each textual information object of the training set may be stored at storage area 125. A feature reduction process (or more generally, a feature reduction facility) 120 accepts the data structures 300, either from the storage area 125 or the feature extraction process 110, and reduces the number of features of the feature vector 310 from "n" to "z", where "z" is less than "n", and may be as much as three (3) orders of magnitude less than "n" . Referring to Figure 4, a new data structure 400 is generated by the feature reduction process 120. The data structure 400 includes a category vector 320 which corresponds to that of the data structure 300. The data structure 400 also includes a reduced feature vector 410 having fields 412 corresponding to features (or attributes) of the textual information objects of the training set that are determined to be relatively important features by the feature reduction process 120. The feature reduction process 120 also produces a reduced feature set which defines which "z" of the "n" fields 412 of the feature vector are considered to be relatively important. An exemplary feature reduction process 120 is discussed in § 4.2.1.2 below. Although each textual information object was described as being a single, relatively large, feature vector array, in practice, sparse array techniques are used such that each textual information object may be represented by two (2) smaller arrays, rather than one (1) large array, as discussed above. Referring for example to Figure 4B, if the first and 57th features are not considered important, a sparse data structure 410' results. The sparse data structure 410' includes an first array 450 which includes the indices (e.g., 50, 101, and 220) of non-zero, selected, features, and a second array 460 which includes the values of the features indexed by the first array 450. Figure 4D depicts another sparse data structure 410" which has alternating indexing fields and value fields corresponding to the non-zero, selected, features.
An optional (depicted in dashed lines) binarizing process 150 may be used to convert each value of each field 412 of the reduced feature vector 412 of each of the objects of the training set, to a binary value. This conversion of values to binary values may take place by applying the values to a predetermined (or derived) threshold, in the case of each field 412. Thus, a different threshold may be associated with each field 412. Naturally, a single threshold could be applied to all fields 412. Figure 5 illustrates a data structure which may be used to represent the resulting reduced feature vector containing binary values 51C. The vector 510 has the same number of fields "z" as the vector 410 of the data structure 400 of Figure 4. However, in this case, each of the fields 512 include a binary value (e.g., represented by one (1) bit) rather than a more general value. The category vector 320 of the data structure 500 is the same as that of the data structures 300 and 400. An exemplary binarizing process is discussed in § 4.2.1.3 below.
Referring to Figure 5B, if the selected, non-zero, features 101 and 220 pass their threshold(s) but the selected, non-zero, feature 50 does not pass its threshold, a sparse data structure 510' includes a first array 550 which includes the indices (e.g., 101 and 220) of the selected , non-zero, threshold passing features, and a second array 560 which includes the values of the indexed features. As shown in Figure 5C, since the values of the binarized, selected, non-zero, threshold passing features will all be "1", an alternative sparse data structure 510" including only the index array 550 may be used.
Naturally other quantizers can be used to provide q-ary (where q>2) quantization instead of binary quantization.
The reduced feature vector (or binarized reduced feature vector) may be stored at storage unit 135. A general text classifier will include classifiers for each of the "m" different categories. Recall that a weight vector parameter and a monotonic (e.g., sigmoid) function, having adjustable parameters A and B, are to be determined for each of the "m" classifiers. Accordingly, "m" processes 140 (which may include a support vector machine training facility and a function generator) accept the reduced feature vector (or binarized reduced feature vector) from storage unit 135 or from the feature reduction process 120 (or binarizing process 150). Although "m" parallel processes 140 are depicted in Figure 1, a single process 140 could be used to process "m" successive categories (or p processes could be run m/p times) . In any event, a weight vector parameter and a monotonic (e.g., sigmoid) function are generated for each of the "m" categories. An exemplary process for generating parameters of a text classifier is described in § 4.2.1.4 below.
§ 4.2.1.1 FEATURE EXTRACTION
Recall from Figure 1 that a feature extraction process 110 generates a set of feature vectors (e.g., stored in data structures 300) from a set of training textual information objects.
Figure 6 is a flow diagram of an exemplary feature extraction process 110'. In step 610, a COUNT value, to be used as an index to textual information objects, is initialized by setting it to 1. Next, a textual information object, as indexed by the COUNT value, is accepted as shown in step 620. As shown in step 630 a FEATURE_COUNT value is initialized to 1 and a VALUE value, as indexed by the FEATURE_COUNT value, is initialized to 0. Next, as shown in step 640 the VALUE value, as indexed by the FEATURE_COUNT value, of the textual information object, as indexed by the COUNT value, is determined. For example, if the 5th feature is the word "sports", and the word "sports" appears three (3) times in the first textual information object, then the VALUE (5) of OBJECT (1) is three (3). As shown in steps 650 and 660, the values for each feature of a given object are determined. As shown in steps 670 and 680, once the values of the features of one object have been determined, the values of the features of the next object are then determined. When the values of all of the features of all of the objects have been determined, the process 110' terminates via return node 690.
To reiterate, step 640 determines a value of a feature of a textual information object. This determination may be performed as follows. A text analyzer may be used to parse each textual information object into constituent tokens. A token is any textual component, such as a word, letter, internal punctuation mark or the like, that is separated from another such component by a blank (white) space or leading (following) punctuation marks. Syntactic phrases and normalized representations for times and dates are also extracted by the text analysis module. Domain-specific features may also be defined and extracted as appropriate. Each of the features may define a word, phrase or letter grouping. Thus, a feature vector is associated with each of the textual information objects .
If a sparse data structure having an index array and a value array (see, e.g., Figure 3B) or having an array with alternating index fields and value fields (see, e.g., Figure 3c) is used, in step 640, if VALUE (FEATURE_COUNT) is non-zero, then (i) the next element of the index array (or next index field) is set to FEATURE_COUNT, and (ii) the next element of the value array (or next value field) is set to VALUE (FEATURE COUNT) .
§ 4.2.1.2 FEATURE REDUCTION (HIGH LEVEL)
As discussed above, a feature vector, associated with a textual information object, may contain millions of feature elements. To make training the textual classifier more feasible, the feature vector may be reduced to fewer elements. Figure 7 is a high level flow diagram of an exemplary feature reduction process 120' ("feature reduction" may also be referred to as "feature selection") which may be used in the present invention. First, as shown in step 710, the present invention may use a category independent feature reduction step (such as ZIPF or frequency) (known but discussed in § 4.2.1.2.1 below for the reader's convenience). Then, as shown in step 720, a category dependent feature reduction technique (some of which are discussed in § 4.2.1.2.2 below) may be performed. The process 120' then terminates via return node 730.
§ 4.2.1.2.1 CATEGORY INDEPENDENT FEATURE
REDUCTION
A matrix of feature vectors of a number of training example textual information objects may be reduced by applying collection-wide frequency statistics. Zipf's Law, which is known in the art, concerns the distribution of different words in text and basically states that, the product of a word's rank (r) and frequency (f) is a constant (C) i.e., (r*f=C) , as is the product of frequency and number of occurrences (N) (f*N=C) . That is, many words will occur only once in a collection; half that number will occur twice; a third that number will occur three times, and so on. It is assumed that words occurring only a small number of times will not be very helpful in classifying new items, and they are thus often omitted. At the other end of the spectrum, words occurring very often are not likely to provide good discrimination among categories and are sometimes omitted. By applying Zipf frequency thresholds, a reduced feature vector results for each of the textual information object training examples. The document-frequency thresholding proposed by the Yang-Pedersen article is similar to our Zipf selection. However, the method discussed in the Yang-Pedersen article uses different selection criteria, keeping only the most frequent terms.
§ 4.2.1.2.2 CATEGORY DEPENDENT FEATURE REDUCTION
After the feature vectors of the textual information objects of the training examples are reduced based on a category independent feature reduction technique, such as Zipf's law for example, they may be further reduced by a category dependent feature reduction technique. Such category dependent feature reduction may use known techniques such as the mutual information technique discussed in § 4.2.1.2.2.1 below, or the information gain technique discussed in § 4.2.1.2.2.2 below. Other known category dependent feature reduction techniques, such as chi-square, correlation coefficient, and SVD (all known and not discussed here) may be used instead. Category dependent feature reduction may also be carried out in accordance with a proprietary technique such as the Bayesian score technique discussed in § 4.2.1.2.2.3 below. §4.2.1.2.2.1 MUTUAL INFORMATION
Mutual information is a measure of the extent to which a feature is associated with a class, in this instance to a text category. Features that exhibit high mutual information measures are good discriminators of the class, and are therefore chosen. In an exemplary embodiment, the features may be words and the classes may be categories. In such a case, the mutual information "MI" for a given feature (fj.) , such as a word for example, in a given category c± can be expressed as:
Figure imgf000036_0001
where P(f+c+) is the probability that a document has feature f and belongs to category c;
P(f"c+) is the probability that a document does not have feature f, but belongs to category c; P(f+c~) is the probability that a document has feature f but does not belong to category c; and P(f~c") is the probability that a document does not have feature f and does not belong to category c;
Figure 9 depicts a two-way contingency table 900 of a feature "f" 950 and a category "c" 960. Each textual information object exhibits one (1) of the following four (4) characteristics: (i) has the feature "f" 950 and belongs to the category "c" 960 (f+c+ 910); (ii) has the feature "f" 950 but does not belong to the category "c" 960 (f+c~ 920); (iii) does not have the feature "f" 950 but belongs to the category "c" 960 (f"c+ 930); or (iv) does not have the feature "f" 950 and does not belong to the category "c" 960 (f~c~ 940) . If A is a number of textual information objects exhibiting characteristic f+c+ 910, B is a number of textual information objects exhibiting characteristic f+c" 920, C is a number of textual information objects exhibiting characteristic f"c+ 930, D is a number of textual information objects exhibiting characteristic f~c~ 940, and N is a total number of textual information objects, then mutual information (MI) may be expressed as:
Figure imgf000037_0001
Features (e.g., words) can be reduced to those features (words) for which the mutual information is the highest .
Given a training corpus, mutual information is computed for each unique term. Those terms whose mutual information value is less than some predetermined threshold may be removed from the feature vector. Alternatively, if the number of features is to be reduced to a predetermined number "x", then the features with the x highest mutual information values are kept while other features are not considered.
§4.2.1.2.2.2 INFORMATION GAIN
Information gain is a term selection criterion which measures the number of bits of information obtained for category prediction by knowing the presence or absence of a term in a document. For the case of a binary classification task (i.e., in category (c+) vs. not in category (c~) ) and binary feature values (i.e., feature present (f+) vs. feature absent (f~) ) , the information gain (G(f)) of a feature (f) is defined as:
G(f) = ~ ∑ (c')log2P(C')+ (r) ∑ (/V)log2 ( V) ιe{+,-} >e{+--}
+P(f-) ∑ (/V)log2 ( V)
where: P(c+) is the probability that a document belongs to category c;
P(c~) is the probability that a document does not belong to category c;
P(f+) is the probability that a document contains feature f;
P(f") is the probability that a document does not contain feature f;
Pff^c1) is the probability that a document has feature f and belongs to category c1 -- for binary classification i takes on values + and - ;
P(f~c1) is the probability that a document does not have feature f and belongs to category c1 — for binary classification i takes on values + and -;
This measure can be generalized to m-ary classification by summing over m possible categories.
Given a training corpus, information gain is computed for each unique feature. Those features whose information gain value is less than some predetermined threshold may be removed from the feature vector. Alternatively, if the number of features is to be reduced to a predetermined number "x", then the features with the x highest information gain values are kept while other features are not considered.
§4.2.1.2.2.3 BAYESIAN SCORE
The feature vector may be reduced based on a Bayesian Score (BS) . The determination of a Bayesian Score is described in patent application serial no. 08/240,019, entitled "Method for Improving Belief Networks" (incorporated herein by reference) . The Bayesian Score (BS) of the ith feature f may be defined as: BSf ι =igr(^+i) +igr(5+i)+igr(c+i)+igr(D+i)
-\%T(A + B + 2)-\gT{C + D + 2)-\zT{A + C + 2)-\gT{B + D + 2) (6)
+igr(N+ 4)
where T(n) = nr(n-l), for all n>0;
Figure imgf000040_0001
r(n+l) = n!, for n=0,l,2,...;
A is a number of textual information objects exhibiting characteristic f+c+ 910, B is a number of textual information objects exhibiting characteristic f+c~ 920, C is a number of textual information objects exhibiting characteristic f"c+ 930, D is a number of textual information objects - xhibiting characteristic f~c~ 940, and N is a total number of textual information objects
If the number of features is to be reduced to a predetermined number "x", then the features with the x highest Bayesian Scores are kept while the other features are not considered. A Bayesian Score threshold may also be used to reduce the number of features .
§ 4.2.1.3 BINARIZING THE REDUCED FEATURE VECTOR
The values held in a reduced feature vector may be applied to one threshold, or to thresholds corresponding to the features, to binarize the values of the reduced feature vector. Alternatively, the values in the feature vector may be applied directly to the classifier, or may be encoded as q-ary (where q>2) code through quantization, as is well known in the neural network art.
§ 4.2.1.4 CLASSIFICATION PARAMETER GENERATION
The foregoing steps (i.e., feature extraction, feature reduction, and binarization) may be thought of as pre-processing steps to permit parameters of a text classifier, such as an SVM text classifier, to be more easily determined. Basically, the classifier will classify a textual information object having a reduced feature vector x, based on the following:
Classification =f(w»x)
where f(z) a monotonic (e.g., sigmoid) function of z; and is a weight vector.
Thus, the classifier has a weight vector parameter, the determination of which is described in § 4.2.1.4.1 below, and a monotonic function having adjustable parameters, the determination of which is described in § 4.2.1.4.2 below. Thus, referring to the flow diagram of an exemplary process 140' for generating parameters for a classification engine depicted in Figure 8, the weight vector is determined, as shown in step 810, and the monotonic function (and its adjustable parameters) is determined, as shown in step 820. The process 140' is terminated via return node 830. § 4.2.1.4.1 WEIGHT VECTOR DETERMINATION
The classifier of each category has a weight vector parameter. For each category, the weight vector parameter may be generated by methods used to train a support vector machine. As discussed above, and as is known in the art, support vector machines (or "SVMs") may be trained with known objects having known classifications to define a hyperplane, or hypersurface, which separates points in n-dimensional feature vector space which are in the category and not in the category. A pair of adjustable parameters, (a "weight vector") and b (a "threshold") can be defined such that all of the training data x, having a known classification y, satisfy the following constraints:
xt » w+b≥ + \ for y, = + l (7)
xt * w+b≤-l for y, =-\ (8)
where: i = 1, . . ., number of training examples; x is the ith input vector; w is a weight vector parameter; b is a threshold parameter; and γ± is a known classification associated with the ith training example and is +1 if the example is "in the class" and -1 if the training example is "not in the class". The inequality conditions of equations (7) and (8) can be combined into the following inequality condition by multiplying each side of the equations by y and +1 or -1, and subtracting 1 from both sides:
>,(*,. •w+δ)-l>0 (9)
The points for which the equality portion of equation (7) hold lie on the hyperplane Xi dot w + b = 1, with normal w, and perpendicular distance to the origin (l-b)/| |w| |, where | |w| | is the Euclidean norm of the vector . Similarly, the points for which the equality portion of equation (8) hold lie on the hyperplane Xi dot + b = -1, with perpendicular distance to the origin (-1 - b) / | |w| |. Hence, the margin is the sum of both of the distances, namely 2/| I I I . By minimizing | | | |2, subject to the constraints of equation (9), the hyperplane providing the maximum margin can therefore be determined.
Thus, training an SVM presents a constrained optimization (e.g., minimization) problem. That is, a function is minimized (referred to as "the objective function") subject to one or more constraints . Although those skilled in the art are familiar with methods for solving constrained optimization problems (see, e.g., the text, Fletcher, Practical Methods of Optimization, 2nd ed., pp. 195-224, John Wiley & Sons (1987)), relevant methods will be introduced below for the reader's convenience. A point that satisfies all constraints is referred to as a "feasible point" and is located in a "feasible region". A constrained optimization (e.g., minimization) problem is solved by a feasible point having no feasible descent directions. Thus, methods for solving constrained minimization problems are often iterative so that a sequence of values converges to a local minimum.
Those skilled in the art recognize that
Lagrange multipliers provide a way to transform equality-constrained optimization problems into unconstrained extremization problems. Lagrange multipliers (α) may be used to find an extreme value of a function f(x) subject to a constraint g(x), such that 0 = V f(x) + α V g(x), where V is a gradient function. Thus, if f (w) = | | | |2 and g(w) = γ (Xj. dot + b) - 1, then the Lagrangian:
1 nte nte Zp=-||w|| -∑α,j/((x,.w+6)+∑α, (10)
^ ι=l 1=1
where "nte" is the number of training examples, results .
As is known from mathematical programming literature, the concept of "duality" allows the provision of an alternative formulation of a mathematical programming problem which is, for example, computationally convenient. Minimizing equation (10) subject to the constraint that "alpha" must be non-negative is referred to as a "primal" (or original) problem. The "dual" (i.e., the transformed) problem maximizes equation (10) subject to the constraints that the gradient of Lp with respect to w and b vanishes and that the "alpha" must be non-negative. This transformation is known as the "Wolfe dual". The dual constraints can be expressed as:
^=∑<W, (H)
∑a,y, = (12)
Substituting the conditions of equations (11) and (12) into equation (10) yields the following Lagrangian:
Figure imgf000045_0001
The solution to the dual quadratic programming problem can be determined by maximizing the Lagrangian dual problem LD.
The support vector machine can therefore be trained by solving the dual Lagrangian quadratic programming problem. There is a Lagrange multiplier x for each example i of a training set. The points for which is greater than zero are the "support vectors". These support vectors are the critical elements of the training set since they lie closest to the decision boundary and therefore define the margin from the decision boundary.
Unfortunately, the quadratic programming problem of equation (13) merely represents the optimization problem with constraints in a more manageable form than the problems of equations (7) and (8). Numerical methods, such as constrained conjugate gradient ascent, projection methods, Bunch-Kaufman decomposition, and interior points methods, may be used to solve the quadratic problem. Suffice to say that these numerical methods are not trivial, particularly when there are a large number of examples i in the training set. Referring to equation (13), solving the quadratic problem involves an "nte" by "nte" matrix (where "nte" is the number of examples in the training set) .
The quadratic programming problem may also be solved with a chunking algorithm. Vladimir Vapnik, who is credited with developing SVMs, introduced an iterative training method referred to as "chunking". Though the "chunking" method of training SVMs is understood by those skilled in the art, it will be introduced here for the reader's convenience. At each step, a subproblem consisting of a union of non-zero Lagrange multipliers from a previous step and a number of worst training examples that violate Kuhn-Tucker conditions is solved. These steps are repeated until all examples fulfill the Kuhn-Tucker conditions. In this way, instead of operating on a matrix the size of the square of the number of training examples, the matrix at each step is now on the order of the number of non-zero Lagrange multipliers, squared. Since the matrices are smaller, the algorithm can be faster. Although chunking makes training SVMs tractable, it is still too slow for some applications. For example, the number of non-zero Lagrange multipliers can still grow into the thousands in certain applications. The quadratic problem is still solved with a numerical quadratic programming technique and each iteration is time consuming.
A reportedly better training algorithm for
SVMs was introduced in the paper, E. Osuna et al . , "An Improved Training Algorithm for Support Vector Machines," Proceedings of IEEE NNSP ' 97 (September 1997) (hereafter referred to as "the Osuna method"). The Osuna method stores a fixed size matrix, rather than a matrix that grows with the number of non-zero Lagrange multipliers. A fixed number of Lagrange multipliers are optimized at every step. With each new step, one of the old Lagrange multipliers is removed from consideration and a new example whose output violates the Kuhn-Tucker conditions is added. The Osuna paper provides a proof that the quadratic problem will decrease with every step such that the Osuna method converges to a solution. However, the Osuna method still disadvantageously requires a numerical solution at each step. Although it converges to a solution, the Osuna method uses an entire numerical quadratic programming problem just to "fix" one Lagrange multiplier that violates the Kuhn-Tucker conditions.
A faster way to determine the weight vector is by using the method disclosed in U.S. Patent Application Serial No. 09/055,477, by John Platt, entitled "Methods and Apparatus for Building a Support Vector Machine Classifier", filed on April 6, 1998 and incorporated herein by reference. In a first disclosed methodology (also referred to as "QP1") of the Platt method, the threshold b is zero (0) . First, the Lagrange multipliers, corresponding to each of examples of the training set, are initialized. For example, all of the Lagrange multipliers may be initialized by setting them to zero (0) .
For each training example from the set of training examples, it is determined whether the solution is optimized for the example. Each example is examined for optimality by determining whether or not a Kuhn-Tucker condition is violated. If the solution is not optimal (e.g., if a Kuhn-Tucker condition is violated) , then a new Lagrange multiplier for the nth training example is determined. More specifically, the new Lagrange multiplier for the nth training example may be determined by the following expression: eeww old + K-. uhn - Tucker Error (14) k(χ„,χ„)
In the expression of equation (14), the Kuhn-Tucker (KT) Error may be expressed as:
1 - Onyn (15)
In the expression of equation (14), the denominator is the kernel function which, in the case of a linear support vector machine, is the dot product of the feature vectors x of the nth training example.
The new Lagrange multiplier is then limited (or clipped) as follows:
new.clψped
Figure imgf000049_0001
so that the inequality constraints are fulfilled.
Once the sweep through a set of training examples is complete, whether any of the Lagrange multipliers have changed during the sweep is determined. If so, then other examples may now violate the Kuhn-Tucker conditions. More specifically, as the Lagrange multipliers are adjusted, the separating hyperplane (or hypersurface) of the SVM changes. Hence, another sweep through the data is carried out. If no Lagrange multipliers were changed, then all the examples must fulfill the Kuhn-Tucker conditions and the QP1 process is terminated.
An initial sweep processes all examples of the training set. Assuming that a Lagrange multiplier was changed, the next sweep processes only the non-boundary examples of the training set. Subsequent sweeps also process only the non-boundary Lagrange multipliers until no Lagrange multipliers change. Then, the next sweep processes the entire set of training examples. If no Lagrange multipliers were changed, then processing ends. If, on the other hand, a Lagrange multiplier was changed, the processing continues as discussed above. Naturally, each sweep could process all of the training examples in the alternative .
In the case of linear SVMs, the entire SVM state may be stored as a weight vector rather than a weighted sum of kernels. In this case, the weight vector is directly updated when the Lagrange multiplier is changed. That is:
w=(a d y^ (17)
The Kuhn-Tucker conditions may be checked using interval arithmetic: namely, if the condition is approximately met within a certain tolerance, then condition is met. The tolerance may be set to le-3. This should be accurate enough for recognition machines, and will prevent the method from spending a large amount of computational resources generating unneeded precision in the Lagrange multipliers.
When the input points consist of sparse binary vectors (e.g., the object to be classified has relatively few of many possible features), the embodiment shown can be dramatically sped up by using a sparse evaluation of the dot product. Namely, the dot product between the weight and the input point is the sum of the weight elements that correspond to the non-zero input dimensions. This can be expressed as:
wx. = Wr (18) f≡allnon-zero features of vector xt
This optimization is useful for training linear SVMs to be used for text classification since most text objects are represented by sparse feature vectors (e.g., text objects have relatively few of all words of a feature vector) .
The alpha vector, which includes the values of the Lagrange multipliers for all training examples, may be stored as a vector of length equal to the size of the training set. That is:
α=α„ . . ., nte (19) wherein "nte" is the number of training examples .
Since, as stated above, most of the Lagrange multipliers will be zero (0), the alpha vector may be stored as two (2) vectors of length equal to the number of non-zero Lagrange multipliers, one storing the identity (index) of the non-zero Lagrange multiplier, the other storing the value of the non-zero Lagrange multiplier .
In a second disclosed methodology (also referred to as "QP2") of the Platt method, the quadratic programming process is subject to a linear equality constraint (i.e. the threshold b is not necessarily equal to zero (0)).
First, the Lagrange multipliers, corresponding to each of examples of the training set, are initialized. Recall that, since most of the Lagrange multipliers may be zero (0), all of the Lagrange multipliers may be initialized by setting them to zero (0) . For each training example from the set of training examples, it is determined whether a
Kuhn-Tucker condition is violated. If a Kuhn-Tucker condition is violated, another example is selected from the training set (as will be discussed below) , thereby creating a pair of Lagrange multipliers to be jointly optimized. An attempt is made to determine new
Lagrange multipliers such that the examples are jointly optimized. It is determined whether the examples were optimized. If they were not optimized, another feature vector is selected from the training set thereby creating another pair of Lagrange multipliers.
The first step in the joint optimization of the two (2) Lagrange multipliers is to determine the bounds on one of the variables. Either the first or the second multiplier can be bounded. Here, the second is chosen . Let L be the lower bound on the second Lagrange multiplier and H be the higher bound on the second Lagrange multiplier. Let yi be the desired output of the first example and let y2 be the desired output of the second example. Let cti be the current value of the first Lagrange multiplier and α2 be the current value of the second Lagrange multiplier. If yi is the same as y2, then the following bounds are computed:
H = mm.(C,ax +a2) L = max(0, ax + a2 -C) (20)
If yi is the opposite sign as y2, then the bounds are computed as follows:
H = m (C,C-al + 2) L = max(0, a2 - a ) (21)
If the value L is the same as the value H, then no progress can be made and it is determined that the training examples were not optimized. The new optimized value of α2 may be computed via the formula:
αnew = a +
Figure imgf000054_0001
If the new value of the second Lagrange multiplier is less than L, then it is set to L. Conversely, if the new value of the second Lagrange multiplier is greater than H, then it is set to H. If the new clipped (or limited) value of the second Lagrange multiplier is the same as the old value, then no optimization is possible it is determined that the training examples were not optimized. Otherwise, the new value of the first Lagrange multiplier is then derived from the clipped (or limited) value of the second Lagrange multiplier:
ar = al +yiy2 (a2 -ar ' ) (23)
If the support vector machine is linear, then the weights and thresholds are updated to reflect the new Lagrange multipliers so that other violations of the Kuhn Tucker conditions can be detected.
The second Lagrange multiplier pair may be selected based on the following heuristic. The ideal second Lagrange multiplier would change the most upon joint optimization. An easy approximation to the change upon optimization is the absolute value of the numerator in the change in the second Lagrange multiplier:
I yλy2KT ERROR, - KT ERROR, \ = | (0, - yλ)- (02 - y2) | (24)
If the true error (Oι~yι) of the first Lagrange multiplier is positive, then a second Lagrange multiplier that has a large negative true error (02-y2) would be a good candidate for joint optimization. If the first true error is negative, then a second
Lagrange multiplier that has a large positive true error would be a good candidate for optimization. Therefore, the QP2 process seeks out a non-boundary Lagrange multiplier ( ≠ 0 or C) that has a true error that is the most opposite of the true error of the first Lagrange multiplier.
There are some degenerate cases where different examples have the same input feature vectors. This could prevent the joint optimization from progressing. These redundant examples could be filtered out. Alternatively, a hierarchy of heuristics may be used to find a second example to make forward progress on the joint optimization step. If the first heuristic described above fails, the QP2 process will select a non-boundary Lagrange multiplier as the other Lagrange multiplier of the pair to be jointly optimized. If this heuristic fails for all non-boundary Lagrange multipliers, any of the other Lagrange multipliers may be selected as the other Lagrange multiplier of the pair.
Once the sweep through the set of training examples is complete, if one or more Lagrange multipliers changed with the sweep through the entire data set, the non-boundary Lagrange multipliers are then optimized. If, on the other hand, no Lagrange multipliers changed during the sweep, then all of the examples obey the Kuhn-Tucker conditions and the QP2 process is terminated.
As can been appreciated, the QP2 process first sweeps through all training examples of the training set. Assuming that a Lagrange multiplier was changed, the next sweep processes only the non-boundary training examples of the training set. Subsequent sweeps also process only the non-boundary Lagrange multipliers until no Lagrange multipliers change. Then, the next sweep processes the entire set of training examples. If no Lagrange multipliers change, processing ends. If, on the other hand, a Lagrange multiplier changes, the processing continues as discussed above. Naturally, in an alternative methodology, all training examples of the training set could be processed on every sweep.
The Lagrange multipliers can be stored as a either one array having a size corresponding to the number of training examples (also referred to as "the full array") or two (2) arrays that represent a larger sparse array.
The real error mentioned used when jointly optimizing a pair of Lagrange multipliers is the desired output of the SVM less the actual output of the SVM (yi - Oi) . These real errors may be cached for the examples that have a non-boundary (i.e., non-zero or a non-C) Lagrange multiplier. Caching the real errors allows an second example to be intelligently chosen.
For a linear SVM, the linear SVM may be stored as one weight vector, rather than as a linear superpositions of a set of input points. If the input vector is sparse, the update to the weight vector can be accomplished in a known sparse manner, also.
Thus, the weight vector w for a given category c can be determined by training an SVM, for example, based on one of the foregoing (e.g., Chunking, Osuna, and Platt) methods.
§ 4.2.1.4.2 MONOTONIC FUNCTION PARAMETER DETERMINATION
As discussed above, the text classifier employs a monotonic (e.g., sigmoid) function to classify textual information objects. A sigmoid function may be expressed in the form: /(«) = (25) l + e"
The characteristics of the sigmoid function may be adjusted using constants A and B (also referred to as "adjustable parameters") such that:
Figure imgf000058_0001
Techniques for solving for A and B of the monotonic function are described below.
§ 4.2.1.4.2.1 OPTIMIZATION
(MAXIMUM LIKELIHOOD)
The constants A and B may be determined by using a maximum likelihood on the set of training examples of textual information objects. In order to fit the sigmoid to the training data, the logarithm of the likelihood that the training targets y were generated by the probability distribution f, assuming independence between training examples, is computed. The log likelihood is described by:
nte ∑ y, log(/( ,))+ (l-^ )log(l-/(x,)) (28)
1=1
where γ is a known result (+1 or -1); and nte is the number of training examples. The optimal sigmoid parameters are then derived by maximizing the log likelihood over possible sigmoid parameters . § 4.2.1.4.2.1.1 KNOWN
OPTIMIZATION METHODS
Expression (28) may be maximized based on known unconstrained optimization techniques such as gradient descent, Levenberg-Marquardt, Newton-Raphson, conjugate gradient, and variable metric methods for example .
§ 4.2.1.4.2.1.2 LAPLACE ' S RULE OF SUCCESSION
In one embodiment, in expression (28) above, the YJ. values may be replaced by target values tx, which may be expressed as:
Figure imgf000059_0001
where N+ is the number of textual information objects of the training set in the class; and N- is the number of textual information objects of the training set not in the class. By using the target values tx rather than γl r the resulting sigmoid function is no more precise than the data used to determine it. That is, the sigmoid function is not "overfit" to past data, particularly in categories with little training data. In this way, a sigmoid function is determined that matches unknown data given a prior that all probabilities are equally likely. Although the use of the target function tx of equation (30) has been presented in the context of a text classifier, it is applicable to other classifiers as well. Naturally, other target functions, which depend on the number of training examples in the positive and negative classes, could be used instead of the target function defined in equation (29) .
In one embodiment of fitting a sigmoid to the training data, the target values (ttx) expressed as:
Figure imgf000060_0001
are used. This version of the target function limits the effect of Laplace' s rule of succession to be no farther away than 0.01 from the true target y. This has been empirically found to be desirable.
§ 4.2.2 EXEMPLARY STRUCTURE OF TEXT CLASSIFIER
Figure 2 is a diagram of processes that may be used by a text classifier based on the present invention. As shown in Figure 2, a textual information object 205 to be classified is applied to a feature extraction process (or more generally, a feature extractor) 210. As described in § 4.2.2.1 below, this feature extraction process 210 may be the same as the feature extraction process 110 discussed in § 4.2.1.1 above. Then, a feature reduction process (or more generally, a feature reduction facility) 220, described in § 4.2.2.2 below, using the set of reduced features generated by the feature reduction process 120 discussed in § 4.2.1.2 above, reduces the size of the feature vector associated with the textual information object 205. Thus, a reduced feature vector associated with the textual information object 205 is generated. As shown in Figure 2, a binarizing process 230 may then be applied to the reduced feature vector.
The reduced feature vector generated by the feature reduction process 220 (or the binarized reduced feature vector generated by the binarizing process 230) is applied to one or more classifiers 250. These classifiers 250 may use a classification process described in § 4.2.2.4 below. The weight parameter w and the parameters of the parameterizable monotonic function for each classifier 250 are those determined by the classification engine parameter generation process(es) 140 discussed in § 4.2.1.4 above. A probability that the textual information object 205 belongs to a particular class is output. § 4.2.2.1 FEATURE EXTRACTION
The feature extraction process 210 may be analogous to the feature extraction process 110 discussed in § 4.2.1.1 above with reference to Figures 1 and 6.
§ 4.2.2.2 FEATURE REDUCTION
An exemplary feature reduction process 220 is now described. Recall that the feature reduction process 120 discussed above reduces (e.g., based on category independent and category dependent techniques) the feature vectors corresponding to the textual information objects of the set of training examples 115. This process 120 also produces a reduced feature set, which may be an index to features deemed relevant or important for the text classification. For example, if the original feature vector included 100 elements, the reduced feature set may include an index to the 2nd, 5th, 20th, 21st, and 89th elements of the feature of the vector. Thus, the 100 element feature vector is reduced to a five (5) element reduced feature vector. Typically, the number of elements in the feature vector and reduced feature vector will be much larger than those used in this illustration.
The exemplary feature reduction process 220 uses the index of relevant or important features to map the values of the features to sequential indices of the reduced feature vector. Following the foregoing simplified example, the value of the 2nd element would be placed in the first element of the reduced feature vector, the value of the 5th element would be placed in the second element of the reduced feature vector, the value of the 20th element would be placed in the third element of the reduced feature vector, the value of the 21st element would be placed in the fourth element of the reduced feature vector and the value of the 89th element would be placed in the fifth element of the reduced feature vector. As discussed in § 4.2.1 above, in practice, a textual information object is not represented by a single feature vector array, but rather, using sparse techniques known to those skilled in the art, it is represented by smaller feature index and value arrays.
Note that the reduced feature set may be provided to the feature extraction process 210. In this case, the feature extraction process only extracts those features belonging to the reduced feature set. In this way, a feature reduction is carried out while the features are being extracted.
§ 4.2.2.3 BINARIZATION
The optional binarizing process 230 may be analogous to the binarizing process 150 discussed in § 4.2.1.3 above with reference to Figure 1. § 4.2.2.4 CLASSIFICATION
The classification processes 250 may be effected as follows. The probability that the textual information object 205 (as represented by reduced feature vector x) is in category i can be expressed as:
Figure imgf000064_0001
This text classifier has a number of advantages. First, the probabilities output by the classifiers of the various categories can be compared across categories. Second, if the target technique, described in § 4.2.1.4.2.1.2 above is used for determining the monotonic function and its adjustable parameters, that function will not be overfit when relatively small amounts of training data are available for a particular category.
§4.2.3 EXEMPLARY ARCHITECTURE
Figure 11A and the following discussion provide a brief, general description of an exemplary apparatus in which at least some aspects of the present invention may be implemented. The present invention will be described in the general context of computer-executable instructions, such as program modules, being executed by a personal computer. However, the methods of the present invention may be effected by other apparatus. Program modules may include routines, programs, objects, components, data structures, etc. that perform a task(s) or implement particular abstract data types. Moreover, those skilled in the art will appreciate that at lease some aspects of the present invention may be practiced with other configurations, including hand-held devices, multiprocessor systems, microprocessor-based or programmable consumer electronics, network computers, minicomputers, set top boxes, mainframe computers, and the like. At least some aspects of the present invention may also be practiced in distributed computing environments where tasks are performed by remote processing devices linked through a communications network. In a distributed computing environment, program modules may be located in local and/or remote memory storage devices.
With reference to Figure 11A, an exemplary apparatus 1100 for implementing at least some aspects of the present invention includes a general purpose computing device in the form of a conventional personal computer 1120. The personal computer 1120 may include a processing unit 1121, a system memory 1122, and a system bus 1123 that couples various system components including the system memory to the processing unit 1121. The system bus 1123 may be any of several types of bus structures including a memory bus or memory controller, a peripheral bus, and a local bus using any of a variety of bus architectures. The system memory may include read only memory (ROM) 1124 and/or random access memory (RAM) 1125. A basic input/output system 1126 (BIOS) , containing basic routines that help to transfer information between elements within the personal computer 1120, such as during start-up, may be stored in ROM 1124. The personal computer 1120 may also include a hard disk drive 1127 for reading from and writing to a hard disk, (not shown) , a magnetic disk drive 1128 for reading from or writing to a (e.g., removable) magnetic disk 1129, and an optical disk drive 1130 for reading from or writing to a removable (magneto) optical disk 1131 such as a compact disk or other (magneto) optical media. The hard disk drive 1127, magnetic disk drive 1128, and (magneto) optical disk drive 1130 may be coupled with the system bus 1123 by a hard disk drive interface 1132, a magnetic disk drive interface 1133, and a (magneto) optical drive interface 1134, respectively. The drives and their associated storage media provide nonvolatile storage of machine readable instructions, data structures, program modules and other data for the personal computer 1120. Although the exemplary environment described herein employs a hard disk, a removable magnetic disk 1129 and a removable optical disk 1131, those skilled in the art will appreciate that other types of storage media, such as magnetic cassettes, flash memory cards, digital video disks, Bernoulli cartridges, random access memories (RAMs) , read only memories (ROM) , and the like, may be used instead of, or in addition to, the storage devices introduced above.
A number of program modules may be stored on the hard disk 1123, magnetic disk 1129, (magneto) optical disk 1131, ROM 1124 or RAM 1125, such as an operating system 1135, one or more application programs 1136, other program modules 1137, and/or program data 1138 for example. A user may enter commands and information into the personal computer 1120 through input devices, such as a keyboard 1140 and pointing device 1142 for example. Other input devices (not shown) such as a microphone, joystick, game pad, satellite dish, scanner, or the like may also be included. These and other input devices are often connected to the processing unit 1121 through a serial port interface 1146 coupled to the system bus. However, input devices may be connected by other interfaces, such as a parallel port, a game port or a universal serial bus (USB) . A monitor 1147 or other type of display device may also be connected to the system bus 1123 via an interface, such as a video adapter 1148 for example. In addition to the monitor, the personal computer 1120 may include other peripheral output devices (not shown) , such as speakers and printers for example.
The personal computer 1120 may operate in a networked environment which defines logical connections to one or more remote computers, such as a remote computer 1149. The remote computer 1149 may be another personal computer, a server, a router, a network PC, a peer device or other common network node, and may include many or all of the elements described above relative to the personal computer 1120, although only a memory storage device 1150 has been illustrated in Figure 11A. The logical connections depicted in Figure 11A include a local area network (LAN) 1151 and a wide area network (WAN) 1152, an intranet and the Internet.
When used in a LAN, the personal computer 1120 may be connected to the LAN 1151 through a network interface adapter (or "NIC") 1153. When used in a WAN, such as the Internet, the personal computer 1120 may include a modem 1154 or other means for establishing communications over the wide area network 1152. The modem 1154, which may be internal or external, may be connected to the system bus 1123 via the serial port interface 1146. In a networked environment, at least some of the program modules depicted relative to the personal computer 1120 may be stored in the remote memory storage device. The network connections shown are exemplary and other means of establishing a communications link between the computers may be used.
Figure 11B is a more general machine 1100 which may effect one or more of the processes discussed above. The machine 1100 basically includes a processor(s) 1102, an input/output interface unit(s) 1104, a storage device (s) 1106, and a system bus or network 1108 for facilitating data and control communications among the coupled elements . The processor (s) 1102 may execute machine-executable instructions to effect one or more aspects of the present invention. At least a portion of the machine executable instructions may be stored (temporarily or more permanently) on the storage devices (1106) and/or may be received from an external source via an input interface unit 1104.
§ 4.2.4 CONCLUSIONS
Thus, methods and apparatus for classifying text and building a text classifier have been disclosed. The methods and apparatus advantageously: (i) reduce the size of a feature vector representing a textual information object while maintaining the features important for classification; (ii) avoid generating a classifier overfit to relatively low amounts of training data; (iii) train relatively fast; and/or (iv) classify relatively fast.

Claims

WHAT IS CLAIMED IS:
1. A method for generating, from a training set of textual information objects, each either belonging to a category or not, parameters of a classifier for determining whether or not a textual information object belongs to the category, the method comprising steps of: a) extracting features from the textual information objects to generate, for each textual information object of the training set, an associated feature vector having a number of features; b) reducing the number of features of each of the feature vectors to generated, for each textual information object of the training set, an associated reduced feature vector; c) determining a weight vector for the classifier by training a support vector machine; and d) determining a monotonic function to be used by the classifier.
2. The method of claim 1 wherein the step of reducing the number of features includes sub-steps of: i) reducing the number of features based on Zipf's Law; and ii) further reducing the number of features based on at least one criteria selected from a group of criteria consisting of (a) mutual information, (b) information gain, (c) Bayesian score, (d) chi-square, (e) correlation coefficient, and (f) singular value decomposition.
3. The method of claim 1 wherein the step of determining a monotonic function to be used by the classifier includes a step of determining parameters of the monotonic function by maximizing a likelihood of the training set of textual information objects given an output of the monotonic function.
4. The method of claim 3 wherein the step of determining parameters of the monotonic function by maximizing a likelihood of the training set of textual information objects given an output of the monotonic function uses a target value ti defined as:
Figure imgf000071_0001
where Ν+ is a number of textual information objects of the training set in the category and N- is a number of textual information objects not in the category.
5. The method of claim 3 wherein the step of determining parameters of the monotonic function by maximizing a likelihood of the training set of textual information objects given an output of the monotonic function uses a target value derived from a number of textual information objects of the training set in the category and a number of textual information objects of the training set not in the category.
6. A method for generating, from a set of reduced feature vectors associated with a training set of textual information objects, each either belonging to a category or not, parameters of a classifier for determining whether or not a textual information object belongs to the category, the method comprising steps of: a) determining a weight vector for the classifier by training a support vector machine; and b) determining a monotonic function to be used by the classifier.
7. The method of claim 6 wherein the step of determining a monotonic function to be used by the classifier includes a step of determining parameters of the monotonic function by maximizing a likelihood of the training set of textual information objects given an output of the monotonic function.
8. The method of claim 7 wherein the step of determining parameters of the monotonic function by maximizing a likelihood of the training set of textual information objects given an output of the monotonic function uses a target value ti defined as: ╬╣1
Figure imgf000073_0001
where Ν+ is a number of textual information objects of the training set in the category and N_ is a number of textual information objects not in the category.
9. The method of claim 7 wherein the step of determining parameters of the monotonic function by maximizing a likelihood of the training set of textual information objects given an output of the monotonic function uses a target value derived from a number of textual information objects of the training set in the category and a number of textual information objects of the training set not in the category.
10. A method for generating, from a set of feature vectors associated with a training set of textual information objects, each either belonging to a category or not, parameters of a classifier for determining whether or not a textual information object belongs to the category, the method comprising steps of: a) reducing the number of features of each of the feature vectors to generated, for each textual information object of the training set, an associated reduced feature vector; b) determining a weight vector for the classifier by training support vector machine; and c) determining a monotonic function to be used by the classifier.
11. The method of claim 10 wherein the step of reducing the number of features includes sub-steps of: i) reducing the number of features based on Zipf ' s Law; and ii) further reducing the number of features based on at least one criteria selected from a group of criteria consisting of (a) mutual information, (b) information gain, (c) Bayesian score, (d) chi-square, (e) correlation coefficient, and (f) singular value decomposition.
12. The method of claim 10 wherein the step of determining a monotonic function to be used by the classifier includes a step of determining parameters of the monotonic function by maximizing a likelihood of the training set of textual information objects given an output of the monotonic function.
13. The method of claim 12 wherein the step of determining parameters of the monotonic function by maximizing a likelihood of the training set of textual information objects given an output of the monotonic function uses a target value ti defined as:
Figure imgf000075_0001
where Ν+ is a number of textual information objects of the training set in the category and N- is a number of textual information objects not in the category.
14. The method of claim 12 wherein the step of determining parameters of the monotonic function by maximizing a likelihood of the training set of textual information objects given an output of the monotonic function uses a target value derived from a number of textual information objects of the training set in the category and a number of textual information objects of the training set not in the category.
15. An apparatus for generating, from a training set of textual information objects, each either belonging to a category or not, parameters of a classifier for determining whether or not a textual information object belongs to the category, the apparatus comprising: a) a feature extractor for extracting features from the textual information objects to generate, for each textual information object of the training set, an associated feature vector having a number of features; b) a feature reduction facility for reducing the number of features of each of the feature vectors to generated, for each textual information object of the training set, an associated reduced feature vector; c) a support vector machine training facility for determining a weight vector for the classifier; and d) a function generator for determining a monotonic function to be used by the classifier.
16. The apparatus of claim 17 wherein the function generator includes means for determining parameters of the monotonic function by maximizing a likelihood of the training set of textual information objects given an output of the monotonic function.
17. The apparatus of claim 16 wherein the means for determining parameters of the monotonic function by maximizing a likelihood of the training set of textual information objects given an output of the monotonic function use a target value tx defined as:
il
Figure imgf000076_0001
where N+ is a number of textual information objects of the training set in the category and N_ is a number of textual information objects not in the category.
18. The apparatus of claim 16 wherein the means for determining parameters of the monotonic function by maximizing a likelihood of the training set of textual information objects given an output of the monotonic function use a target value derived from a number of textual information objects of the training set in the category and a number of textual information objects of the training set not in the category.
19. An apparatus for generating, from a set of reduced feature vectors associated with a training set of textual information objects, each either belonging to a category or not, parameters of a classifier for determining whether or not a textual information object belongs to the category, the apparatus comprising: a) a support vector machine training facility for determining a weight vector for the classifier; and b) a function generator for determining a monotonic function to be used by the classifier.
20. The apparatus of claim 19 wherein the function generator includes means for determining parameters of the monotonic function by maximizing a likelihood of the training set of textual information objects given an output of the monotonic function.
21. The apparatus of claim 20 wherein the means for determining parameters of the monotonic function by maximizing a likelihood of the training set of textual information objects given an output of the monotonic function uses a target value ti defined as:
Figure imgf000078_0001
where Ν+ is a number of textual information objects of the training set in the category and N- is a number of textual information objects not in the category.
22. The apparatus of claim 20 wherein the step of determining parameters of the monotonic function by maximizing a likelihood of the training set of textual information objects given an output of the monotonic function uses a target value derived from a number of textual information objects of the training set in the category and a number of textual information objects of the training set not in the category.
23. An apparatus for generating, from a set of feature vectors associated with a training set of textual information objects, each either belonging to a category or not, parameters of a classifier for determining whether or not a textual information object belongs to the category, the apparatus comprising: a) a feature reduction facility for reducing the number of features of each of the feature vectors to generated, for each textual information object of the training set, an associated reduced feature vector; b) a support vector machine training facility for determining a weight vector for the classifier; and c) a function generator for determining a monotonic function to be used by the classifier.
24. The apparatus of claim 23 wherein the function generator includes means for determining parameters of the monotonic function by maximizing a likelihood of the training set of textual information objects given an output of the monotonic function.
25. The apparatus of claim 24 wherein the means for determining parameters of the monotonic function by maximizing a likelihood of the training set of textual information objects given an output of the monotonic function uses a target value ti defined as:
Figure imgf000080_0001
where Ν+ is a number of textual information objects of the training set in the category and N- is a number of textual information objects not in the category.
26. The apparatus of claim 24 wherein the means for determining parameters of the monotonic function by maximizing a likelihood of the training set of textual information objects given an output of the monotonic function uses a target value derived from a number of textual information objects of the training set in the category and a number of textual information objects of the training set not in the category.
27. A method for classifying a textual information object as belonging to or not belonging to a category, the method comprising steps of: a) extracting features from the textual information object to generate a feature vector associated with the textual information object; b) reducing features of the feature vector to generate a reduced feature vector; and c) applying the reduced feature vector and a weight vector to a classification function defined as: f(j) = l + e A(w┬╗x)+B
28. The method of claim 27 wherein the weight vector is determined from a training set of textual information objects having a predetermined classification applied to a support vector machine training facility.
29. A method for classifying a reduced feature vector, associated with a textual information object, as belonging to or not belonging to a category, the method comprising steps of: a) accepting a weight vector parameter; b) accepting a classification function defined as;
f(J) = ^ + eA(ή.i)+B
and c) applying the reduced feature vector and the accepted weight vector parameter to the accepted classification function.
30. The method of claim 29 wherein the weight vector is determined from a training set of textual information objects having a predetermined classification applied to a support vector machine training facility.
31. A method for classifying a feature vector, associated with a textual information object, as belonging to or not belonging to a category, the method comprising steps of: a) reducing features of the feature vector to generate a reduced feature vector; and b) applying the reduced feature vector and a weight vector to a classification function defined as:
f(i)= A(1w'x)+B l+e
32. The method of claim 31 wherein the weight vector is determined from a training set of textual information objects having a predetermined classification applied to a support vector machine training facility.
33. An apparatus for classifying a textual information object as belonging to or not belonging to a category, the apparatus comprising: a) a feature extractor for extracting features from the textual information object to generate a feature vector associated with the textual information object; b) a feature reduction facility for reducing features of the feature vector to generate a reduced feature vector; and c) a classifier for applying the reduced feature vector and a weight vector to a classification function defined as:
f(i)= A( lwx)+B l+e
34. The apparatus of claim 33 wherein the weight vector is determined from a training set of textual information objects having a predetermined classification applied to a support vector machine training facility.
35. An apparatus for classifying a reduced feature vector, associated with a textual information object, as belonging to or not belonging to a category, the apparatus comprising: a) a first input for accepting a weight vector parameter; b) a second input for accepting a classification function ; and c) a classifier for applying the reduced feature vector and the accepted weight vector parameter to the accepted classification function defined as:
/(i)= l+e A( lw┬╗x)+B
36. The apparatus of claim 35 wherein the weight vector is determined from a training set of textual information objects having a predetermined classification applied to a support vector machine training facility.
37. An apparatus for classifying a feature vector, associated with a textual information object, as belonging to or not belonging to a category, the apparatus comprising: a) a feature reduction facility for reducing features of the feature vector to generate a reduced feature vector; and b) a classifier for applying the reduced feature vector and a weight vector to a classification function defined as:
f(i)= A(*w┬╗x)+B l+e
38. The apparatus of claim 37 wherein the weight vector is determined from a training set of textual information objects having a predetermined classification applied to a support vector machine training facility.
39. A method for generating, from a training set of objects, each object either belonging to a category or not, parameters of a sigmoid function of a classifier for determining whether or not an object belongs to the category, the method comprising steps of: a) determining a target value based on a number of objects of the training set in the category and a number of objects of the training set not in the category; and b) determining the parameters of the sigmoid function by maximizing a likelihood of the training set of textual information objects given an output of the monotonic function using the determined target value.
40. The method of claim 39 wherein the target value ti is defined as:
Figure imgf000085_0001
where Ν+ is a number of textual information objects of the training set in the category and N- is a number of textual information objects not in the category.
41. The method of claim 39 wherein the target is limited to a maximum of the target value (ti) and 0.99 if the object belongs to the category and the minimum of the target value (ti) and 0.01 if the object does not belong to the category.
42. An apparatus for generating, from a training set of objects, each object either belonging to a category or not, parameters of a sigmoid function of a classifier for determining whether or not an object belongs to the category, the apparatus comprising: a) a target value generator for determining a target value based on a number of objects of the training set in the category and a number of objects of the training set not in the category; and b) means for determining parameters of the sigmoid functions by maximizing a likelihood of the training set of objects using the determined target value.
43. The apparatus of claim 42 wherein the target value tx is defined as:
Figure imgf000086_0001
where Ν+ is a number of textual information objects of the training set in the category and N- is a number of textual information objects not in the category.
44. The apparatus of claim 43 wherein the target value is limited to a maximum of the target value (tx) and
0.99 if the object belongs to the category and the minimum of the target value (ti) and 0.01 if the object does not belong to the category.
PCT/US1999/014086 1998-06-23 1999-06-22 Methods and apparatus for classifying text and for building a text classifier WO1999067728A1 (en)

Priority Applications (2)

Application Number Priority Date Filing Date Title
EP99933534.2A EP1090365B1 (en) 1998-06-23 1999-06-22 Methods and apparatus for classifying text and for building a text classifier
JP2000556321A JP4514082B2 (en) 1998-06-23 1999-06-22 Method and apparatus for building a text classifier and text classifier

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US09/102,946 1998-06-23
US09/102,946 US6192360B1 (en) 1998-06-23 1998-06-23 Methods and apparatus for classifying text and for building a text classifier

Publications (1)

Publication Number Publication Date
WO1999067728A1 true WO1999067728A1 (en) 1999-12-29

Family

ID=22292549

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US1999/014086 WO1999067728A1 (en) 1998-06-23 1999-06-22 Methods and apparatus for classifying text and for building a text classifier

Country Status (5)

Country Link
US (1) US6192360B1 (en)
EP (1) EP1090365B1 (en)
JP (1) JP4514082B2 (en)
CN (1) CN1307579C (en)
WO (1) WO1999067728A1 (en)

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2000019335A1 (en) * 1998-09-30 2000-04-06 Siemens Aktiengesellschaft Method and device for assigning an object to at least one class
EP1154358A2 (en) * 2000-05-12 2001-11-14 Applied Psychology Research Limited Automatic text classification system
WO2001090921A2 (en) * 2000-05-25 2001-11-29 Kanisa, Inc. System and method for automatically classifying text
WO2002010957A2 (en) * 2000-07-31 2002-02-07 Eliyon Technologies Corporation Computer method and apparatus for determining content types of web pages
JP2003535407A (en) * 2000-05-31 2003-11-25 レキシス ネクシス Computer-based system and method for finding the rule of law in text
EP1430378A2 (en) * 2001-05-08 2004-06-23 Sunflare Co. Ltd. New differential lsi space-based probabilistic document classifier
EP1480167A1 (en) * 2002-02-27 2004-11-24 NEC Corporation Pattern feature selection method, classification method, judgment method, program, and device
US6868411B2 (en) 2001-08-13 2005-03-15 Xerox Corporation Fuzzy text categorizer
WO2009132188A1 (en) * 2008-04-24 2009-10-29 Boston Scientific Scimed, Inc. Methods, systems, and devices for tissue characterization by spectral similarity of intravascular ultrasound signals
US9549713B2 (en) 2008-04-24 2017-01-24 Boston Scientific Scimed, Inc. Methods, systems, and devices for tissue characterization and quantification using intravascular ultrasound signals
CN107193836A (en) * 2016-03-15 2017-09-22 腾讯科技(深圳)有限公司 A kind of recognition methods and device
US10456105B2 (en) 2015-05-05 2019-10-29 Boston Scientific Scimed, Inc. Systems and methods with a swellable material disposed over a transducer of an ultrasound imaging system

Families Citing this family (326)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
AU3969093A (en) 1992-04-30 1993-11-29 Apple Computer, Inc. Method and apparatus for organizing information in a computer system
US7475048B2 (en) * 1998-05-01 2009-01-06 Health Discovery Corporation Pre-processed feature ranking for a support vector machine
US7805388B2 (en) * 1998-05-01 2010-09-28 Health Discovery Corporation Method for feature selection in a support vector machine using feature ranking
WO2003040949A1 (en) * 2001-11-07 2003-05-15 Biowulf Technologies, Llc Pre-processed Feature Ranking for a support Vector Machine
JP2000029902A (en) * 1998-07-15 2000-01-28 Nec Corp Structure document classifying device and recording medium where program actualizing same structured document classifying device by computer is recorded, and structured document retrieval system and recording medium where program actualizing same structured document retrieval system by computer is recorded
US6405188B1 (en) * 1998-07-31 2002-06-11 Genuity Inc. Information retrieval system
JP2000285140A (en) * 1998-12-24 2000-10-13 Ricoh Co Ltd Device and method for processing document, device and method for classifying document, and computer readable recording medium recorded with program for allowing computer to execute these methods
US6654787B1 (en) 1998-12-31 2003-11-25 Brightmail, Incorporated Method and apparatus for filtering e-mail
US6622134B1 (en) * 1999-01-05 2003-09-16 International Business Machines Corporation Method of constructing data classifiers and classifiers constructed according to the method
US8095581B2 (en) * 1999-02-05 2012-01-10 Gregory A Stobbs Computer-implemented patent portfolio analysis method and apparatus
US6477551B1 (en) * 1999-02-16 2002-11-05 International Business Machines Corporation Interactive electronic messaging system
US6510406B1 (en) * 1999-03-23 2003-01-21 Mathsoft, Inc. Inverse inference engine for high performance web search
US6862710B1 (en) 1999-03-23 2005-03-01 Insightful Corporation Internet navigation using soft hyperlinks
EP1049030A1 (en) * 1999-04-28 2000-11-02 SER Systeme AG Produkte und Anwendungen der Datenverarbeitung Classification method and apparatus
US6502082B1 (en) * 1999-06-01 2002-12-31 Microsoft Corp Modality fusion for object tracking with training system and method
FI991490A0 (en) * 1999-06-30 1999-06-30 Bayes Information Technology O visualization method
US6622160B1 (en) * 1999-07-30 2003-09-16 Microsoft Corporation Methods for routing items for communications based on a measure of criticality
US7194681B1 (en) * 1999-07-30 2007-03-20 Microsoft Corporation Method for automatically assigning priorities to documents and messages
US6714967B1 (en) * 1999-07-30 2004-03-30 Microsoft Corporation Integration of a computer-based message priority system with mobile electronic devices
US6539391B1 (en) * 1999-08-13 2003-03-25 At&T Corp. Method and system for squashing a large data set
US6816847B1 (en) * 1999-09-23 2004-11-09 Microsoft Corporation computerized aesthetic judgment of images
US6772149B1 (en) * 1999-09-23 2004-08-03 Lexis-Nexis Group System and method for identifying facts and legal discussion in court case law documents
GB2354609B (en) * 1999-09-25 2003-07-16 Ibm Method and system for predicting transactions
US7970718B2 (en) * 2001-05-18 2011-06-28 Health Discovery Corporation Method for feature selection and for evaluating features identified as significant for classifying data
US8041632B1 (en) * 1999-10-28 2011-10-18 Citibank, N.A. Method and system for using a Bayesian belief network to ensure data integrity
US6728690B1 (en) * 1999-11-23 2004-04-27 Microsoft Corporation Classification system trainer employing maximum margin back-propagation with probabilistic outputs
US8271316B2 (en) * 1999-12-17 2012-09-18 Buzzmetrics Ltd Consumer to business data capturing system
US6560582B1 (en) * 2000-01-05 2003-05-06 The United States Of America As Represented By The Secretary Of The Navy Dynamic memory processor
US6697769B1 (en) * 2000-01-21 2004-02-24 Microsoft Corporation Method and apparatus for fast machine training
EP1128278B1 (en) * 2000-02-23 2003-09-17 SER Solutions, Inc Method and apparatus for processing electronic documents
US6757646B2 (en) * 2000-03-22 2004-06-29 Insightful Corporation Extended functionality for an inverse inference engine based web search
US6532467B1 (en) * 2000-04-10 2003-03-11 Sas Institute Inc. Method for selecting node variables in a binary decision tree structure
AUPQ684400A0 (en) * 2000-04-11 2000-05-11 Telstra R & D Management Pty Ltd A gradient based training method for a support vector machine
US7912868B2 (en) * 2000-05-02 2011-03-22 Textwise Llc Advertisement placement method and system using semantic analysis
US20020099730A1 (en) * 2000-05-12 2002-07-25 Applied Psychology Research Limited Automatic text classification system
US6438519B1 (en) * 2000-05-31 2002-08-20 Motorola, Inc. Apparatus and method for rejecting out-of-class inputs for pattern classification
US20040073617A1 (en) 2000-06-19 2004-04-15 Milliken Walter Clark Hash-based systems and methods for detecting and preventing transmission of unwanted e-mail
US6785683B1 (en) * 2000-07-06 2004-08-31 International Business Machines Corporation Categorization and presentation tool for code resources
EP1182577A1 (en) * 2000-08-18 2002-02-27 SER Systeme AG Produkte und Anwendungen der Datenverarbeitung Associative memory
US9177828B2 (en) 2011-02-10 2015-11-03 Micron Technology, Inc. External gettering method and device
US7197470B1 (en) * 2000-10-11 2007-03-27 Buzzmetrics, Ltd. System and method for collection analysis of electronic discussion methods
US7185065B1 (en) * 2000-10-11 2007-02-27 Buzzmetrics Ltd System and method for scoring electronic messages
AU2000278962A1 (en) * 2000-10-19 2002-04-29 Copernic.Com Text extraction method for html pages
US6640228B1 (en) 2000-11-10 2003-10-28 Verizon Laboratories Inc. Method for detecting incorrectly categorized data
US6778941B1 (en) 2000-11-14 2004-08-17 Qualia Computing, Inc. Message and user attributes in a message filtering method and system
US6798912B2 (en) * 2000-12-18 2004-09-28 Koninklijke Philips Electronics N.V. Apparatus and method of program classification based on syntax of transcript information
US6751628B2 (en) * 2001-01-11 2004-06-15 Dolphin Search Process and system for sparse vector and matrix representation of document indexing and retrieval
US6721737B2 (en) * 2001-04-04 2004-04-13 International Business Machines Corporation Method of ranking items using efficient queries
US7593920B2 (en) * 2001-04-04 2009-09-22 West Services, Inc. System, method, and software for identifying historically related legal opinions
US6823323B2 (en) * 2001-04-26 2004-11-23 Hewlett-Packard Development Company, L.P. Automatic classification method and apparatus
US20020174132A1 (en) * 2001-05-04 2002-11-21 Allresearch, Inc. Method and system for detecting unauthorized trademark use on the internet
US6925433B2 (en) * 2001-05-09 2005-08-02 International Business Machines Corporation System and method for context-dependent probabilistic modeling of words and documents
US7269546B2 (en) * 2001-05-09 2007-09-11 International Business Machines Corporation System and method of finding documents related to other documents and of finding related words in response to a query to refine a search
US6768815B2 (en) 2001-05-10 2004-07-27 The United States Of America As Represented By The Secretary Of The Navy Color sensor
FR2825168A1 (en) * 2001-05-23 2002-11-29 France Telecom METHOD FOR DISCRECING ATTRIBUTES OF A DATABASE
US6898737B2 (en) 2001-05-24 2005-05-24 Microsoft Corporation Automatic classification of event data
US7076527B2 (en) * 2001-06-14 2006-07-11 Apple Computer, Inc. Method and apparatus for filtering email
US7298903B2 (en) * 2001-06-28 2007-11-20 Microsoft Corporation Method and system for separating text and drawings in digital ink
US7043492B1 (en) 2001-07-05 2006-05-09 Requisite Technology, Inc. Automated classification of items using classification mappings
US20030033263A1 (en) * 2001-07-31 2003-02-13 Reel Two Limited Automated learning system
US7874841B1 (en) 2001-08-08 2011-01-25 Lycas Geoffrey S Method and apparatus for personal awareness and growth
US6609124B2 (en) 2001-08-13 2003-08-19 International Business Machines Corporation Hub for strategic intelligence
US7526425B2 (en) * 2001-08-14 2009-04-28 Evri Inc. Method and system for extending keyword searching to syntactically and semantically annotated data
US7398201B2 (en) * 2001-08-14 2008-07-08 Evri Inc. Method and system for enhanced data searching
US7283951B2 (en) * 2001-08-14 2007-10-16 Insightful Corporation Method and system for enhanced data searching
US7295977B2 (en) * 2001-08-27 2007-11-13 Nec Laboratories America, Inc. Extracting classifying data in music from an audio bitstream
ES2375403T3 (en) 2001-08-27 2012-02-29 BDGB Enterprise Software Sàrl A METHOD FOR THE AUTOMATIC INDEXATION OF DOCUMENTS.
KR100483602B1 (en) * 2001-10-12 2005-04-15 (주)이캐빈 Method and system for monitoring e-mail
US6944616B2 (en) 2001-11-28 2005-09-13 Pavilion Technologies, Inc. System and method for historical database training of support vector machines
US7155608B1 (en) * 2001-12-05 2006-12-26 Bellsouth Intellectual Property Corp. Foreign network SPAM blocker
JP3726263B2 (en) * 2002-03-01 2005-12-14 ヒューレット・パッカード・カンパニー Document classification method and apparatus
US8578480B2 (en) 2002-03-08 2013-11-05 Mcafee, Inc. Systems and methods for identifying potentially malicious messages
US20060015942A1 (en) 2002-03-08 2006-01-19 Ciphertrust, Inc. Systems and methods for classification of messaging entities
JP3873135B2 (en) * 2002-03-08 2007-01-24 インターナショナル・ビジネス・マシーンズ・コーポレーション Data processing method, information processing system and program using the same
US7903549B2 (en) * 2002-03-08 2011-03-08 Secure Computing Corporation Content-based policy compliance systems and methods
US8561167B2 (en) 2002-03-08 2013-10-15 Mcafee, Inc. Web reputation scoring
JP4226261B2 (en) * 2002-04-12 2009-02-18 三菱電機株式会社 Structured document type determination system and structured document type determination method
US20030225763A1 (en) * 2002-04-15 2003-12-04 Microsoft Corporation Self-improving system and method for classifying pages on the world wide web
US7110596B2 (en) * 2002-04-25 2006-09-19 Microsoft Corporation System and method facilitating document image compression utilizing a mask
US7043079B2 (en) 2002-04-25 2006-05-09 Microsoft Corporation “Don't care” pixel interpolation
US7392472B2 (en) * 2002-04-25 2008-06-24 Microsoft Corporation Layout analysis
US7263227B2 (en) * 2002-04-25 2007-08-28 Microsoft Corporation Activity detector
US7024039B2 (en) * 2002-04-25 2006-04-04 Microsoft Corporation Block retouching
US7120297B2 (en) * 2002-04-25 2006-10-10 Microsoft Corporation Segmented layered image system
US7164797B2 (en) * 2002-04-25 2007-01-16 Microsoft Corporation Clustering
US6996575B2 (en) * 2002-05-31 2006-02-07 Sas Institute Inc. Computer-implemented system and method for text-based document processing
US7165068B2 (en) * 2002-06-12 2007-01-16 Zycus Infotech Pvt Ltd. System and method for electronic catalog classification using a hybrid of rule based and statistical method
US8046832B2 (en) 2002-06-26 2011-10-25 Microsoft Corporation Spam detector with challenges
US7016884B2 (en) * 2002-06-27 2006-03-21 Microsoft Corporation Probability estimate for K-nearest neighbor
US7415445B2 (en) * 2002-09-24 2008-08-19 Hewlett-Packard Development Company, L.P. Feature selection for two-class classification systems
US7266559B2 (en) * 2002-12-05 2007-09-04 Microsoft Corporation Method and apparatus for adapting a search classifier based on user queries
US7386527B2 (en) * 2002-12-06 2008-06-10 Kofax, Inc. Effective multi-class support vector machine classification
US6732157B1 (en) * 2002-12-13 2004-05-04 Networks Associates Technology, Inc. Comprehensive anti-spam system, method, and computer program product for filtering unwanted e-mail messages
US20040122660A1 (en) * 2002-12-20 2004-06-24 International Business Machines Corporation Creating taxonomies and training data in multiple languages
US7435306B2 (en) * 2003-01-22 2008-10-14 The Boeing Company Method for preparing rivets from cryomilled aluminum alloys and rivets produced thereby
US8335683B2 (en) * 2003-01-23 2012-12-18 Microsoft Corporation System for using statistical classifiers for spoken language understanding
US20040148170A1 (en) * 2003-01-23 2004-07-29 Alejandro Acero Statistical classifiers for spoken language understanding and command/control scenarios
US20040148201A1 (en) * 2003-01-27 2004-07-29 Smith Tracy Lee Insurance management system
US7720781B2 (en) * 2003-01-29 2010-05-18 Hewlett-Packard Development Company, L.P. Feature selection method and apparatus
US20040148267A1 (en) * 2003-01-29 2004-07-29 Forman George Henry Evaluation methodology and apparatus
US7249162B2 (en) * 2003-02-25 2007-07-24 Microsoft Corporation Adaptive junk message filtering system
US7219148B2 (en) * 2003-03-03 2007-05-15 Microsoft Corporation Feedback loop for spam prevention
US7483947B2 (en) * 2003-05-02 2009-01-27 Microsoft Corporation Message rendering for identification of content features
US7831667B2 (en) * 2003-05-15 2010-11-09 Symantec Corporation Method and apparatus for filtering email spam using email noise reduction
US7272853B2 (en) * 2003-06-04 2007-09-18 Microsoft Corporation Origination/destination features and lists for spam prevention
US7734627B1 (en) 2003-06-17 2010-06-08 Google Inc. Document similarity detection
US7519668B2 (en) * 2003-06-20 2009-04-14 Microsoft Corporation Obfuscation of spam filter
US7711779B2 (en) * 2003-06-20 2010-05-04 Microsoft Corporation Prevention of outgoing spam
US8533270B2 (en) 2003-06-23 2013-09-10 Microsoft Corporation Advanced spam detection techniques
US7155484B2 (en) * 2003-06-30 2006-12-26 Bellsouth Intellectual Property Corporation Filtering email messages corresponding to undesirable geographical regions
US7051077B2 (en) * 2003-06-30 2006-05-23 Mx Logic, Inc. Fuzzy logic voting method and system for classifying e-mail using inputs from multiple spam classifiers
CA2542937A1 (en) * 2003-07-01 2005-01-13 Cardiomag Imaging, Inc. (Cmi) Machine learning for classification of magneto cardiograms
US20090100138A1 (en) * 2003-07-18 2009-04-16 Harris Scott C Spam filter
US8650187B2 (en) * 2003-07-25 2014-02-11 Palo Alto Research Center Incorporated Systems and methods for linked event detection
US7577654B2 (en) * 2003-07-25 2009-08-18 Palo Alto Research Center Incorporated Systems and methods for new event detection
US7769626B2 (en) * 2003-08-25 2010-08-03 Tom Reynolds Determining strategies for increasing loyalty of a population to an entity
US8301482B2 (en) * 2003-08-25 2012-10-30 Tom Reynolds Determining strategies for increasing loyalty of a population to an entity
US20050060643A1 (en) * 2003-08-25 2005-03-17 Miavia, Inc. Document similarity detection and classification system
US7330112B1 (en) 2003-09-09 2008-02-12 Emigh Aaron T Location-aware services
US7818317B1 (en) * 2003-09-09 2010-10-19 James Roskind Location-based tasks
DE10345948B4 (en) * 2003-10-02 2018-08-23 Robert Bosch Gmbh Method for evaluation and temporal stabilization of classification results
US7257564B2 (en) * 2003-10-03 2007-08-14 Tumbleweed Communications Corp. Dynamic message filtering
US7664812B2 (en) * 2003-10-14 2010-02-16 At&T Intellectual Property I, L.P. Phonetic filtering of undesired email messages
US7610341B2 (en) * 2003-10-14 2009-10-27 At&T Intellectual Property I, L.P. Filtered email differentiation
US7451184B2 (en) * 2003-10-14 2008-11-11 At&T Intellectual Property I, L.P. Child protection from harmful email
US20050080642A1 (en) * 2003-10-14 2005-04-14 Daniell W. Todd Consolidated email filtering user interface
US7930351B2 (en) * 2003-10-14 2011-04-19 At&T Intellectual Property I, L.P. Identifying undesired email messages having attachments
US7996470B2 (en) * 2003-10-14 2011-08-09 At&T Intellectual Property I, L.P. Processing rules for digital messages
US20050149546A1 (en) * 2003-11-03 2005-07-07 Prakash Vipul V. Methods and apparatuses for determining and designating classifications of electronic documents
US7519565B2 (en) * 2003-11-03 2009-04-14 Cloudmark, Inc. Methods and apparatuses for classifying electronic documents
US7245765B2 (en) * 2003-11-11 2007-07-17 Sri International Method and apparatus for capturing paper-based information on a mobile computing device
US20050108203A1 (en) * 2003-11-13 2005-05-19 Chunqiang Tang Sample-directed searching in a peer-to-peer system
JP2005198970A (en) * 2004-01-19 2005-07-28 Konica Minolta Medical & Graphic Inc Medical image processor
US7532755B2 (en) * 2004-01-22 2009-05-12 Lexmark International, Inc. Image classification using concentration ratio
US8214438B2 (en) * 2004-03-01 2012-07-03 Microsoft Corporation (More) advanced spam detection features
US20050204005A1 (en) * 2004-03-12 2005-09-15 Purcell Sean E. Selective treatment of messages based on junk rating
US20050204006A1 (en) * 2004-03-12 2005-09-15 Purcell Sean E. Message junk rating interface
US7725414B2 (en) 2004-03-16 2010-05-25 Buzzmetrics, Ltd An Israel Corporation Method for developing a classifier for classifying communications
US7499588B2 (en) 2004-05-20 2009-03-03 Microsoft Corporation Low resolution OCR for camera acquired documents
US7565369B2 (en) * 2004-05-28 2009-07-21 International Business Machines Corporation System and method for mining time-changing data streams
US8484295B2 (en) 2004-12-21 2013-07-09 Mcafee, Inc. Subscriber reputation filtering method for analyzing subscriber activity and detecting account misuse
US7953814B1 (en) * 2005-02-28 2011-05-31 Mcafee, Inc. Stopping and remediating outbound messaging abuse
US7680890B1 (en) 2004-06-22 2010-03-16 Wei Lin Fuzzy logic voting method and system for classifying e-mail using inputs from multiple spam classifiers
US7664819B2 (en) * 2004-06-29 2010-02-16 Microsoft Corporation Incremental anti-spam lookup and update service
US7904517B2 (en) * 2004-08-09 2011-03-08 Microsoft Corporation Challenge response systems
CN100353355C (en) * 2004-08-12 2007-12-05 上海交通大学 Cross merge method for reducing support vector and training time
US7660865B2 (en) 2004-08-12 2010-02-09 Microsoft Corporation Spam filtering with probabilistic secure hashes
US7698339B2 (en) * 2004-08-13 2010-04-13 Microsoft Corporation Method and system for summarizing a document
US7440944B2 (en) * 2004-09-24 2008-10-21 Overture Services, Inc. Method and apparatus for efficient training of support vector machines
US7523085B2 (en) 2004-09-30 2009-04-21 Buzzmetrics, Ltd An Israel Corporation Topical sentiments in electronically stored communications
US8635690B2 (en) 2004-11-05 2014-01-21 Mcafee, Inc. Reputation based message processing
US7447362B2 (en) * 2004-11-08 2008-11-04 Dspv, Ltd. System and method of enabling a cellular/wireless device with imaging capabilities to decode printed alphanumeric characters
US7593903B2 (en) * 2004-12-03 2009-09-22 Hewlett-Packard Development Company, L.P. Method and medium for feature selection of partially labeled data
US7437334B2 (en) * 2004-12-03 2008-10-14 Hewlett-Packard Development Company, L.P. Preparing data for machine learning
US9015472B1 (en) 2005-03-10 2015-04-21 Mcafee, Inc. Marking electronic messages to indicate human origination
US9160755B2 (en) * 2004-12-21 2015-10-13 Mcafee, Inc. Trusted communication network
US8738708B2 (en) * 2004-12-21 2014-05-27 Mcafee, Inc. Bounce management in a trusted communication network
US7551782B2 (en) * 2005-02-15 2009-06-23 Dspv, Ltd. System and method of user interface and data entry from a video call
US9769354B2 (en) 2005-03-24 2017-09-19 Kofax, Inc. Systems and methods of processing scanned data
US9137417B2 (en) 2005-03-24 2015-09-15 Kofax, Inc. Systems and methods for processing video data
US7739253B1 (en) * 2005-04-21 2010-06-15 Sonicwall, Inc. Link-based content ratings of pages
US20060248094A1 (en) * 2005-04-28 2006-11-02 Microsoft Corporation Analysis and comparison of portfolios by citation
US20060248055A1 (en) * 2005-04-28 2006-11-02 Microsoft Corporation Analysis and comparison of portfolios by classification
US9792359B2 (en) * 2005-04-29 2017-10-17 Entit Software Llc Providing training information for training a categorizer
US9047290B1 (en) 2005-04-29 2015-06-02 Hewlett-Packard Development Company, L.P. Computing a quantification measure associated with cases in a category
US10445359B2 (en) * 2005-06-07 2019-10-15 Getty Images, Inc. Method and system for classifying media content
US20070112839A1 (en) * 2005-06-07 2007-05-17 Anna Bjarnestam Method and system for expansion of structured keyword vocabulary
US9158855B2 (en) 2005-06-16 2015-10-13 Buzzmetrics, Ltd Extracting structured data from weblogs
US7930353B2 (en) * 2005-07-29 2011-04-19 Microsoft Corporation Trees of classifiers for detecting email spam
US20070100779A1 (en) * 2005-08-05 2007-05-03 Ori Levy Method and system for extracting web data
US7599556B2 (en) * 2005-08-25 2009-10-06 Joseph Stanley Czyszczewski Apparatus, system, and method for scanning segmentation
GB2445688A (en) 2005-09-01 2008-07-16 Zvi Haim Lev System and method for reliable content access using a cellular/wireless device with imaging capabilities
US7406450B2 (en) * 2005-09-28 2008-07-29 Nec Laboratories America, Inc. Spread kernel support vector machine
US20070118759A1 (en) * 2005-10-07 2007-05-24 Sheppard Scott K Undesirable email determination
US8065370B2 (en) 2005-11-03 2011-11-22 Microsoft Corporation Proofs to filter spam
US20090017765A1 (en) * 2005-11-04 2009-01-15 Dspv, Ltd System and Method of Enabling a Cellular/Wireless Device with Imaging Capabilities to Decode Printed Alphanumeric Characters
CN101310294A (en) * 2005-11-15 2008-11-19 伯纳黛特·加纳 Method for training neural networks
EP1949273A1 (en) 2005-11-16 2008-07-30 Evri Inc. Extending keyword searching to syntactically and semantically annotated data
US8380696B1 (en) 2005-12-20 2013-02-19 Emc Corporation Methods and apparatus for dynamically classifying objects
WO2007081234A1 (en) * 2006-01-12 2007-07-19 Otkrytoe Aktsionernoe Obschestvo 'bineuro' Device for encoding semantics of text-based documents
US7333965B2 (en) * 2006-02-23 2008-02-19 Microsoft Corporation Classifying text in a code editor using multiple classifiers
US8862680B2 (en) * 2006-05-01 2014-10-14 The Boeing Company Methods and systems for data prioritization
US8417783B1 (en) * 2006-05-31 2013-04-09 Proofpoint, Inc. System and method for improving feature selection for a spam filtering model
US7937345B2 (en) * 2006-07-12 2011-05-03 Kofax, Inc. Data classification methods using machine learning techniques
US7761391B2 (en) 2006-07-12 2010-07-20 Kofax, Inc. Methods and systems for improved transductive maximum entropy discrimination classification
US20080086432A1 (en) * 2006-07-12 2008-04-10 Schmidtler Mauritius A R Data classification methods using machine learning techniques
US7958067B2 (en) * 2006-07-12 2011-06-07 Kofax, Inc. Data classification methods using machine learning techniques
US8014591B2 (en) * 2006-09-13 2011-09-06 Aurilab, Llc Robust pattern recognition system and method using socratic agents
US8966389B2 (en) * 2006-09-22 2015-02-24 Limelight Networks, Inc. Visual interface for identifying positions of interest within a sequentially ordered information encoding
US8396878B2 (en) 2006-09-22 2013-03-12 Limelight Networks, Inc. Methods and systems for generating automated tags for video files
US8204891B2 (en) * 2007-09-21 2012-06-19 Limelight Networks, Inc. Method and subsystem for searching media content within a content-search-service system
US9015172B2 (en) 2006-09-22 2015-04-21 Limelight Networks, Inc. Method and subsystem for searching media content within a content-search service system
US7917492B2 (en) * 2007-09-21 2011-03-29 Limelight Networks, Inc. Method and subsystem for information acquisition and aggregation to facilitate ontology and language-model generation within a content-search-service system
US7660783B2 (en) * 2006-09-27 2010-02-09 Buzzmetrics, Inc. System and method of ad-hoc analysis of data
US20080097982A1 (en) * 2006-10-18 2008-04-24 Yahoo! Inc. System and method for classifying search queries
US8738359B2 (en) * 2006-10-18 2014-05-27 Honda Motor Co., Ltd. Scalable knowledge extraction
US20080104066A1 (en) * 2006-10-27 2008-05-01 Yahoo! Inc. Validating segmentation criteria
US8260602B1 (en) 2006-11-02 2012-09-04 The Math Works, Inc. Timer analysis and identification
US8224905B2 (en) 2006-12-06 2012-07-17 Microsoft Corporation Spam filtration utilizing sender activity data
US8065307B2 (en) * 2006-12-20 2011-11-22 Microsoft Corporation Parsing, analysis and scoring of document content
US7873583B2 (en) * 2007-01-19 2011-01-18 Microsoft Corporation Combining resilient classifiers
US8364617B2 (en) * 2007-01-19 2013-01-29 Microsoft Corporation Resilient classification of data
US8763114B2 (en) 2007-01-24 2014-06-24 Mcafee, Inc. Detecting image spam
US7779156B2 (en) 2007-01-24 2010-08-17 Mcafee, Inc. Reputation based load balancing
US8214497B2 (en) 2007-01-24 2012-07-03 Mcafee, Inc. Multi-dimensional reputation scoring
CA2717462C (en) 2007-03-14 2016-09-27 Evri Inc. Query templates and labeled search tip system, methods, and techniques
CN100565559C (en) * 2007-03-14 2009-12-02 中国科学院自动化研究所 Image text location method and device based on connected component and support vector machine
US8180713B1 (en) 2007-04-13 2012-05-15 Standard & Poor's Financial Services Llc System and method for searching and identifying potential financial risks disclosed within a document
US20080281922A1 (en) * 2007-05-09 2008-11-13 Microsoft Corporation Automatic generation of email previews and summaries
CN101382934B (en) * 2007-09-06 2010-08-18 华为技术有限公司 Search method for multimedia model, apparatus and system
WO2009052308A1 (en) 2007-10-17 2009-04-23 Roseman Neil S Nlp-based content recommender
US8594996B2 (en) 2007-10-17 2013-11-26 Evri Inc. NLP-based entity recognition and disambiguation
US8185930B2 (en) 2007-11-06 2012-05-22 Mcafee, Inc. Adjusting filter or classification control settings
US20090125980A1 (en) * 2007-11-09 2009-05-14 Secure Computing Corporation Network rating
US8862582B2 (en) * 2007-11-15 2014-10-14 At&T Intellectual Property I, L.P. System and method of organizing images
US8250469B2 (en) * 2007-12-03 2012-08-21 Microsoft Corporation Document layout extraction
US8392816B2 (en) * 2007-12-03 2013-03-05 Microsoft Corporation Page classifier engine
US8347326B2 (en) 2007-12-18 2013-01-01 The Nielsen Company (US) Identifying key media events and modeling causal relationships between key events and reported feelings
US7853599B2 (en) * 2008-01-21 2010-12-14 Microsoft Corporation Feature selection for ranking
JP5330720B2 (en) * 2008-03-24 2013-10-30 株式会社エムティーアイ Chord identification method, chord identification device, and learning device
US8589503B2 (en) * 2008-04-04 2013-11-19 Mcafee, Inc. Prioritizing network traffic
US20090274376A1 (en) * 2008-05-05 2009-11-05 Yahoo! Inc. Method for efficiently building compact models for large multi-class text classification
CN100583101C (en) * 2008-06-12 2010-01-20 昆明理工大学 Text categorization feature selection and weight computation method based on field knowledge
JP5423676B2 (en) 2008-07-30 2014-02-19 日本電気株式会社 Data classification system, data classification method, and data classification program
JP5500070B2 (en) * 2008-07-30 2014-05-21 日本電気株式会社 Data classification system, data classification method, and data classification program
US10354229B2 (en) 2008-08-04 2019-07-16 Mcafee, Llc Method and system for centralized contact management
US8799279B2 (en) 2008-12-31 2014-08-05 At&T Intellectual Property I, L.P. Method and apparatus for using a discriminative classifier for processing a query
JP2010165160A (en) * 2009-01-15 2010-07-29 Kddi Corp Musical piece classification device, musical piece retrieval system and computer program
US9349046B2 (en) 2009-02-10 2016-05-24 Kofax, Inc. Smart optical input/output (I/O) extension for context-dependent workflows
US9576272B2 (en) 2009-02-10 2017-02-21 Kofax, Inc. Systems, methods and computer program products for determining document validity
US8774516B2 (en) 2009-02-10 2014-07-08 Kofax, Inc. Systems, methods and computer program products for determining document validity
US9767354B2 (en) 2009-02-10 2017-09-19 Kofax, Inc. Global geographic information retrieval, validation, and normalization
US8958605B2 (en) 2009-02-10 2015-02-17 Kofax, Inc. Systems, methods and computer program products for determining document validity
US8339680B2 (en) * 2009-04-02 2012-12-25 Xerox Corporation Printer image log system for document gathering and retention
CA2796408A1 (en) * 2009-04-16 2010-10-21 Evri Inc. Enhanced advertisement targeting
CN101901345B (en) * 2009-05-27 2013-02-27 复旦大学 Classification method of differential proteomics
JP5305241B2 (en) * 2009-06-05 2013-10-02 株式会社リコー Classification parameter generation apparatus, generation method, and generation program
US8738354B2 (en) * 2009-06-19 2014-05-27 Microsoft Corporation Trans-lingual representation of text documents
CN101587493B (en) * 2009-06-29 2012-07-04 中国科学技术大学 Text classification method
CN102023986B (en) * 2009-09-22 2015-09-30 日电(中国)有限公司 The method and apparatus of text classifier is built with reference to external knowledge
US9213756B2 (en) * 2009-11-02 2015-12-15 Harry Urbschat System and method of using dynamic variance networks
US9158833B2 (en) * 2009-11-02 2015-10-13 Harry Urbschat System and method for obtaining document information
US9152883B2 (en) * 2009-11-02 2015-10-06 Harry Urbschat System and method for increasing the accuracy of optical character recognition (OCR)
US8321357B2 (en) * 2009-09-30 2012-11-27 Lapir Gennady Method and system for extraction
CA2779208C (en) * 2009-10-30 2016-03-22 Evri, Inc. Improving keyword-based search engine results using enhanced query strategies
KR20110064082A (en) * 2009-12-07 2011-06-15 삼성전자주식회사 Method for providing function to write text and function to scrap, and electronic apparatus
CN102163190B (en) * 2010-02-20 2013-09-11 三星电子(中国)研发中心 Method and apparatus of making text feature extracting strategy, method and apparatus of text classification
US9710556B2 (en) 2010-03-01 2017-07-18 Vcvc Iii Llc Content recommendation based on collections of entities
CN102200981B (en) * 2010-03-25 2013-07-17 三星电子(中国)研发中心 Feature selection method and feature selection device for hierarchical text classification
US8645125B2 (en) 2010-03-30 2014-02-04 Evri, Inc. NLP-based systems and methods for providing quotations
US8140567B2 (en) 2010-04-13 2012-03-20 Microsoft Corporation Measuring entity extraction complexity
US8429098B1 (en) * 2010-04-30 2013-04-23 Global Eprocure Classification confidence estimating tool
US8621638B2 (en) 2010-05-14 2013-12-31 Mcafee, Inc. Systems and methods for classification of messaging entities
US8874727B2 (en) 2010-05-31 2014-10-28 The Nielsen Company (Us), Llc Methods, apparatus, and articles of manufacture to rank users in an online social network
US8838633B2 (en) 2010-08-11 2014-09-16 Vcvc Iii Llc NLP-based sentiment analysis
US9405848B2 (en) 2010-09-15 2016-08-02 Vcvc Iii Llc Recommending mobile device activities
US8924391B2 (en) 2010-09-28 2014-12-30 Microsoft Corporation Text classification using concept kernel
US8725739B2 (en) 2010-11-01 2014-05-13 Evri, Inc. Category-based content recommendation
CN101976270B (en) * 2010-11-29 2012-04-25 南京师范大学 Uncertain reasoning-based text hierarchy classification method and device
US9589254B2 (en) 2010-12-08 2017-03-07 Microsoft Technology Licensing, Llc Using e-mail message characteristics for prioritization
CN102073704B (en) * 2010-12-24 2013-09-25 华为终端有限公司 Text classification processing method, system and equipment
US9116995B2 (en) 2011-03-30 2015-08-25 Vcvc Iii Llc Cluster-based identification of news stories
US8650198B2 (en) * 2011-08-15 2014-02-11 Lockheed Martin Corporation Systems and methods for facilitating the gathering of open source intelligence
US9483794B2 (en) 2012-01-12 2016-11-01 Kofax, Inc. Systems and methods for identification document processing and business workflow integration
US9058580B1 (en) 2012-01-12 2015-06-16 Kofax, Inc. Systems and methods for identification document processing and business workflow integration
US9158967B2 (en) 2012-01-12 2015-10-13 Kofax, Inc. Systems and methods for mobile image capture and processing
US10146795B2 (en) 2012-01-12 2018-12-04 Kofax, Inc. Systems and methods for mobile image capture and processing
US9058515B1 (en) 2012-01-12 2015-06-16 Kofax, Inc. Systems and methods for identification document processing and business workflow integration
US10354187B2 (en) * 2013-01-17 2019-07-16 Hewlett Packard Enterprise Development Lp Confidentiality of files using file vectorization and machine learning
US9483566B2 (en) * 2013-01-23 2016-11-01 Google Inc. System and method for determining the legitimacy of a listing
EP2973226A4 (en) 2013-03-13 2016-06-29 Kofax Inc Classifying objects in digital images captured using mobile devices
US9355312B2 (en) 2013-03-13 2016-05-31 Kofax, Inc. Systems and methods for classifying objects in digital images captured using mobile devices
US9208536B2 (en) 2013-09-27 2015-12-08 Kofax, Inc. Systems and methods for three dimensional geometric reconstruction of captured image data
US20140316841A1 (en) 2013-04-23 2014-10-23 Kofax, Inc. Location-based workflows and services
EP2992481A4 (en) 2013-05-03 2017-02-22 Kofax, Inc. Systems and methods for detecting and classifying objects in video captured using mobile devices
JP6436086B2 (en) * 2013-09-18 2018-12-12 日本電気株式会社 Classification dictionary generation device, classification dictionary generation method, and program
WO2015073920A1 (en) 2013-11-15 2015-05-21 Kofax, Inc. Systems and methods for generating composite images of long documents using mobile video data
FR3016981A1 (en) * 2014-01-28 2015-07-31 Deadia METHOD OF SEMANTICALLY ANALYZING A TEXT
CN105138527B (en) * 2014-05-30 2019-02-12 华为技术有限公司 A kind of data classification homing method and device
US9760788B2 (en) 2014-10-30 2017-09-12 Kofax, Inc. Mobile document detection and orientation based on reference object characteristics
CN104391828B (en) * 2014-11-11 2017-11-17 百度在线网络技术(北京)有限公司 The method and apparatus for determining short text similarity
US10331782B2 (en) 2014-11-19 2019-06-25 Lexisnexis, A Division Of Reed Elsevier Inc. Systems and methods for automatic identification of potential material facts in documents
CN104572631B (en) * 2014-12-03 2018-04-13 北京捷通华声语音技术有限公司 The training method and system of a kind of language model
US10127284B2 (en) * 2014-12-18 2018-11-13 International Business Machines Corporation Scoring attributes in deep question answering systems based on algorithmic source code influences
US10929774B2 (en) * 2015-02-12 2021-02-23 Koninklijke Philips N.V. Robust classifier
CN106294506B (en) * 2015-06-10 2020-04-24 华中师范大学 Domain-adaptive viewpoint data classification method and device
CN105045812B (en) * 2015-06-18 2019-01-29 上海高欣计算机系统有限公司 The classification method and system of text subject
US10242285B2 (en) 2015-07-20 2019-03-26 Kofax, Inc. Iterative recognition-guided thresholding and data extraction
CN105117384A (en) * 2015-08-19 2015-12-02 小米科技有限责任公司 Classifier training method, and type identification method and apparatus
JP6611053B2 (en) * 2015-09-17 2019-11-27 パナソニックIpマネジメント株式会社 Subject estimation system, subject estimation method and program
CN105488025B (en) 2015-11-24 2019-02-12 小米科技有限责任公司 Template construction method and device, information identifying method and device
CN106933847B (en) * 2015-12-30 2019-12-27 北京神州泰岳软件股份有限公司 Method and device for establishing data classification model
US9779296B1 (en) 2016-04-01 2017-10-03 Kofax, Inc. Content-based detection and three dimensional geometric reconstruction of objects in image and video data
CN107818076B (en) * 2016-09-12 2021-11-12 微软技术许可有限责任公司 Semantic processing for natural language
US10949772B2 (en) * 2017-01-24 2021-03-16 International Business Machines Corporation System for evaluating journal articles
US11100422B2 (en) 2017-01-24 2021-08-24 International Business Machines Corporation System for evaluating journal articles
CN107316065B (en) * 2017-06-26 2021-03-02 刘艳 Sparse feature extraction and classification method based on fractional subspace model
US10878144B2 (en) 2017-08-10 2020-12-29 Allstate Insurance Company Multi-platform model processing and execution management engine
US11755949B2 (en) 2017-08-10 2023-09-12 Allstate Insurance Company Multi-platform machine learning systems
CN107368611B (en) * 2017-08-11 2018-06-26 同济大学 A kind of short text classification method
CN107506434A (en) * 2017-08-23 2017-12-22 北京百度网讯科技有限公司 Method and apparatus based on artificial intelligence classification phonetic entry text
CN108230359B (en) * 2017-11-12 2021-01-26 北京市商汤科技开发有限公司 Object detection method and apparatus, training method, electronic device, program, and medium
US10803350B2 (en) 2017-11-30 2020-10-13 Kofax, Inc. Object detection and image cropping using a multi-detector approach
RU2678716C1 (en) * 2017-12-11 2019-01-31 Общество с ограниченной ответственностью "Аби Продакшн" Use of autoencoders for learning text classifiers in natural language
US10303771B1 (en) * 2018-02-14 2019-05-28 Capital One Services, Llc Utilizing machine learning models to identify insights in a document
US11055492B2 (en) * 2018-06-02 2021-07-06 Apple Inc. Privatized apriori algorithm for sequential data discovery
CN109241770B (en) * 2018-08-10 2021-11-09 深圳前海微众银行股份有限公司 Information value calculation method and device based on homomorphic encryption and readable storage medium
CN109325357B (en) * 2018-08-10 2021-12-14 深圳前海微众银行股份有限公司 RSA-based information value calculation method, device and readable storage medium
CN109165300B (en) * 2018-08-31 2020-08-11 中国科学院自动化研究所 Text inclusion recognition method and device
US11087088B2 (en) * 2018-09-25 2021-08-10 Accenture Global Solutions Limited Automated and optimal encoding of text data features for machine learning models
US11093619B2 (en) 2018-10-27 2021-08-17 International Business Machines Corporation Automated fixups based on partial goal satisfaction
CN109918497A (en) * 2018-12-21 2019-06-21 厦门市美亚柏科信息股份有限公司 A kind of file classification method, device and storage medium based on improvement textCNN model
US11914606B2 (en) * 2019-03-04 2024-02-27 Walmart Apollo, Llc Systems and methods for a machine learning framework
US11012730B2 (en) 2019-03-29 2021-05-18 Wipro Limited Method and system for automatically updating video content
US11354609B2 (en) * 2019-04-17 2022-06-07 International Business Machines Corporation Dynamic prioritization of action items
US11270077B2 (en) * 2019-05-13 2022-03-08 International Business Machines Corporation Routing text classifications within a cross-domain conversational service
CN110110087A (en) * 2019-05-15 2019-08-09 济南浪潮高新科技投资发展有限公司 A kind of Feature Engineering method for Law Text classification based on two classifiers
US11537602B2 (en) 2020-05-12 2022-12-27 International Business Machines Corporation Computer implemented live cross walks in compliance mappings in response to regulatory changes and assessing risks of changes
US20220101190A1 (en) * 2020-09-30 2022-03-31 Alteryx, Inc. System and method of operationalizing automated feature engineering
US11620453B2 (en) * 2020-10-16 2023-04-04 CS Disco, Inc. System and method for artificial intelligence driven document analysis, including searching, indexing, comparing or associating datasets based on learned representations
US11455324B2 (en) 2020-10-23 2022-09-27 Settle Smart Ltd. Method for determining relevant search results
US11630853B2 (en) * 2021-01-29 2023-04-18 Snowflake Inc. Metadata classification
US11348617B1 (en) 2021-03-08 2022-05-31 Bank Of America Corporation System for implementing content retrofitting using information vectorization
US11281858B1 (en) * 2021-07-13 2022-03-22 Exceed AI Ltd Systems and methods for data classification
WO2023027109A1 (en) * 2021-08-25 2023-03-02 富士フイルム株式会社 Device for generating data merging rule for machine learning model, operation method and program for device for generating data merging rule, learning device for machine learning model, and operation method and program for learning device
CN113761231B (en) * 2021-09-07 2022-07-12 浙江传媒学院 Text character feature-based text data attribution description and generation method

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP0747846A2 (en) * 1995-06-07 1996-12-11 AT&T IPM Corp. Method and apparatus for training a text classifier

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5999639A (en) * 1997-09-04 1999-12-07 Qualia Computing, Inc. Method and system for automated detection of clustered microcalcifications from digital mammograms
US6115052A (en) * 1998-02-12 2000-09-05 Mitsubishi Electric Information Technology Center America, Inc. (Ita) System for reconstructing the 3-dimensional motions of a human figure from a monocularly-viewed image sequence
US6115708A (en) * 1998-03-04 2000-09-05 Microsoft Corporation Method for refining the initial conditions for clustering with applications to small and large database clustering

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP0747846A2 (en) * 1995-06-07 1996-12-11 AT&T IPM Corp. Method and apparatus for training a text classifier

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
JOACHIMS T: "Text categorization with support vector machines: learning with many relevant features", MACHINE LEARNING: ECML-98. 10TH EUROPEAN CONFERENCE ON MACHINE LEARNING. PROCEEDINGS, MACHINE LEARNING: ECML-98 10TH EUROPEAN CONFERENCE ON MACHINE LEARNING. PROCEEDINGS, CHEMNITZ, GERMANY, 21-23 APRIL 1998, 1998, Berlin, Germany, Springer-Verlag, Germany, pages 137 - 142, XP002119808, ISBN: 3-540-64417-2 *
OSUNA E ET AL: "An improved training algorithm for support vector machines", NEURAL NETWORKS FOR SIGNAL PROCESSING VII. PROCEEDINGS OF THE 1997 IEEE SIGNAL PROCESSING SOCIETY WORKSHOP (CAT. NO.97TH8330), NEURAL NETWORKS FOR SIGNAL PROCESSING VII. PROCEEDINGS OF THE 1997 IEEE SIGNAL PROCESSING SOCIETY WORKSHOP, AMELIA ISLAND,, 1997, New York, NY, USA, IEEE, USA, pages 276 - 285, XP002119807, ISBN: 0-7803-4256-9 *
Y. YANG & J. O. PEDERSEN: "A Comparative Study on Feature Selection in Text Categorization", THE FOURTEENTH INTERNATIONAL CONFERENCE ON MACHINE LEARNING, ICML '97, 8 July 1997 (1997-07-08) - 12 July 1997 (1997-07-12), Nashville, TN, USA, XP002119806, Retrieved from the Internet <URL:http://www.cs.cmu.edu/~yiming/papers.yy/ml97.ps> [retrieved on 19991018] *

Cited By (27)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2000019335A1 (en) * 1998-09-30 2000-04-06 Siemens Aktiengesellschaft Method and device for assigning an object to at least one class
EP1154358A3 (en) * 2000-05-12 2003-08-06 Applied Psychology Research Limited Automatic text classification system
EP1154358A2 (en) * 2000-05-12 2001-11-14 Applied Psychology Research Limited Automatic text classification system
WO2001090921A2 (en) * 2000-05-25 2001-11-29 Kanisa, Inc. System and method for automatically classifying text
WO2001090921A3 (en) * 2000-05-25 2003-12-24 Kanisa Inc System and method for automatically classifying text
JP4773672B2 (en) * 2000-05-31 2011-09-14 レクシスネクシス ア ディヴィジョン オブ リード エルザヴィア インコーポレイテッド Computer-based system and method for finding the rule of law in text
JP2003535407A (en) * 2000-05-31 2003-11-25 レキシス ネクシス Computer-based system and method for finding the rule of law in text
US6778986B1 (en) 2000-07-31 2004-08-17 Eliyon Technologies Corporation Computer method and apparatus for determining site type of a web site
US7054886B2 (en) 2000-07-31 2006-05-30 Zoom Information, Inc. Method for maintaining people and organization information
WO2002010982A3 (en) * 2000-07-31 2003-07-10 Eliyon Technologies Corp Computer system for collecting information from web sites
WO2002010957A3 (en) * 2000-07-31 2003-04-10 Eliyon Technologies Corp Computer method and apparatus for determining content types of web pages
WO2002010957A2 (en) * 2000-07-31 2002-02-07 Eliyon Technologies Corporation Computer method and apparatus for determining content types of web pages
WO2002010982A2 (en) * 2000-07-31 2002-02-07 Eliyon Technologies Corporation Computer system for collecting information from web sites
US7356761B2 (en) 2000-07-31 2008-04-08 Zoom Information, Inc. Computer method and apparatus for determining content types of web pages
US7065483B2 (en) 2000-07-31 2006-06-20 Zoom Information, Inc. Computer method and apparatus for extracting data from web pages
US6983282B2 (en) 2000-07-31 2006-01-03 Zoom Information, Inc. Computer method and apparatus for collecting people and organization information from Web sites
US6618717B1 (en) 2000-07-31 2003-09-09 Eliyon Technologies Corporation Computer method and apparatus for determining content owner of a website
EP1430378A4 (en) * 2001-05-08 2008-04-30 Sunflare Co Ltd New differential lsi space-based probabilistic document classifier
EP1430378A2 (en) * 2001-05-08 2004-06-23 Sunflare Co. Ltd. New differential lsi space-based probabilistic document classifier
US6868411B2 (en) 2001-08-13 2005-03-15 Xerox Corporation Fuzzy text categorizer
EP1480167A4 (en) * 2002-02-27 2007-06-20 Nec Corp Pattern feature selection method, classification method, judgment method, program, and device
EP1480167A1 (en) * 2002-02-27 2004-11-24 NEC Corporation Pattern feature selection method, classification method, judgment method, program, and device
US7634140B2 (en) 2002-02-27 2009-12-15 Nec Corporation Pattern feature selection method, classification method, judgment method, program, and device
WO2009132188A1 (en) * 2008-04-24 2009-10-29 Boston Scientific Scimed, Inc. Methods, systems, and devices for tissue characterization by spectral similarity of intravascular ultrasound signals
US9549713B2 (en) 2008-04-24 2017-01-24 Boston Scientific Scimed, Inc. Methods, systems, and devices for tissue characterization and quantification using intravascular ultrasound signals
US10456105B2 (en) 2015-05-05 2019-10-29 Boston Scientific Scimed, Inc. Systems and methods with a swellable material disposed over a transducer of an ultrasound imaging system
CN107193836A (en) * 2016-03-15 2017-09-22 腾讯科技(深圳)有限公司 A kind of recognition methods and device

Also Published As

Publication number Publication date
CN1310825A (en) 2001-08-29
EP1090365A1 (en) 2001-04-11
US6192360B1 (en) 2001-02-20
JP4514082B2 (en) 2010-07-28
EP1090365B1 (en) 2013-07-24
CN1307579C (en) 2007-03-28
JP2002519766A (en) 2002-07-02

Similar Documents

Publication Publication Date Title
EP1090365B1 (en) Methods and apparatus for classifying text and for building a text classifier
Goudjil et al. A novel active learning method using SVM for text classification
US6327581B1 (en) Methods and apparatus for building a support vector machine classifier
Crammer et al. Adaptive regularization of weight vectors
US8566270B2 (en) Sparse representations for text classification
Harish et al. Representation and classification of text documents: A brief review
US6571225B1 (en) Text categorizers based on regularizing adaptations of the problem of computing linear separators
US7689531B1 (en) Automatic charset detection using support vector machines with charset grouping
He et al. On machine learning methods for Chinese document categorization
Tyagi et al. Second order training and sizing for the multilayer perceptron
US8560466B2 (en) Method and arrangement for automatic charset detection
Bouguila et al. A discrete mixture-based kernel for SVMs: application to spam and image categorization
Pengcheng et al. Chinese calligraphic style representation for recognition
US7836000B2 (en) System and method for training a multi-class support vector machine to select a common subset of features for classifying objects
WO2023055858A1 (en) Systems and methods for machine learning-based data extraction
WO2022035942A1 (en) Systems and methods for machine learning-based document classification
Sethuraman et al. An improved feature selection based on naive bayes with kernel density estimator for opinion mining
Lhazmir et al. Feature extraction based on principal component analysis for text categorization
Mansour et al. Text vectorization method based on concept mining using clustering techniques
Wei et al. Study of text classification methods for data sets with huge features
Berahmand et al. An Improved Deep Text Clustering via Local Manifold of an Autoencoder Embedding
Yao et al. Collaborative text categorization via exploiting sparse coefficients
Dang et al. Unsupervised threshold autoencoder to analyze and understand sentence elements
US11599588B1 (en) Apparatus and method of entity data aggregation
Abdulkhaliq et al. Sentiment Analysis Using Hybrid Feature Selection Techniques

Legal Events

Date Code Title Description
WWE Wipo information: entry into national phase

Ref document number: 99808930.3

Country of ref document: CN

AK Designated states

Kind code of ref document: A1

Designated state(s): CA CN JP

AL Designated countries for regional patents

Kind code of ref document: A1

Designated state(s): AT BE CH CY DE DK ES FI FR GB GR IE IT LU MC NL PT SE

121 Ep: the epo has been informed by wipo that ep was designated in this application
DFPE Request for preliminary examination filed prior to expiration of 19th month from priority date (pct application filed before 20040101)
ENP Entry into the national phase

Ref document number: 2000 556321

Country of ref document: JP

Kind code of ref document: A

WWE Wipo information: entry into national phase

Ref document number: 1999933534

Country of ref document: EP

WWP Wipo information: published in national office

Ref document number: 1999933534

Country of ref document: EP