US20150052113A1 - Answer Determination for Natural Language Questioning - Google Patents
Answer Determination for Natural Language Questioning Download PDFInfo
- Publication number
- US20150052113A1 US20150052113A1 US14/479,765 US201414479765A US2015052113A1 US 20150052113 A1 US20150052113 A1 US 20150052113A1 US 201414479765 A US201414479765 A US 201414479765A US 2015052113 A1 US2015052113 A1 US 2015052113A1
- Authority
- US
- United States
- Prior art keywords
- candidate
- question
- natural language
- approach
- search
- Prior art date
- Legal status (The legal status 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 status listed.)
- Abandoned
Links
Images
Classifications
-
- G06F17/30401—
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/24—Querying
- G06F16/242—Query formulation
- G06F16/243—Natural language query formulation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/33—Querying
- G06F16/3331—Query processing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/951—Indexing; Web crawling techniques
-
- G06F17/30864—
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
- G06F40/35—Discourse or dialogue representation
Definitions
- the invention relates generally to communication systems. More specifically, the invention relates to finding factoid answers to natural language questions within an electronic operating environment.
- Search engines have become one tool for discriminating mounds of information.
- Conventional systems allow a user to enter keywords, and the engines search the vast amounts of data to find corresponding results.
- the results may be based upon the frequency that other users access that data or it may be based upon the approximated correlation between the keyword(s) searched and the different data.
- search engines provide a means for narrowing the amount of data a user must review.
- search engines allow a user to answer specific questions.
- the web site “Ask Jeeves,” by Ask Jeeves, Inc. of Oakland, Calif. allows a user to type in a question and the system will perform a search and retrieval function to attempt to find an answer to the question.
- a web site takes an input question of a user, such as “What is the capital of California?,” searches the Internet, and retrieves web pages that may include the answer to the question.
- an introduction portion of the web site may be included on a results page.
- a concise answer is not provided alone. The answer is provided within link to a web site.
- numerous irrelevant links to web pages are retrieved, requiring a user to wade through the data and perform her personal determination of the answer.
- Question answering is the task of finding a concise answer to a Natural Language question.
- Question answering systems use a search engine but differ from a web search task by the type of its input and output.
- Input to a search engine is a query while a QA system takes a natural language question as an input. For example, if a user wants to find out, “Who first broke the sound barrier?”, the user of a search engine would convert the question into a query like first AND broke AND “sound barrier,” while the user of a QA system would type or say the question.
- the output of the QA system is a concise answer: Yeager, in the case of the above question while the search engine produces snippets with link to the web pages.
- semantic role labeling is applied to the QA task for factoid questions, answers to a factoid question being a single word or a short phrase.
- the traditional QA system modular architecture consisting of question classification, query generation, search, answer extraction, and re-ranking may be utilized.
- FIG. 1 illustrates an example block diagram of a Baseline system architecture in accordance with at least one aspect of the present invention
- FIG. 2 illustrates an example block diagram of a question answering (QASR) system architecture in accordance with at least one aspect of the present invention
- FIG. 3 an example block diagram of a 2-tier cascaded approach in accordance with at least one aspect of the present invention
- FIG. 4 is an example block diagram of a 3-tier cascaded approach in accordance with at least one aspect of the present invention.
- FIG. 5 is a flow chart of an illustrative example of a method for determining a factoid answer to a natural language question in accordance with at least one aspect of the present invention.
- FIGS. 6A-6C are a flowchart of an illustrative example of a method for determining a searched argument classification corresponding to a natural language question in accordance with at least one aspect of the present invention.
- a system may comprise one or more memories configured to store computer readable instructions as described herein. There memories are computer-readable media having computer-executable commands for performing one or more methods. In one or more configurations, these computer readable instructions may be software modules. In addition, the system may further comprise one or more processors configured to execute the computer readable instructions stored on the one or more memories. In one or more configurations of the present invention, a computer may be used to implement one or more aspects of the present invention as described herein.
- semantic role labeling aims to identify the predicate/argument relations within a sentence.
- question and candidate sentences extracted using a search engine are processed to identify predicate/argument structure.
- the semantic role labeler described in Pradhan et al., Semantic Role Labeling Using Different Syntactic Views , In Proceedings of the Association for Computational Linguistics 43rd annual meeting (ACL-2005) may be used.
- the present invention is not so limited to any one semantic role labeler and that any number of different semantic role labelers may be utilized in accordance with aspects of the present invention.
- semantic role labeling refers to techniques for classifying word(s) or portions of a sentence into object and predicate designations. Those skilled in the art include individuals with an understanding of ontologies and semantic principles of languages.
- TARGET tags are assigned to the predicate and ARG0, ARG1, ARG2 . . . , ARGM-TMP, and ARGM-LOC tags are assigned to arguments.
- ARG0 serves as an agent
- ARG1, ARG2 . . . are objects.
- Arguments of the type ARGM-TMP, which represents temporal argument, and ARGM-LOC, which represents a location argument, are shared over all predicates.
- semantic role labeling may be seen in the following example.
- a sentence “Nostradamus was born in 1503 in the south of France” is tagged by a semantic role labeling program as [ARG1 Nostradamus] was [TARGET born] [ARGM-TMP in 1503] [ARGM-LOC in the south of France].
- “born” is identified as a target predicate with three arguments, object “Nostradamus”, temporal argument “in 1503,” and location argument “in the south of France.”
- the ability to extract the predicate-argument structure is a technique useful in many applications.
- a question answering system that uses semantic role labeling is utilized and evaluated using a combination of automatic and manual evaluation on a set of Who/When/Where/What questions.
- the searched argument of the question is heuristically determined and this argument is extracted from the candidate sentences where the argument is applied to the matching predicate.
- the candidate sentence is tagged.
- Garfield is a popular comic strip created by Jim Davis featuring the cat Garfield . . .
- Garfield becomes analyzed as Garfield is [ARG1 a popular comic strip] [TARGET created] [ARG0 by Jim Davis] featuring the cat Garfield . . .
- ARG0 corresponds to the “agent,” and the answer to the “who” question is expected to be an agent. Therefore, we can extract Jim Davis as the answer.
- predicate/argument extraction approach using semantic role labeling has been successfully applied to the task of spoken language understanding where predicate-argument extraction eliminates the need for the costly data labeling in the call-type classification system, achieving the performance close to the one of the manually labeled system.
- the method has successfully been applied to an Information Extraction task where, again, using predicate/argument extraction, the system's performance is close to the one using manually created patterns.
- Some Question Answering system focus on using a fixed size document set to find answers, while others use the Internet as a resource.
- a fixed set of documents is generally more reliable and accurate, as one can choose which documents to use, but it often lacks redundancy.
- the Internet in the alternative, lacks accuracy while providing a greater redundancy. Redundancy of the Internet information benefits a QA system because evaluation of the system occurs on a range of the number of snippets used for the answer extraction. Finding an answer in a small set of documents is a harder problem because the questions are less likely to appear in the form of the question phrase and it requires more careful analysis of the question and query generation in order to widen the range of potential answers in order to achieve a similar performance as using Internet data.
- Some systems incorporate a combination of sophisticated NLP (natural language processing) techniques including syntactic parsing, named entity detection, lexical chains derived from extended WordNet ontologies, and logical prover.
- Predicate/argument identification in the question and the candidate answer improves the performance of the answers extraction module from the candidate sentences.
- a grammatic relations approach aims at the extraction of the answers from candidate sentences. Grammatic relations extraction may be based on the grammatic annotations from the Penn Treebank. Semantic roles in a sentence represent a predicate/argument structure while grammatic roles are an approximation to it.
- Semantic symmetry occurs when the meaning of a sentence can not be derived using lexical information alone.
- the sentences “The bird ate the snake” and “The snake ate the bird” are similar lexically as both of the nouns are animate objects.
- Ambiguous modification occurs when it is difficult to determine which adjective is being modified. For example, in the sentence, “the planet's largest volcano,” a system must be able to discern whether “largest” applies to the planet or to the volcano.
- the Internet is used as a data source for search and extraction of candidate documents.
- a search engine such as the search engine by GoogleTM of Mountain View, Calif.
- GoogleTM search engine by GoogleTM of Mountain View, Calif.
- FIG. 1 illustrates an example block diagram of a Baseline system architecture in accordance with at least one aspect of the present invention. The stages of the system correspond to the different modules shown in FIG. 1 .
- the work flow progresses from a natural language question to a phrase extraction module 101 to generate a phrase from the natural language question.
- An outputted phrase is used to extract documents by a search module 103 .
- Outputted candidate sentences are found from the returned documents after proceeding through a sentence extraction module 105 . Potential answers of the candidate sentences are then extracted from an answer extraction module 107 . Finally, a list of the answer candidates is ranked by an answer ranking module 109 . Illustrative examples of each of the different modules are described more fully below.
- a user inputs a natural language question into the system.
- Phrase generation module 101 may be configured to generate a phrase from the natural language question.
- the phrase generation module 101 then passes the phrase to a search module 103 for the document retrieval.
- FIG. 2 illustrates an example block diagram of a Question Answering (QASR) system architecture in accordance with at least one aspect of the present invention.
- the QASR system uses two techniques of phrase generation, exact phrase and conjunction of sub-phrases/inexact phrase.
- the two techniques of phrase generation may be parts of a cascaded approach.
- the system may first attempt to find an answer using search results of an exact phrase technique. If an answer is not found from the exact phrase technique, an inexact phrase technique may then be utilized.
- a natural language question contains an auxiliary verb, such as is, was, were, are, did, do, or does
- a phrase is generated by dropping all words between the ‘wh’ word and the auxiliary verb and matching the tense of the predicate with the auxiliary verb.
- the natural language question “When did Vesuvius last erupt?” yields a phrase, “Vesuvius last erupted,” where the predicate “erupt” is changed to “erupted.”
- a phrase, “Martha Stewart lives in” is yielded.
- a phrase is generated from a natural language question using the output of a semantic role labeler, which identifies phrases as arguments of the predicate.
- the output of the semantic role labeler for the question, “Who invented the electric guitar?,” is [ARG0 Who] [TARGET invented] [ARG1 the electric guitar].
- the conjunction of the two identified phrases, except for the ‘wh’ word, “invented” and “the electric guitar” is used for the inexact search. If some words are not identified as phrases by the semantic role labeler, they may be added into conjunction as single words.
- Search module 103 may be configured to operate using a conventional search engine, such as the GoogleTM search engine.
- One or more web-sites or databases may be filtered to avoid using other question answering web-sites.
- the search is done using an exact phrase technique and/or a conjunction of sub-phrases/inexact phrase technique.
- the candidate documents are extracted, the documents are sent to the sentence extraction module 105 where the candidate documents are sentence split by a tool for extracting sentence split text from the web pages. AnswerBus by Zhiping Zheng and as found at http://www.answerbus.com/index.shtml provides such an example tool.
- Sentence extraction module 105 may utilize one of three methods of sentence extraction: exact phrase; conjunction of sub-phrases; or predicate.
- exact phrase search the candidate sentences are extracted with the exact phrase by scanning the candidate documents for the sentences with occurrences of the exact phrase.
- Predicate extraction may also be used by selecting all sentences containing the predicate.
- inexact search conjunction of the sub-phrases and predicate candidate sentence extraction may be utilized.
- an answer is expected to appear on either side of the search phrase depending on the question type.
- the phrase is, [ . . . ] “invented sillypuddy.”
- the answer is expected to appear right before the phrase. All words from the beginning of the sentence are considered an answer.
- a sentence splitting algorithm may be utilized. Such a method is applicable to an exact search/phrase sentence extraction method.
- Answer extraction module 107 filters out of all words that are “stop-words,” such as he, she, and who. A shortened version is generated for every candidate answer by removing spaces, punctuation, and stop-words. Such a mechanism allows similar answers, like “Beatles” and “the Beatles,” to be counted as the same answer. The confidence of an answer is a frequency of its shortened version with respect to all candidate answers. A threshold is used in the cascaded approach to eliminate the baseline answers with the confidence lower than the threshold. Answer ranking module 109 performs the corresponding ranking of answer candidates in accordance with any of a number of different ranking methods.
- semantic role labeler (SRL) modules 201 and 203 may be configured to replace the operation of answer extraction module 107 .
- Both a natural language question and candidate answers are labeled by one or more semantic role labeler modules 201 and/or 203 .
- the natural language question is parsed to find a predicate and the type of the argument of ‘wh’ word in the sentence.
- the accuracy of the SRL module 201 on the “wh” words may be limited and the labels of the “wh” words are not relied upon to determine the argument type of the candidate answer.
- a heuristics system for combining the type of the “wh” natural language question and question classification may be used.
- the natural language question may be parsed as:
- semantic role labeler 201 outputs [ARG0 Who] [TARGET won] [ARG1 the nobel prize] in literature [ARGM-TMP in 1988].
- “won” is identified as a predicate. Because the natural language question is identified as a “who” question, ARG0 is used as the argument type.
- semantic role labeler 203 identifies an ARG0 of the predicate “won.”
- the phrase automatically extracted by the phrase extraction module 101 from the natural language sentence “won the Nobel prize in literature in 1988,” is passed to the search module 103 to retrieve candidate documents.
- Candidate documents are sentence split in the sentence extraction module 105 and the sentences that contain all of the nonstop words from the phrase are considered as the candidate sentences containing the answer.
- the candidate sentences retrieved by search module may be:
- Both of the candidate sentences contain the correct answer, Mahfouz.
- a baseline system fails to find the answer as it expects it to appear right before the search phrase.
- the word “who” is assumed to be the answer by the baseline system, but in this example, this is a stop word and the system returns no answer.
- the candidate sentences are passed to the semantic role labeling module 203 which parses the sentences as:
- the two candidate answers identified are “the Egyptian” and “Mahfouz.” Both of the answers are technically correct, although one of them is more specific and considered correct according to the answer patterns, that being “Mahfouz.”
- a Semantic role labeling technique that may be applied to the candidate sentences.
- a cascaded approach may be used of a combination of an exact phrase search with the phrase extraction and/or a conjunction of sub-phrases/inexact search with predicate sentence extraction.
- Algorithms may be combined using a cascaded approach to maximize the Mean Reciprocal Ranking (MRR) performance of the system by combining the higher precision/lower recall version of the algorithm, SRL on exact, with the lower precision/higher recall version, SRL on inexact.
- MRR Mean Reciprocal Ranking
- Question Answering on a limited domain task assumes that the document set for question answering is fixed. This limits the question range as well as the redundancy of answer occurrences.
- a limited domain allows for the preprocessing of the data offline.
- data processing e.g., tagging and parsing
- Limited domain allows the processing of all sentences. Thus, candidate sentence selection is improved because now this decision is made using extra information.
- the natural language question and the candidate documents are sentence split and all of the data is labeled with the semantic roles.
- a score is assigned for each sentence in the data set. If a predicate or any of its synonyms is not present in the candidate sentence, a score of 0 is assigned.
- aspects of the present invention utilize factoid type Who/When/Where/What natural language questions containing a predicate.
- the system uses absolute accuracy and Mean Reciprocal Ranking (MRR), where a system provides multiple (5) answers and each question receives a score equal to inverse of the correct answer's index or zero if correct answer is not found.
- MRR Mean Reciprocal Ranking
- An evaluation script and correct answer patterns may be provided with the data to automatically evaluate the system.
- the first automatically-marked-correct answers may be reviewed manually and the percentage of these answers that contains irrelevant information or is not grammatically correct may be computed.
- One motivation behind such an evaluation is for the voice enabled QA system where irrelevant and ungrammatical answers may decrease the user's comprehension even in the case of the correct answer.
- sentence and answer extraction techniques for each module of the system may be evaluated separately.
- search result and sentence extraction are evaluated using pattern match of correct answers on the retrieved documents.
- Table 1 shows the results of one such example.
- the system is configurable for combining the two answer extraction algorithms, baseline and SRL, and the two document/sentence extraction methods, exact phrase and conjunction of sub-phrases/inexact phrase, in various ways.
- Table 2 presents a precision of correct answers and recall values for each of the systems.
- the precision of correct answers corresponds to the number of questions that have a correct answer in one of the five (5) top answers over the total questions answered.
- the recall corresponds to the number of questions answered over the total number of questions.
- the results show that a SRL system on the exactly matched candidates gives higher precision 0.84 and lower recall 0.44, while the SRL algorithm results in lower precision 0.42 and higher recall 0.82.
- a Baseline SRL search corr prec/recall corr prec/recall Exact 0.60/0.53 0.84/0.44 Inexact n/a 0.42/0.8 Exact + Inex n/a 0.53/0.82
- Table 3 presents an example table of the Mean Reciprocal Ranking (MRR) and precision values for the Baseline system, SRL system on exact match, SRL system on inexact match, and the combination of SRL on exact and inexact matched candidate sentences.
- MRR Mean Reciprocal Ranking
- the combination of the SRL on Exact Search and Inexact search raises the MRR to 0.35, compared with the baseline of 0.24.
- Such an increase is due to two factors, the MRR of the SRL system on exact search candidates is higher than the MRR of the baseline system on the exact match candidates, and the SRL system has higher coverage because it also uses the inexact match candidates to retrieve the answers.
- the Baseline algorithm does not address the inexact candidates because these candidates do not contain a match phrase.
- a 3-tier cascaded approach may be used as is illustrated in FIG. 4 .
- the 3-tier cascaded approach includes running a baseline approach in combination with the SRL approach.
- the Baseline approach is the fastest as it does not require deep semantic processing of the candidate answers.
- Such a 3-tier cascade approach may produces a similar MRR and precision result to a 2-tier cascade approach as illustrated in FIG. 3 but it decreases the conciseness and precision of the correct answers described in the following section.
- the average length of the answers may be computed in the number of words. As shown in Table 4, the results show that an SRL approach decreases the average length of an answer from 9.1 to 4.5 words per answer. Comparing the correct answers of the Baseline and the SRL systems, it is found that in 26% of the first correct answers on the Baseline system, some information irrelevant to the question is contained. Information may be considered irrelevant if it does not correspond to the question asked. In addition, 17% of the first correct answers of the Baseline system are grammatically incorrect.
- 7% of the first correct answers contain irrelevant information and 2% are grammatically incorrect.
- the answer to a natural language question “Where was Tesla born?” using a Baseline system is “to Serbia parents in smiljan, croatia”, and using SRL system is “in smiljan, croatia.”
- a Baseline system's answer contains the correct pattern “Croatia” as well as some irrelevant information “to Serbia parents.” This is not necessarily incorrect, but may affect the system performance when used via the voice interface.
- “smiljan” is not considered irrelevant as it is elaboration on the location.
- An example of a grammatically incorrect answer may be found from the natural language question, “Who painted titania?”
- a Baseline's answer is “had Manet.” Such a grammatically incorrect answer may affect user comprehension in the case of the voice enabled interface.
- a Baseline system extracts part of a sentence next to the search phrase, which may contain the correct answer along with other irrelevant information. Truncation of a sentence does not necessarily produce a grammatically correct sentence, while if the semantic role parse is correct, a complete phrase should be identified and the final answer is more likely to be concise and grammatically correct. In addition to improving the automatically evaluated system performance, an SRL approach produces more concise and grammatically correct answers.
- FIG. 5 is a flow chart of an illustrative example of a method for determining a factoid answer to a natural language question in accordance with at least one aspect of the present invention.
- the process starts at step 501 where a natural language question is received by the system. Proceeding to step 503 , semantic role labeling in accordance with aspects of the present invention are applied to the natural language question and a predicate and searched argument are outputted at step 505 before the process proceeds to step 515 .
- step 507 the natural language question is analyzed to extract a phrase.
- the extracted phrase is used as part of a search of an open domain source, such as the Internet. It should be understood that a fixed domain source may also be used and that the present invention is not so limited to an open domain source.
- candidate documents are determined from the search and then candidate sentences are extracted from the candidate documents at step 513 before the process proceeds to step 515 .
- step 515 semantic role labeling is applied to the candidate sentences based upon the predicate and the searched argument.
- Candidate answers are then outputted in step 517 .
- An answer ranking module may then rank the candidate answers in step 519 based upon any of a number of different methods for ranking data.
- step 521 the highest ranking candidate answer is outputted as the answer to the factoid natural language question.
- FIGS. 6A-6C are a flowchart of an illustrative example of a method for determining a searched argument classification corresponding to a natural language question in accordance with at least one aspect of the present invention. Aspects of the description of FIGS. 6A-6C may be utilized within step 503 of FIG. 5 .
- analysis of the natural language factoid question is initiated.
- a determination is then made at step 603 as to whether the natural language question is a “who” question, meaning, does the question begin with “who.” If the natural language question is a “who” question, the process moves to step 605 where the search argument is classified as ARG0 and the process ends. If the natural language question is not a “who” question, the process proceeds to step 607 .
- step 627 the search argument is classified as ARG1 and the process ends. If the question class is not HUM and/or the question does not contain do, did, or does, the process proceeds to step 629 .
- step 629 a determination is made as to whether the question class is identified as HUM and does the question not contain do, did, or does, meaning does the “what” question seek a human for an answer and is the verb something other than do, did, or does. If the question class is HUM and the question does not contain do, did, or does, the process proceeds to step 631 where the search argument is classified as ARG0 and the process ends. If the question class is not HUM and/or the question does contain do, did, or does, the process proceeds to step 633 as shown in FIG. 6C .
- step 633 a determination is made as to whether the question class is identified as ENTY, meaning does the “what” question seek an entity for an answer. If the question class is ENTY, the process proceeds to step 635 where the search argument is classified as ARG1 and the process ends. If the question class is not ENTY, the process proceeds to step 637 .
- step 637 a determination is made as to whether the question class is identified as ABBR or DESC, meaning does the “what” question seek an answer for an abbreviation of description. If the question class is ABBR or DESC, the process proceeds to step 639 where the search argument is classified as ARG2 and the process ends. If the question class is not ABBR or DESC, the process proceeds to step 641 .
- step 641 a determination is made as to whether the question class is identified as UNMARKED and whether the question contains state, city, or country, meaning does the “what” question seek a city, state, or country for an answer. If the question class is UNMARKED and the question contains state, city, or country, the process proceeds to step 643 where the search argument is classified as ARGM-LOC and the process ends. If the question class is not UNMARKED and/or the question does not contain state, city, or country, the process proceeds to step 645 where the search argument is classified as ARG0 before the process ends.
Abstract
Description
- The present application is a continuation of U.S. patent application Ser. No. 11/319,188, filed Dec. 28, 2005, which claims priority to provisional U.S. application Ser. No. 60/740,632, filed Nov. 30, 2005, herein incorporated by reference in its entirety for all purposes.
- The invention relates generally to communication systems. More specifically, the invention relates to finding factoid answers to natural language questions within an electronic operating environment.
- As the use of computers has increased in the work environment as well as home, the demand for faster and more accurate information has followed. Users want information at their fingertips and they want the information to be correct the first time. However, with the introduction and boom of the Internet, users can be overwhelmed in information, much of it highly irrelevant. A user is then left with the task of finding the proverbial needle in the haystack of information they seek.
- Search engines have become one tool for discriminating mounds of information. Conventional systems allow a user to enter keywords, and the engines search the vast amounts of data to find corresponding results. The results may be based upon the frequency that other users access that data or it may be based upon the approximated correlation between the keyword(s) searched and the different data. In short, search engines provide a means for narrowing the amount of data a user must review.
- Some specific types of search engines allow a user to answer specific questions. For example, the web site “Ask Jeeves,” by Ask Jeeves, Inc. of Oakland, Calif., allows a user to type in a question and the system will perform a search and retrieval function to attempt to find an answer to the question. In operation, such a web site takes an input question of a user, such as “What is the capital of California?,” searches the Internet, and retrieves web pages that may include the answer to the question. As part of the retrieval process, an introduction portion of the web site may be included on a results page. However, a concise answer is not provided alone. The answer is provided within link to a web site. In addition, numerous irrelevant links to web pages are retrieved, requiring a user to wade through the data and perform her personal determination of the answer.
- Question answering (QA) is the task of finding a concise answer to a Natural Language question. Question answering systems use a search engine but differ from a web search task by the type of its input and output. Input to a search engine is a query while a QA system takes a natural language question as an input. For example, if a user wants to find out, “Who first broke the sound barrier?”, the user of a search engine would convert the question into a query like first AND broke AND “sound barrier,” while the user of a QA system would type or say the question. The output of the QA system is a concise answer: Yeager, in the case of the above question while the search engine produces snippets with link to the web pages.
- Providing a concise and accurate answer to a natural language question allows a user to retrieve an answer and continue with her work more efficiently. Semantic role labeling of a natural language factoid question allows for such a result. In accordance with aspects of the present invention, semantic role labeling is applied to the QA task for factoid questions, answers to a factoid question being a single word or a short phrase. For a baseline system approach, the traditional QA system modular architecture consisting of question classification, query generation, search, answer extraction, and re-ranking may be utilized.
- This Summary is provided to introduce a selection of concepts in a simplified form that is further described below in the Detailed Description. The Summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter.
- The following detailed description of illustrative embodiments is better understood when read in conjunction with the accompanying drawings, which are included by way of example, and not by way of limitation with regard to the claimed invention.
-
FIG. 1 illustrates an example block diagram of a Baseline system architecture in accordance with at least one aspect of the present invention; -
FIG. 2 illustrates an example block diagram of a question answering (QASR) system architecture in accordance with at least one aspect of the present invention; -
FIG. 3 an example block diagram of a 2-tier cascaded approach in accordance with at least one aspect of the present invention; -
FIG. 4 is an example block diagram of a 3-tier cascaded approach in accordance with at least one aspect of the present invention; -
FIG. 5 is a flow chart of an illustrative example of a method for determining a factoid answer to a natural language question in accordance with at least one aspect of the present invention; and -
FIGS. 6A-6C are a flowchart of an illustrative example of a method for determining a searched argument classification corresponding to a natural language question in accordance with at least one aspect of the present invention. - In the following description of various illustrative embodiments, reference is made to the accompanying drawings, which form a part hereof, and in which is shown, by way of illustration, various embodiments in which the invention may be practiced. It is to be understood that other embodiments may be utilized and structural and functional modifications may be made without departing from the scope of the present invention.
- In accordance with at least one aspect of the present invention, a system may comprise one or more memories configured to store computer readable instructions as described herein. There memories are computer-readable media having computer-executable commands for performing one or more methods. In one or more configurations, these computer readable instructions may be software modules. In addition, the system may further comprise one or more processors configured to execute the computer readable instructions stored on the one or more memories. In one or more configurations of the present invention, a computer may be used to implement one or more aspects of the present invention as described herein.
- In accordance with at least one aspect of the present invention, semantic role labeling aims to identify the predicate/argument relations within a sentence. In accordance with aspects of the present invention, question and candidate sentences extracted using a search engine are processed to identify predicate/argument structure. In accordance with one embodiment, the semantic role labeler described in Pradhan et al., Semantic Role Labeling Using Different Syntactic Views, In Proceedings of the Association for Computational Linguistics 43rd annual meeting (ACL-2005) may be used. However, it should be understood by those skilled in the art that the present invention is not so limited to any one semantic role labeler and that any number of different semantic role labelers may be utilized in accordance with aspects of the present invention. As used herein, semantic role labeling refers to techniques for classifying word(s) or portions of a sentence into object and predicate designations. Those skilled in the art include individuals with an understanding of ontologies and semantic principles of languages.
- Tags are assigned to predicates and argument phrases. For example, a TARGET tag is assigned to the predicate and ARG0, ARG1, ARG2 . . . , ARGM-TMP, and ARGM-LOC tags are assigned to arguments. The meaning of an argument type depends on a class of the predicate that it appears with, but generally ARG0 serves as an agent and ARG1, ARG2 . . . are objects. Arguments of the type ARGM-TMP, which represents temporal argument, and ARGM-LOC, which represents a location argument, are shared over all predicates.
- Application of semantic role labeling may be seen in the following example. A sentence “Nostradamus was born in 1503 in the south of France” is tagged by a semantic role labeling program as [ARG1 Nostradamus] was [TARGET born] [ARGM-TMP in 1503] [ARGM-LOC in the south of France]. In this example, “born” is identified as a target predicate with three arguments, object “Nostradamus”, temporal argument “in 1503,” and location argument “in the south of France.” The ability to extract the predicate-argument structure is a technique useful in many applications.
- In accordance with aspects of the present invention, a question answering system that uses semantic role labeling is utilized and evaluated using a combination of automatic and manual evaluation on a set of Who/When/Where/What questions. The searched argument of the question is heuristically determined and this argument is extracted from the candidate sentences where the argument is applied to the matching predicate. To demonstrate the predicate/argument extraction for the QA task, consider a question “Who created a comic strip Garfield?” and a candidate sentence: “Garfield is a popular comic strip created by Jim Davis featuring the cat Garfield . . . ” Without deep semantic processing and finding predicate/argument relations, one could extract the answer “Jim Davis” by creating an example-specific template. However, it is not feasible to create templates for each anticipated predicate/answer candidate pair because the number of predicates covered by an open-domain question answering system is unlimited as well as the variation of candidate sentences.
- With the knowledge of the predicate/argument structure identified by the semantic role labeler, the candidate sentence is tagged. For example, “Garfield is a popular comic strip created by Jim Davis featuring the cat Garfield . . . ” becomes analyzed as Garfield is [ARG1 a popular comic strip] [TARGET created] [ARG0 by Jim Davis] featuring the cat Garfield . . . In this example, ARG0 corresponds to the “agent,” and the answer to the “who” question is expected to be an agent. Therefore, we can extract Jim Davis as the answer.
- A predicate/argument extraction approach using semantic role labeling has been successfully applied to the task of spoken language understanding where predicate-argument extraction eliminates the need for the costly data labeling in the call-type classification system, achieving the performance close to the one of the manually labeled system. The method has successfully been applied to an Information Extraction task where, again, using predicate/argument extraction, the system's performance is close to the one using manually created patterns.
- Some Question Answering system focus on using a fixed size document set to find answers, while others use the Internet as a resource. A fixed set of documents is generally more reliable and accurate, as one can choose which documents to use, but it often lacks redundancy. The Internet, in the alternative, lacks accuracy while providing a greater redundancy. Redundancy of the Internet information benefits a QA system because evaluation of the system occurs on a range of the number of snippets used for the answer extraction. Finding an answer in a small set of documents is a harder problem because the questions are less likely to appear in the form of the question phrase and it requires more careful analysis of the question and query generation in order to widen the range of potential answers in order to achieve a similar performance as using Internet data.
- Some systems incorporate a combination of sophisticated NLP (natural language processing) techniques including syntactic parsing, named entity detection, lexical chains derived from extended WordNet ontologies, and logical prover. Predicate/argument identification in the question and the candidate answer improves the performance of the answers extraction module from the candidate sentences. A grammatic relations approach aims at the extraction of the answers from candidate sentences. Grammatic relations extraction may be based on the grammatic annotations from the Penn Treebank. Semantic roles in a sentence represent a predicate/argument structure while grammatic roles are an approximation to it.
- Other systems match questions and answers on the level of syntactic relations. Their approach uses linguistically-uninformed techniques as a foundation of the Question Answering while sophisticated NLP processing is applied in the cases known to improve the performance. Linguistic phenomena that are difficult to handle for the linguistically-uninformed techniques have been identified. They are semantic symmetry and ambiguous modification.
- Semantic symmetry occurs when the meaning of a sentence can not be derived using lexical information alone. For example, the sentences “The bird ate the snake” and “The snake ate the bird” are similar lexically as both of the nouns are animate objects. However, without knowing the syntactic information that the bird is the subject and the snake is the object, it is difficult to determine automatically who ate whom. Ambiguous modification occurs when it is difficult to determine which adjective is being modified. For example, in the sentence, “the planet's largest volcano,” a system must be able to discern whether “largest” applies to the planet or to the volcano.
- In accordance with aspects of the present invention, the Internet is used as a data source for search and extraction of candidate documents. To find the candidate documents, a search engine, such as the search engine by Google™ of Mountain View, Calif., may be utilized. It should be understood by those skilled in the art that the present invention is not so limited to any particular search engine and that any number of different search engines for use with the Internet may be utilized.
FIG. 1 illustrates an example block diagram of a Baseline system architecture in accordance with at least one aspect of the present invention. The stages of the system correspond to the different modules shown inFIG. 1 . The work flow progresses from a natural language question to aphrase extraction module 101 to generate a phrase from the natural language question. An outputted phrase is used to extract documents by asearch module 103. Outputted candidate sentences are found from the returned documents after proceeding through asentence extraction module 105. Potential answers of the candidate sentences are then extracted from ananswer extraction module 107. Finally, a list of the answer candidates is ranked by ananswer ranking module 109. Illustrative examples of each of the different modules are described more fully below. - In accordance with aspects of the present invention, a user inputs a natural language question into the system.
Phrase generation module 101 may be configured to generate a phrase from the natural language question. Thephrase generation module 101 then passes the phrase to asearch module 103 for the document retrieval.FIG. 2 illustrates an example block diagram of a Question Answering (QASR) system architecture in accordance with at least one aspect of the present invention. The QASR system uses two techniques of phrase generation, exact phrase and conjunction of sub-phrases/inexact phrase. The two techniques of phrase generation may be parts of a cascaded approach. In one embodiment, the system may first attempt to find an answer using search results of an exact phrase technique. If an answer is not found from the exact phrase technique, an inexact phrase technique may then be utilized. - For an exact phrase technique, the following illustrative heuristics may be utilized by the system. If a natural language question contains an auxiliary verb, such as is, was, were, are, did, do, or does, a phrase is generated by dropping all words between the ‘wh’ word and the auxiliary verb and matching the tense of the predicate with the auxiliary verb. For example, the natural language question, “When did Vesuvius last erupt?” yields a phrase, “Vesuvius last erupted,” where the predicate “erupt” is changed to “erupted.” For the example sentence, “What state does Martha Stewart live in?,” a phrase, “Martha Stewart lives in” is yielded. The word “state,” which appears between “What” and “does,” is dropped. If the natural language question does not contain an auxiliary verb, a phrase is generated by dropping the ‘wh’ word. For example, the natural language question “Who invented the radio?,” yields the exact phrase “invented the radio.”
- For a conjunction of sub-phrases/inexact phrase technique, a phrase is generated from a natural language question using the output of a semantic role labeler, which identifies phrases as arguments of the predicate. The output of the semantic role labeler for the question, “Who invented the electric guitar?,” is [ARG0 Who] [TARGET invented] [ARG1 the electric guitar]. The conjunction of the two identified phrases, except for the ‘wh’ word, “invented” and “the electric guitar” is used for the inexact search. If some words are not identified as phrases by the semantic role labeler, they may be added into conjunction as single words.
-
Search module 103 may be configured to operate using a conventional search engine, such as the Google™ search engine. One or more web-sites or databases may be filtered to avoid using other question answering web-sites. The search is done using an exact phrase technique and/or a conjunction of sub-phrases/inexact phrase technique. After the candidate documents are extracted, the documents are sent to thesentence extraction module 105 where the candidate documents are sentence split by a tool for extracting sentence split text from the web pages. AnswerBus by Zhiping Zheng and as found at http://www.answerbus.com/index.shtml provides such an example tool. -
Sentence extraction module 105 may utilize one of three methods of sentence extraction: exact phrase; conjunction of sub-phrases; or predicate. In the case of the exact phrase search, the candidate sentences are extracted with the exact phrase by scanning the candidate documents for the sentences with occurrences of the exact phrase. Predicate extraction may also be used by selecting all sentences containing the predicate. In the case of the inexact search, conjunction of the sub-phrases and predicate candidate sentence extraction may be utilized. - In the baseline approach, an answer is expected to appear on either side of the search phrase depending on the question type. For example, for the natural language question, “Who invented sillypuddy?,” the phrase is, [ . . . ] “invented sillypuddy.” In this example, the answer is expected to appear right before the phrase. All words from the beginning of the sentence are considered an answer. In this case, a sentence splitting algorithm may be utilized. Such a method is applicable to an exact search/phrase sentence extraction method.
-
Answer extraction module 107 filters out of all words that are “stop-words,” such as he, she, and who. A shortened version is generated for every candidate answer by removing spaces, punctuation, and stop-words. Such a mechanism allows similar answers, like “Beatles” and “the Beatles,” to be counted as the same answer. The confidence of an answer is a frequency of its shortened version with respect to all candidate answers. A threshold is used in the cascaded approach to eliminate the baseline answers with the confidence lower than the threshold. Answer rankingmodule 109 performs the corresponding ranking of answer candidates in accordance with any of a number of different ranking methods. - In accordance with one embodiment and the description in
FIG. 2 , semantic role labeler (SRL)modules answer extraction module 107. Both a natural language question and candidate answers are labeled by one or more semanticrole labeler modules 201 and/or 203. Initially, the natural language question is parsed to find a predicate and the type of the argument of ‘wh’ word in the sentence. The accuracy of theSRL module 201 on the “wh” words may be limited and the labels of the “wh” words are not relied upon to determine the argument type of the candidate answer. Instead, a heuristics system for combining the type of the “wh” natural language question and question classification may be used. For example, in accordance with one embodiment of the present invention, the natural language question may be parsed as: -
- If the natural language question is a “Who” question, use ARG0 as the argument type. For example, “Who wrote Pride and Prejudice?,” is parsed as [ARG0 Who] [TARGET wrote] [ARG1 War and Peace].
- If the natural language question is a “When” question, use ARGM-TMP as the argument type. For example, “When was the Constitution ratified?,” is parsed as [ARGM-TMP When] [TARGET was ratified] [ARG1 the Constitution].
- If the natural language question is a “Where” question, use ARGM-LOC as the argument type. For example, “Where was Lincoln shot?,” is parsed as [ARGM-LOC Where] [TARGET was shot] [ARG1 Lincoln].
- If the natural language question is a “What” question and the Question Class is LOC, use ARGM-LOC as the argument type. For example, “What is the capital of Texas?,” is parsed as [ARGM-LOC What] [TARGET is] [ARG1 the capital of Texas].
- If the natural language question is a “What” question and the Question Class is “NUM_date”, use ARGM-TMP as the argument type. For example, “What year was Grant born?,” is parsed as [ARGM-TMP What] [TARGET was born] [ARG1 year].
- If the natural language question is a “What” question, the Question Class is “HUM”, and the natural language question contains did, do, or does, use ARG1 as the argument type. For example, “What actor did the jury nominate?,” is parsed as [ARG1 What] [TARGET nominated] [ARG0 the jury].
- If the natural language question is a “What” question, the Question Class is “HUM”, and the natural language question does not contain did, do, or does, use ARG0 as the argument type. For example, “What actor received the prize?,” is parsed as [ARG0 What] [TARGET received] [ARG1 the prize].
- If the natural language question is a “What” question and the Question Class is ENTY, use ARG1 as the argument type. For example, “What do you call a professional map drawer?,” is parsed as [ARG1 What] [TARGET is called] [ARG0 a professional map drawer].
- If the natural language question is a “What” question and the Question Class is ABBR or DESC, use ARG2 as the argument type. For example, “What does DEC stand for?,” is parsed as [ARG2 What] [TARGET stands for] [ARG0 DEC].
- If the natural language question is a “What” question, the Question Class is UNMARKED, and the natural language question has state, city, or country, use ARGM-LOC as the argument type. For example, “What state is Boston in?,” is parsed as [ARGM-LOC What] [TARGET is in] [ARG1 state].
- In all other cases, use ARG0 as the argument type. For example, “What is 4 plus 4/,” is parsed as [ARG0 What] [TARGET is] [ARG1 4 plus 4].
- So, if a user enters a natural language question, “Who won the Nobel prize in literature in 1988?,”
semantic role labeler 201 outputs [ARG0 Who] [TARGET won] [ARG1 the nobel prize] in literature [ARGM-TMP in 1988]. “won” is identified as a predicate. Because the natural language question is identified as a “who” question, ARG0 is used as the argument type. When analyzing candidate sentences,semantic role labeler 203 identifies an ARG0 of the predicate “won.” - The phrase, automatically extracted by the
phrase extraction module 101 from the natural language sentence “won the Nobel prize in literature in 1988,” is passed to thesearch module 103 to retrieve candidate documents. Candidate documents are sentence split in thesentence extraction module 105 and the sentences that contain all of the nonstop words from the phrase are considered as the candidate sentences containing the answer. - With respect to this example of “Who won the Nobel prize in literature in 1988?,” the candidate sentences retrieved by search module may be:
-
- “The veritable terrorism of which he is a target is unjustifiable, indefensible,” wrote Naguib Mahfouz, the Egyptian who won the Nobel Prize in Literature in 1988.
- Mahfouz, who won the Nobel Prize in Literature in 1988, is known for his intimate stories of Egyptian life. The Cairo Trilogy is his most celebrated work and his national prestige is absolute, like Victor Hugo is in nineteenth-century Paris.
- Both of the candidate sentences contain the correct answer, Mahfouz. A baseline system fails to find the answer as it expects it to appear right before the search phrase. The word “who” is assumed to be the answer by the baseline system, but in this example, this is a stop word and the system returns no answer.
- In accordance with aspects of the present invention, the candidate sentences are passed to the semantic
role labeling module 203 which parses the sentences as: -
- “The veritable terrorism of which he is a target is unjustifiable, indefensible,” wrote Naguib Mahfouz [ARG0 the Egyptian] [R-ARG0 who] [TARGET won] [ARG1 the Nobel Prize] [ARGM-LOC in Literature in 1988].
- [ARG0 Mahfouz] [R-ARG0 who] [TARGET won] [ARG1 the Nobel Prize in Literature] [ARGM-TMP in 1988] is known for his intimate stories of Egyptian life. The Cairo Trilogy is his most celebrated work and his national prestige is absolute like Victor Hugo is in nineteenth-century Paris.
- The two candidate answers identified are “the Egyptian” and “Mahfouz.” Both of the answers are technically correct, although one of them is more specific and considered correct according to the answer patterns, that being “Mahfouz.”
- Aspects of the present invention provide a Semantic role labeling technique that may be applied to the candidate sentences. As necessary, a cascaded approach may be used of a combination of an exact phrase search with the phrase extraction and/or a conjunction of sub-phrases/inexact search with predicate sentence extraction. Algorithms may be combined using a cascaded approach to maximize the Mean Reciprocal Ranking (MRR) performance of the system by combining the higher precision/lower recall version of the algorithm, SRL on exact, with the lower precision/higher recall version, SRL on inexact. The cascaded approach is depicted in
FIG. 3 . First, an answer extraction algorithm on exact phrase search results is run. If ‘no answer’ is returned, an SRL algorithm on conjunction of sub-phrases/inexact search result is run. - Question Answering on a limited domain task assumes that the document set for question answering is fixed. This limits the question range as well as the redundancy of answer occurrences. At the same time, a limited domain allows for the preprocessing of the data offline. In the case of using the Internet as a domain, data processing, e.g., tagging and parsing, has to happen online on the selected sentences. Limited domain allows the processing of all sentences. Thus, candidate sentence selection is improved because now this decision is made using extra information.
- Under a limited domain approach, the natural language question and the candidate documents are sentence split and all of the data is labeled with the semantic roles. A score is assigned for each sentence in the data set. If a predicate or any of its synonyms is not present in the candidate sentence, a score of 0 is assigned.
- Aspects of the present invention utilize factoid type Who/When/Where/What natural language questions containing a predicate. The system uses absolute accuracy and Mean Reciprocal Ranking (MRR), where a system provides multiple (5) answers and each question receives a score equal to inverse of the correct answer's index or zero if correct answer is not found. An evaluation script and correct answer patterns may be provided with the data to automatically evaluate the system. The first automatically-marked-correct answers may be reviewed manually and the percentage of these answers that contains irrelevant information or is not grammatically correct may be computed. One motivation behind such an evaluation is for the voice enabled QA system where irrelevant and ungrammatical answers may decrease the user's comprehension even in the case of the correct answer. In order to select an algorithm for combining search methods, sentence and answer extraction techniques for each module of the system may be evaluated separately.
- In one embodiment of the present invention, search result and sentence extraction are evaluated using pattern match of correct answers on the retrieved documents. Table 1 shows the results of one such example.
-
TABLE 1 Evaluation of the Search and Sentence Extraction methods: Proportion of number of sentences containing a correct answer/proportion of the questions containing zero correct answers in the candidates Search Sentence Sentence Result Extraction Extraction Sentence Search/Extraction (phrase) (conj. of subph) (predicate) Extraction Exact phrase 8419/76 .30/.51 n/a 0.05/.46 Conjunction of 18814/43 n/a .18/.51 0.08/.35 sub-phrases - This analysis detects that an exact phrase search technique using exact phrase extraction produces the highest precision, 30% of the candidate sentences contain the correct answer, but low recall, only 49% of the questions contain at least one correct answer in the candidate sentences. A conjunction of sub-phrases/inexact search technique with predicate method of sentence extraction has the highest recall, 69% of the candidate sentences contain a correct answer, but only 10% of all candidate sentences have a correct answer. These numbers correspond to the precision and recall of the system performance on exact/inexact search results discussed in the next section and serve as a motivation for the cascaded approach which combines higher precision/lower recall version of the algorithm, SRL on exact, with the lower precision/higher recall version, SRL on inexact.
- The system is configurable for combining the two answer extraction algorithms, baseline and SRL, and the two document/sentence extraction methods, exact phrase and conjunction of sub-phrases/inexact phrase, in various ways. Table 2 presents a precision of correct answers and recall values for each of the systems. The precision of correct answers corresponds to the number of questions that have a correct answer in one of the five (5) top answers over the total questions answered. The recall corresponds to the number of questions answered over the total number of questions. The results show that a SRL system on the exactly matched candidates gives higher precision 0.84 and lower recall 0.44, while the SRL algorithm results in lower precision 0.42 and higher recall 0.82.
-
TABLE 2 Evaluation precision and recall: “corr prec” stands for the Precision of correct answers = number of questions that have a correct answers in one of the 5 top answers/total questions answered. Recall is the number of questions answered over the total number of questions. A Baseline SRL search corr prec/recall corr prec/recall Exact 0.60/0.53 0.84/0.44 Inexact n/a 0.42/0.8 Exact + Inex n/a 0.53/0.82 - Table 3 presents an example table of the Mean Reciprocal Ranking (MRR) and precision values for the Baseline system, SRL system on exact match, SRL system on inexact match, and the combination of SRL on exact and inexact matched candidate sentences. The combination of the SRL on Exact Search and Inexact search raises the MRR to 0.35, compared with the baseline of 0.24. Such an increase is due to two factors, the MRR of the SRL system on exact search candidates is higher than the MRR of the baseline system on the exact match candidates, and the SRL system has higher coverage because it also uses the inexact match candidates to retrieve the answers. The Baseline algorithm does not address the inexact candidates because these candidates do not contain a match phrase.
-
TABLE 3 Evaluation of the Search and Answer Extraction methods: MRR/precision. Cascaded approach combining baseline, SRL on exact and inexact searched sentences gives the best result MRR = 0.35 /answer Baseline SRL Baseline + SRL extr search MRR/Prec. MRR/Prec. MRR/Prec. Exact 0.24/0.19 0.29/0.24 .026/0.2 Inexact n/a 0.23/0.16 n/a Exact + Inex n/a .035/0.3 0.35/0.3 - In accordance with one embodiment, a 3-tier cascaded approach may be used as is illustrated in
FIG. 4 . The 3-tier cascaded approach includes running a baseline approach in combination with the SRL approach. The Baseline approach is the fastest as it does not require deep semantic processing of the candidate answers. Such a 3-tier cascade approach may produces a similar MRR and precision result to a 2-tier cascade approach as illustrated inFIG. 3 but it decreases the conciseness and precision of the correct answers described in the following section. - The average length of the answers may be computed in the number of words. As shown in Table 4, the results show that an SRL approach decreases the average length of an answer from 9.1 to 4.5 words per answer. Comparing the correct answers of the Baseline and the SRL systems, it is found that in 26% of the first correct answers on the Baseline system, some information irrelevant to the question is contained. Information may be considered irrelevant if it does not correspond to the question asked. In addition, 17% of the first correct answers of the Baseline system are grammatically incorrect.
- Under the SRL system, 7% of the first correct answers contain irrelevant information and 2% are grammatically incorrect. For example, the answer to a natural language question “Where was Tesla born?” using a Baseline system is “to Serbia parents in smiljan, croatia”, and using SRL system is “in smiljan, croatia.” A Baseline system's answer contains the correct pattern “Croatia” as well as some irrelevant information “to Serbia parents.” This is not necessarily incorrect, but may affect the system performance when used via the voice interface. In this example, “smiljan” is not considered irrelevant as it is elaboration on the location. An example of a grammatically incorrect answer may be found from the natural language question, “Who painted Olimpia?” A Baseline's answer is “had Manet.” Such a grammatically incorrect answer may affect user comprehension in the case of the voice enabled interface.
- A Baseline system extracts part of a sentence next to the search phrase, which may contain the correct answer along with other irrelevant information. Truncation of a sentence does not necessarily produce a grammatically correct sentence, while if the semantic role parse is correct, a complete phrase should be identified and the final answer is more likely to be concise and grammatically correct. In addition to improving the automatically evaluated system performance, an SRL approach produces more concise and grammatically correct answers.
-
TABLE 4 Manual Evaluation of the Irrelevant Information, Ungrammatical Structure and Average Length of Correct Answers Baseline SRL on exact total 35 total 44 System correct correct Contains irrelevant information 0.26 0.07 Not Grammatically correct 0.17 0.02 Average answer length (in words) 9.1 4.5 -
FIG. 5 is a flow chart of an illustrative example of a method for determining a factoid answer to a natural language question in accordance with at least one aspect of the present invention. The process starts atstep 501 where a natural language question is received by the system. Proceeding to step 503, semantic role labeling in accordance with aspects of the present invention are applied to the natural language question and a predicate and searched argument are outputted atstep 505 before the process proceeds to step 515. - Concurrent to step 503, at
step 507, the natural language question is analyzed to extract a phrase. Proceeding to step 509, the extracted phrase is used as part of a search of an open domain source, such as the Internet. It should be understood that a fixed domain source may also be used and that the present invention is not so limited to an open domain source. Atstep 511, candidate documents are determined from the search and then candidate sentences are extracted from the candidate documents atstep 513 before the process proceeds to step 515. - At
step 515, semantic role labeling is applied to the candidate sentences based upon the predicate and the searched argument. Candidate answers are then outputted instep 517. An answer ranking module may then rank the candidate answers instep 519 based upon any of a number of different methods for ranking data. Finally, atstep 521, the highest ranking candidate answer is outputted as the answer to the factoid natural language question. -
FIGS. 6A-6C are a flowchart of an illustrative example of a method for determining a searched argument classification corresponding to a natural language question in accordance with at least one aspect of the present invention. Aspects of the description ofFIGS. 6A-6C may be utilized withinstep 503 ofFIG. 5 . As the process begins inFIG. 6A , atstep 601, analysis of the natural language factoid question is initiated. A determination is then made atstep 603 as to whether the natural language question is a “who” question, meaning, does the question begin with “who.” If the natural language question is a “who” question, the process moves to step 605 where the search argument is classified as ARG0 and the process ends. If the natural language question is not a “who” question, the process proceeds to step 607. - At
step 607, a determination is made as to whether the natural language question is a “when” question, meaning, does the question begin with “when.” If the natural language question is a “when” question, the process moves to step 609 where the search argument is classified as ARGM-TMP and the process ends. If the natural language question is not a “when” question, the process proceeds to step 611. A determination is made atstep 611 as to whether the natural language question is a “where” question, meaning, does the question begin with “where.” If the natural language question is a “where” question, the process moves to step 613 where the search argument is classified as ARGM-LOC and the process ends. If the natural language question is not a “where” question, the process proceeds to step 615. - At
step 615, a determination is made as to whether the natural language question is a “what” question, meaning, does the question begin with “what.” If the natural language question is a “what” question, the process moves to step 617 as described inFIG. 6B . If the natural language question is not a “what” question, the process ends. As shown inFIG. 6B , atstep 617, a determination is made as to whether the question class is identified as LOC, meaning does the “what” question seek a location answer. If the question class is LOC, the process proceeds to step 619 where the search argument is classified as ARGM-LOC and the process ends. If the question class is not LOC, the process proceeds to step 621. - At
step 621, a determination is made as to whether the question class is identified as NUM_date, meaning does the “what” question seek a temporal answer. If the question class is NUM_date, the process proceeds to step 623 where the search argument is classified as ARGM-TMP and the process ends. If the question class is not NUM_date, the process proceeds to step 625. Atstep 625, a determination is made as to whether the question class is identified as HUM and does the question contain do, did, or does, meaning does the “what” question seek a human for an answer and is the verb do, did, or does. If the question class is HUM and the question does contain do, did, or does, the process proceeds to step 627 where the search argument is classified as ARG1 and the process ends. If the question class is not HUM and/or the question does not contain do, did, or does, the process proceeds to step 629. - At
step 629, a determination is made as to whether the question class is identified as HUM and does the question not contain do, did, or does, meaning does the “what” question seek a human for an answer and is the verb something other than do, did, or does. If the question class is HUM and the question does not contain do, did, or does, the process proceeds to step 631 where the search argument is classified as ARG0 and the process ends. If the question class is not HUM and/or the question does contain do, did, or does, the process proceeds to step 633 as shown inFIG. 6C . - At
step 633, a determination is made as to whether the question class is identified as ENTY, meaning does the “what” question seek an entity for an answer. If the question class is ENTY, the process proceeds to step 635 where the search argument is classified as ARG1 and the process ends. If the question class is not ENTY, the process proceeds to step 637. Atstep 637, a determination is made as to whether the question class is identified as ABBR or DESC, meaning does the “what” question seek an answer for an abbreviation of description. If the question class is ABBR or DESC, the process proceeds to step 639 where the search argument is classified as ARG2 and the process ends. If the question class is not ABBR or DESC, the process proceeds to step 641. - Proceeding to step 641, a determination is made as to whether the question class is identified as UNMARKED and whether the question contains state, city, or country, meaning does the “what” question seek a city, state, or country for an answer. If the question class is UNMARKED and the question contains state, city, or country, the process proceeds to step 643 where the search argument is classified as ARGM-LOC and the process ends. If the question class is not UNMARKED and/or the question does not contain state, city, or country, the process proceeds to step 645 where the search argument is classified as ARG0 before the process ends.
- While illustrative systems and methods as described herein embodying various aspects of the present invention are shown, it will be understood by those skilled in the art, that the invention is not limited to these embodiments. Modifications may be made by those skilled in the art, particularly in light of the foregoing teachings. For example, each of the elements of the aforementioned embodiments may be utilized alone or in combination or subcombination with elements of the other embodiments. It will also be appreciated and understood that modifications may be made without departing from the true spirit and scope of the present invention. The description is thus to be regarded as illustrative instead of restrictive on the present invention.
Claims (20)
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US14/479,765 US20150052113A1 (en) | 2005-11-30 | 2014-09-08 | Answer Determination for Natural Language Questioning |
Applications Claiming Priority (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US74063205P | 2005-11-30 | 2005-11-30 | |
US11/319,188 US8832064B2 (en) | 2005-11-30 | 2005-12-28 | Answer determination for natural language questioning |
US14/479,765 US20150052113A1 (en) | 2005-11-30 | 2014-09-08 | Answer Determination for Natural Language Questioning |
Related Parent Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US11/319,186 Continuation US20060157888A1 (en) | 2005-01-14 | 2005-12-28 | Process for producing plantar orthoses or insoles taking a footprint while walking and directly molding |
Publications (1)
Publication Number | Publication Date |
---|---|
US20150052113A1 true US20150052113A1 (en) | 2015-02-19 |
Family
ID=37865854
Family Applications (2)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US11/319,188 Active 2026-07-30 US8832064B2 (en) | 2005-11-30 | 2005-12-28 | Answer determination for natural language questioning |
US14/479,765 Abandoned US20150052113A1 (en) | 2005-11-30 | 2014-09-08 | Answer Determination for Natural Language Questioning |
Family Applications Before (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US11/319,188 Active 2026-07-30 US8832064B2 (en) | 2005-11-30 | 2005-12-28 | Answer determination for natural language questioning |
Country Status (3)
Country | Link |
---|---|
US (2) | US8832064B2 (en) |
EP (1) | EP1793318A3 (en) |
CA (1) | CA2566927A1 (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109739987A (en) * | 2018-12-29 | 2019-05-10 | 北京创鑫旅程网络技术有限公司 | A kind of corpus labeling method, construction corpus method and device |
Families Citing this family (81)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9009590B2 (en) | 2001-07-31 | 2015-04-14 | Invention Machines Corporation | Semantic processor for recognition of cause-effect relations in natural language documents |
US8799776B2 (en) | 2001-07-31 | 2014-08-05 | Invention Machine Corporation | Semantic processor for recognition of whole-part relations in natural language documents |
US20040243531A1 (en) * | 2003-04-28 | 2004-12-02 | Dean Michael Anthony | Methods and systems for representing, using and displaying time-varying information on the Semantic Web |
US7707204B2 (en) * | 2005-12-13 | 2010-04-27 | Microsoft Corporation | Factoid-based searching |
US8131536B2 (en) * | 2007-01-12 | 2012-03-06 | Raytheon Bbn Technologies Corp. | Extraction-empowered machine translation |
US7890539B2 (en) * | 2007-10-10 | 2011-02-15 | Raytheon Bbn Technologies Corp. | Semantic matching using predicate-argument structure |
US8332394B2 (en) | 2008-05-23 | 2012-12-11 | International Business Machines Corporation | System and method for providing question and answers with deferred type evaluation |
US8275803B2 (en) | 2008-05-14 | 2012-09-25 | International Business Machines Corporation | System and method for providing answers to questions |
US20140142920A1 (en) * | 2008-08-13 | 2014-05-22 | International Business Machines Corporation | Method and apparatus for Utilizing Structural Information in Semi-Structured Documents to Generate Candidates for Question Answering Systems |
CN102439590A (en) | 2009-03-13 | 2012-05-02 | 发明机器公司 | System and method for automatic semantic labeling of natural language texts |
US20110125734A1 (en) * | 2009-11-23 | 2011-05-26 | International Business Machines Corporation | Questions and answers generation |
US8983989B2 (en) * | 2010-02-05 | 2015-03-17 | Microsoft Technology Licensing, Llc | Contextual queries |
US8903794B2 (en) * | 2010-02-05 | 2014-12-02 | Microsoft Corporation | Generating and presenting lateral concepts |
US20110231395A1 (en) * | 2010-03-19 | 2011-09-22 | Microsoft Corporation | Presenting answers |
US8943051B2 (en) | 2010-09-24 | 2015-01-27 | International Business Machines Corporation | Lexical answer type confidence estimation and application |
JP5816936B2 (en) | 2010-09-24 | 2015-11-18 | 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International Business Machines Corporation | Method, system, and computer program for automatically generating answers to questions |
US9002773B2 (en) | 2010-09-24 | 2015-04-07 | International Business Machines Corporation | Decision-support application and system for problem solving using a question-answering system |
US9569724B2 (en) | 2010-09-24 | 2017-02-14 | International Business Machines Corporation | Using ontological information in open domain type coercion |
US9189541B2 (en) | 2010-09-24 | 2015-11-17 | International Business Machines Corporation | Evidence profiling |
EP2616926A4 (en) | 2010-09-24 | 2015-09-23 | Ibm | Providing question and answers with deferred type evaluation using text with limited structure |
US8892550B2 (en) | 2010-09-24 | 2014-11-18 | International Business Machines Corporation | Source expansion for information retrieval and information extraction |
US8738617B2 (en) | 2010-09-28 | 2014-05-27 | International Business Machines Corporation | Providing answers to questions using multiple models to score candidate answers |
CN103229168B (en) | 2010-09-28 | 2016-10-19 | 国际商业机器公司 | The method and system that evidence spreads between multiple candidate answers during question and answer |
WO2012047530A1 (en) | 2010-09-28 | 2012-04-12 | International Business Machines Corporation | Providing answers to questions using logical synthesis of candidate answers |
EP2622428A4 (en) * | 2010-09-28 | 2017-01-04 | International Business Machines Corporation | Providing answers to questions using hypothesis pruning |
US9020805B2 (en) | 2010-09-29 | 2015-04-28 | International Business Machines Corporation | Context-based disambiguation of acronyms and abbreviations |
US8515986B2 (en) * | 2010-12-02 | 2013-08-20 | Microsoft Corporation | Query pattern generation for answers coverage expansion |
US9842168B2 (en) | 2011-03-31 | 2017-12-12 | Microsoft Technology Licensing, Llc | Task driven user intents |
US9244984B2 (en) | 2011-03-31 | 2016-01-26 | Microsoft Technology Licensing, Llc | Location based conversational understanding |
US9858343B2 (en) | 2011-03-31 | 2018-01-02 | Microsoft Technology Licensing Llc | Personalization of queries, conversations, and searches |
US9760566B2 (en) | 2011-03-31 | 2017-09-12 | Microsoft Technology Licensing, Llc | Augmented conversational understanding agent to identify conversation context between two humans and taking an agent action thereof |
US10642934B2 (en) | 2011-03-31 | 2020-05-05 | Microsoft Technology Licensing, Llc | Augmented conversational understanding architecture |
US9064006B2 (en) | 2012-08-23 | 2015-06-23 | Microsoft Technology Licensing, Llc | Translating natural language utterances to keyword search queries |
US9069814B2 (en) * | 2011-07-27 | 2015-06-30 | Wolfram Alpha Llc | Method and system for using natural language to generate widgets |
CN102903008B (en) | 2011-07-29 | 2016-05-18 | 国际商业机器公司 | For the method and system of computer question and answer |
US8510328B1 (en) * | 2011-08-13 | 2013-08-13 | Charles Malcolm Hatton | Implementing symbolic word and synonym English language sentence processing on computers to improve user automation |
US9117194B2 (en) | 2011-12-06 | 2015-08-25 | Nuance Communications, Inc. | Method and apparatus for operating a frequently asked questions (FAQ)-based system |
US9679568B1 (en) | 2012-06-01 | 2017-06-13 | Google Inc. | Training a dialog system using user feedback |
US9123338B1 (en) | 2012-06-01 | 2015-09-01 | Google Inc. | Background audio identification for speech disambiguation |
US20140006012A1 (en) * | 2012-07-02 | 2014-01-02 | Microsoft Corporation | Learning-Based Processing of Natural Language Questions |
US9965472B2 (en) * | 2012-08-09 | 2018-05-08 | International Business Machines Corporation | Content revision using question and answer generation |
US10621880B2 (en) | 2012-09-11 | 2020-04-14 | International Business Machines Corporation | Generating secondary questions in an introspective question answering system |
US10095692B2 (en) * | 2012-11-29 | 2018-10-09 | Thornson Reuters Global Resources Unlimited Company | Template bootstrapping for domain-adaptable natural language generation |
JP6007088B2 (en) * | 2012-12-05 | 2016-10-12 | Kddi株式会社 | Question answering program, server and method using a large amount of comment text |
US9064001B2 (en) * | 2013-03-15 | 2015-06-23 | Nuance Communications, Inc. | Method and apparatus for a frequently-asked questions portal workflow |
US9965548B2 (en) | 2013-12-05 | 2018-05-08 | International Business Machines Corporation | Analyzing natural language questions to determine missing information in order to improve accuracy of answers |
US8751466B1 (en) * | 2014-01-12 | 2014-06-10 | Machine Intelligence Services, Inc. | Customizable answer engine implemented by user-defined plug-ins |
US9633137B2 (en) | 2014-04-24 | 2017-04-25 | International Business Machines Corporation | Managing questioning in a question and answer system |
US9633309B2 (en) | 2014-06-19 | 2017-04-25 | International Business Machines Corporation | Displaying quality of question being asked a question answering system |
US10073673B2 (en) * | 2014-07-14 | 2018-09-11 | Samsung Electronics Co., Ltd. | Method and system for robust tagging of named entities in the presence of source or translation errors |
US9740769B2 (en) * | 2014-07-17 | 2017-08-22 | International Business Machines Corporation | Interpreting and distinguishing lack of an answer in a question answering system |
US9754207B2 (en) | 2014-07-28 | 2017-09-05 | International Business Machines Corporation | Corpus quality analysis |
JP6464604B2 (en) * | 2014-08-08 | 2019-02-06 | 富士通株式会社 | Search support program, search support method, and search support apparatus |
US9720962B2 (en) | 2014-08-19 | 2017-08-01 | International Business Machines Corporation | Answering superlative questions with a question and answer system |
GB2531720A (en) * | 2014-10-27 | 2016-05-04 | Ibm | Automatic question generation from natural text |
WO2016068455A1 (en) * | 2014-10-30 | 2016-05-06 | 주식회사 플런티코리아 | Method and system for providing adaptive keyboard interface and response input method using adaptive keyboard linked with conversation content |
US9720963B2 (en) | 2014-11-05 | 2017-08-01 | International Business Machines Corporation | Answer category data classifying using dynamic thresholds |
US9721004B2 (en) | 2014-11-12 | 2017-08-01 | International Business Machines Corporation | Answering questions via a persona-based natural language processing (NLP) system |
US10061842B2 (en) | 2014-12-09 | 2018-08-28 | International Business Machines Corporation | Displaying answers in accordance with answer classifications |
US11138205B1 (en) * | 2014-12-22 | 2021-10-05 | Soundhound, Inc. | Framework for identifying distinct questions in a composite natural language query |
US9928269B2 (en) * | 2015-01-03 | 2018-03-27 | International Business Machines Corporation | Apply corrections to an ingested corpus |
US10366107B2 (en) | 2015-02-06 | 2019-07-30 | International Business Machines Corporation | Categorizing questions in a question answering system |
US9875296B2 (en) | 2015-03-25 | 2018-01-23 | Google Llc | Information extraction from question and answer websites |
US10795921B2 (en) * | 2015-03-27 | 2020-10-06 | International Business Machines Corporation | Determining answers to questions using a hierarchy of question and answer pairs |
US10586156B2 (en) | 2015-06-25 | 2020-03-10 | International Business Machines Corporation | Knowledge canvassing using a knowledge graph and a question and answer system |
US10170014B2 (en) * | 2015-07-28 | 2019-01-01 | International Business Machines Corporation | Domain-specific question-answer pair generation |
CN106649394A (en) * | 2015-11-03 | 2017-05-10 | 中兴通讯股份有限公司 | Fusion knowledge base processing method and device and knowledge base management system |
US10210317B2 (en) | 2016-08-15 | 2019-02-19 | International Business Machines Corporation | Multiple-point cognitive identity challenge system |
CN108228637B (en) * | 2016-12-21 | 2020-09-04 | 中国电信股份有限公司 | Automatic response method and system for natural language client |
US20180203856A1 (en) * | 2017-01-17 | 2018-07-19 | International Business Machines Corporation | Enhancing performance of structured lookups using set operations |
WO2018165579A1 (en) * | 2017-03-10 | 2018-09-13 | Eduworks Corporation | Automated tool for question generation |
US11182681B2 (en) | 2017-03-15 | 2021-11-23 | International Business Machines Corporation | Generating natural language answers automatically |
CN108829670A (en) * | 2018-06-01 | 2018-11-16 | 北京玄科技有限公司 | Based on single semantic unregistered word processing method, intelligent answer method and device |
CN109582949B (en) * | 2018-09-14 | 2022-11-22 | 创新先进技术有限公司 | Event element extraction method and device, computing equipment and storage medium |
CN110136513A (en) * | 2019-05-13 | 2019-08-16 | 上海乂学教育科技有限公司 | English Listening Comprehension resolution system |
US11341176B2 (en) | 2019-08-19 | 2022-05-24 | International Business Machines Corporation | Question answering |
US11200883B2 (en) * | 2020-01-10 | 2021-12-14 | International Business Machines Corporation | Implementing a domain adaptive semantic role labeler |
JP7323846B2 (en) * | 2020-02-25 | 2023-08-09 | 日本電信電話株式会社 | Document creation support device, document creation support method, and document creation support program |
CN111984774B (en) * | 2020-08-11 | 2024-02-27 | 北京百度网讯科技有限公司 | Searching method, searching device, searching equipment and storage medium |
US20220092352A1 (en) * | 2020-09-24 | 2022-03-24 | International Business Machines Corporation | Label generation for element of business process model |
CN112951207B (en) * | 2021-02-10 | 2022-01-07 | 网易有道信息技术(北京)有限公司 | Spoken language evaluation method and device and related product |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20040117189A1 (en) * | 1999-11-12 | 2004-06-17 | Bennett Ian M. | Query engine for processing voice based queries including semantic decoding |
US20050256700A1 (en) * | 2004-05-11 | 2005-11-17 | Moldovan Dan I | Natural language question answering system and method utilizing a logic prover |
US20060204945A1 (en) * | 2005-03-14 | 2006-09-14 | Fuji Xerox Co., Ltd. | Question answering system, data search method, and computer program |
US7376635B1 (en) * | 2000-07-21 | 2008-05-20 | Ford Global Technologies, Llc | Theme-based system and method for classifying documents |
US8719244B1 (en) * | 2005-03-23 | 2014-05-06 | Google Inc. | Methods and systems for retrieval of information items and associated sentence fragments |
Family Cites Families (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5715445A (en) * | 1994-09-02 | 1998-02-03 | Wolfe; Mark A. | Document retrieval system employing a preloading procedure |
US6754181B1 (en) * | 1996-11-18 | 2004-06-22 | Mci Communications Corporation | System and method for a directory service supporting a hybrid communication system architecture |
US6404877B1 (en) * | 1998-06-02 | 2002-06-11 | At&T Corp. | Automated toll-free telecommunications information service and apparatus |
US7082397B2 (en) * | 1998-12-01 | 2006-07-25 | Nuance Communications, Inc. | System for and method of creating and browsing a voice web |
US6377944B1 (en) * | 1998-12-11 | 2002-04-23 | Avaya Technology Corp. | Web response unit including computer network based communication |
US6665666B1 (en) * | 1999-10-26 | 2003-12-16 | International Business Machines Corporation | System, method and program product for answering questions using a search engine |
AU2001257446A1 (en) * | 2000-04-28 | 2001-11-12 | Global Information Research And Technologies, Llc | System for answering natural language questions |
US7831688B2 (en) * | 2000-06-06 | 2010-11-09 | Adobe Systems Incorporated | Method and system for providing electronic user assistance |
US6813341B1 (en) * | 2000-08-31 | 2004-11-02 | Ivoice, Inc. | Voice activated/voice responsive item locator |
US7027974B1 (en) * | 2000-10-27 | 2006-04-11 | Science Applications International Corporation | Ontology-based parser for natural language processing |
US6766316B2 (en) * | 2001-01-18 | 2004-07-20 | Science Applications International Corporation | Method and system of ranking and clustering for document indexing and retrieval |
US20030004706A1 (en) * | 2001-06-27 | 2003-01-02 | Yale Thomas W. | Natural language processing system and method for knowledge management |
US20060053000A1 (en) * | 2004-05-11 | 2006-03-09 | Moldovan Dan I | Natural language question answering system and method utilizing multi-modal logic |
-
2005
- 2005-12-28 US US11/319,188 patent/US8832064B2/en active Active
-
2006
- 2006-11-01 CA CA002566927A patent/CA2566927A1/en not_active Abandoned
- 2006-11-22 EP EP06124602A patent/EP1793318A3/en not_active Ceased
-
2014
- 2014-09-08 US US14/479,765 patent/US20150052113A1/en not_active Abandoned
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20040117189A1 (en) * | 1999-11-12 | 2004-06-17 | Bennett Ian M. | Query engine for processing voice based queries including semantic decoding |
US7376635B1 (en) * | 2000-07-21 | 2008-05-20 | Ford Global Technologies, Llc | Theme-based system and method for classifying documents |
US20050256700A1 (en) * | 2004-05-11 | 2005-11-17 | Moldovan Dan I | Natural language question answering system and method utilizing a logic prover |
US20060204945A1 (en) * | 2005-03-14 | 2006-09-14 | Fuji Xerox Co., Ltd. | Question answering system, data search method, and computer program |
US8719244B1 (en) * | 2005-03-23 | 2014-05-06 | Google Inc. | Methods and systems for retrieval of information items and associated sentence fragments |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109739987A (en) * | 2018-12-29 | 2019-05-10 | 北京创鑫旅程网络技术有限公司 | A kind of corpus labeling method, construction corpus method and device |
Also Published As
Publication number | Publication date |
---|---|
CA2566927A1 (en) | 2007-05-30 |
EP1793318A2 (en) | 2007-06-06 |
US20070136246A1 (en) | 2007-06-14 |
EP1793318A3 (en) | 2007-09-19 |
US8832064B2 (en) | 2014-09-09 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US8832064B2 (en) | Answer determination for natural language questioning | |
Gupta et al. | A survey of text question answering techniques | |
US10339453B2 (en) | Automatically generating test/training questions and answers through pattern based analysis and natural language processing techniques on the given corpus for quick domain adaptation | |
Alzahrani et al. | Fuzzy semantic-based string similarity for extrinsic plagiarism detection | |
Varma et al. | IIIT Hyderabad at TAC 2009. | |
US20160155058A1 (en) | Non-factoid question-answering system and method | |
US20140032209A1 (en) | Open information extraction | |
Arendarenko et al. | Ontology-based information and event extraction for business intelligence | |
US20160292153A1 (en) | Identification of examples in documents | |
de Abreu et al. | A review on Relation Extraction with an eye on Portuguese | |
Gokul et al. | Sentence similarity detection in Malayalam language using cosine similarity | |
Othman et al. | Enhancing question retrieval in community question answering using word embeddings | |
Pillai et al. | A combined approach using semantic role labelling and word sense disambiguation for question generation and answer extraction | |
Yaghoobzadeh et al. | ISO-TimeML event extraction in Persian text | |
Khairova et al. | Automatic extraction of synonymous collocation pairs from a text corpus | |
Dunn et al. | Language-independent ensemble approaches to metaphor identification | |
Rahat et al. | A recursive algorithm for open information extraction from Persian texts | |
Tran et al. | A model of vietnamese person named entity question answering system | |
Omri et al. | The role of temporal inferences in understanding Arabic text | |
Tanaka et al. | Acquiring and generalizing causal inference rules from deverbal noun constructions | |
Ljunglöf et al. | Assessing the quality of Språkbanken’s annotations | |
Nikolić et al. | Modelling the System of Receiving Quick Answers for e-Government Services: Study for the Crime Domain in the Republic of Serbia | |
Schäfer et al. | Advances in deep parsing of scholarly paper content | |
Hagen et al. | Word sense alignment and disambiguation for historical encyclopedias | |
Ion et al. | A trainable multi-factored QA system |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
AS | Assignment |
Owner name: AT&T INTELLECTUAL PROPERTY II, L.P., GEORGIA Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:STENCHIKOVA, SVETLANA;HAKKANI-TUR, DILEK;TUR, GOKHAN;SIGNING DATES FROM 20060404 TO 20060427;REEL/FRAME:033956/0367 |
|
AS | Assignment |
Owner name: AT&T INTELLECTUAL PROPERTY II, L.P., GEORGIA Free format text: CORRECTIVE ASSIGNMENT TO CORRECT THE ASSIGNOR NAME PREVIOUSLY RECORDED ON REEL 033956 FRAME 0367. ASSIGNOR(S) HEREBY CONFIRMS THE ASSIGNMENT;ASSIGNOR:AT&T PROPERTIES, LLC;REEL/FRAME:039735/0334 Effective date: 20130508 |
|
AS | Assignment |
Owner name: AT&T CORP., NEW YORK Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:STENCHIKOVA, SVETLANA;HAKKANI-TUR, DILEK;TUR, GOKHAN;SIGNING DATES FROM 20060404 TO 20060427;REEL/FRAME:041046/0807 Owner name: AT&T PROPERTIES, LLC, NEVADA Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:AT&T CORP.;REEL/FRAME:041046/0878 Effective date: 20160204 |
|
AS | Assignment |
Owner name: NUANCE COMMUNICATIONS, INC., MASSACHUSETTS Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:AT&T INTELLECTUAL PROPERTY II, L.P.;REEL/FRAME:041512/0608 Effective date: 20161214 |
|
STPP | Information on status: patent application and granting procedure in general |
Free format text: NOTICE OF ALLOWANCE MAILED -- APPLICATION RECEIVED IN OFFICE OF PUBLICATIONS |
|
STCB | Information on status: application discontinuation |
Free format text: ABANDONED -- FAILURE TO PAY ISSUE FEE |