WO2013125286A1 - ノン・ファクトイド型質問応答システム及びコンピュータプログラム - Google Patents
ノン・ファクトイド型質問応答システム及びコンピュータプログラム Download PDFInfo
- Publication number
- WO2013125286A1 WO2013125286A1 PCT/JP2013/051327 JP2013051327W WO2013125286A1 WO 2013125286 A1 WO2013125286 A1 WO 2013125286A1 JP 2013051327 W JP2013051327 W JP 2013051327W WO 2013125286 A1 WO2013125286 A1 WO 2013125286A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- question
- answer
- evaluation
- candidate
- feature
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/10—Machine learning using kernel methods, e.g. support vector machines [SVM]
-
- 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
- G06F16/28—Databases characterised by their database models, e.g. relational or object models
- G06F16/284—Relational databases
- G06F16/285—Clustering or classification
-
- 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
- G06F16/332—Query formulation
- G06F16/3329—Natural language query formulation or dialogue systems
-
- 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
- G06F16/334—Query execution
- G06F16/3347—Query execution using vector based model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/205—Parsing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N7/00—Computing arrangements based on specific mathematical models
Definitions
- the present invention relates to a question answering system, and more particularly to a response system for a question called a non-factoid question, such as a so-called how-type question or why-type question, which asks the reason for being.
- a non-factoid question such as a so-called how-type question or why-type question, which asks the reason for being.
- Non-Patent Document 1 As an example of such a system, there is a system described in Non-Patent Document 1 described later.
- morphological analysis is performed on the question and each sentence in the corpus, and the results are used to determine the document frequency of the words obtained from the question, the word frequency in each sentence, the total number of documents, the document length, etc.
- Is used to calculate a score, and a predetermined number of documents with higher scores are selected from the corpus. Answers to questions based on the score calculated between the words included in the selected candidate and the words included in the answer candidates, with the paragraphs included in the selected document and 1 to 3 consecutive paragraphs as the answer candidates select.
- Non-Patent Document 2 a system as described in Non-Patent Document 2 described later can be considered.
- this system for example, several answer candidates are selected by the technique described in Non-Patent Document 1, and then each answer candidate is re-ranked by a predetermined score.
- Non-Patent Document 2 questions relating to matters not related to facts are referred to as “non-factoid questions”.
- this question answering system 30 stores a corpus including a very large number of sentences (here, limited to Japanese) that can be searched on the Internet, for example, in a corpus storage unit 48.
- the answer processing unit 40 receives a non-factoid type question transmitted from the service using terminal 44 such as a mobile phone capable of text communication, and the answer processing unit 40 sends some answer sentences having a high probability as the answer to the corpus storage unit.
- the selected sentence is selected from a large number of sentences stored in 48, and is returned to the service using terminal 44 as an answer sentence list 50.
- a support vector machine (SVM) 46 is used, and the learning processing unit 42 performs supervised machine learning on the SVM 46 in advance.
- the learning processing unit 42 stores in advance several Japanese QA sentences including a non-factoid type question, a correct or incorrect answer to the non-factoid type question, and a flag indicating whether the answer is correct or not.
- Various combinations of statistical information related to syntactics selected in advance as features to be used for learning of the SVM 46 analyzing the QA sentence stored in the QA sentence storage unit 60, and the QA sentence stored in the QA sentence storage unit 60;
- a learning data generation unit 62 for generating learning data composed of a flag indicating whether the answer to the question is a correct answer to the question, and a learning data storage unit 64 for storing the learning data generated by the learning data generation unit 62
- a learning processing unit 66 that performs supervised machine learning of the SVM 46 using the learning data stored in the learning data storage unit 64.
- the answer processing unit 40 performs a predetermined grammatical analysis on the question sentence, and each of the question sentences included in the question sentence
- the question sentence analysis unit 86 for outputting information (part of speech, utilization form, dependency structure, etc.) necessary for generating a feature for the word
- the service using terminal 44 receiving the question sentence
- a candidate sentence search unit 82 for searching and extracting a predetermined number (for example, 300) of answer candidate sentences for a question from the corpus storage unit 48, and a predetermined number of candidate sentences output by the candidate sentence search unit 82 as grammatical information thereof.
- an answer candidate sentence storage unit 84 for storing together.
- candidate sentences are retrieved from the corpus storage unit 48 and extracted and stored in the answer candidate sentence storage unit 84.
- the entire sentence stored in the corpus storage unit 48 may be set as the answer candidate sentence.
- the candidate sentence search unit 82 only needs to have a function of reading all the sentences stored in the corpus storage unit 48, and the answer candidate sentence storage unit 84 temporarily stores the sentence read by the candidate sentence search unit 82. It is only necessary to fulfill the function of accumulating.
- the corpus storage unit 48 is held locally by the question answering system 30, but the present invention is not limited to such an embodiment.
- the corpus 48 may be remote, or may be stored in a distributed manner in a plurality of storage devices, not limited to one.
- the answer processing unit 40 further generates a feature vector to be given to the SVM 46 based on the combination of the information output from the question sentence analyzing unit 86 and each of the answer candidate sentences stored in the answer candidate sentence storage unit 84.
- a response sentence ranking unit 90 for ranking each of the answer sentences stored in the candidate sentence storage unit 84 and outputting a predetermined number of upper-ranked answer sentence candidates as the answer sentence list 50;
- the SVM 46 mathematically obtains a hyperplane for classifying an object into two classes, determines which class the input belongs to based on the result, and outputs the result as positive / negative polarity information.
- the answer sentence ranking unit 90 uses a combination of this distance and the polarity information output by the SVM 46 as a score of the answer candidate sentence.
- the QA sentence storage unit 60 stores in advance a number of combinations of a question sentence, a sentence suitable as an answer to the question sentence, and an incorrect sentence as an answer to the question sentence. Each combination is manually assigned a flag indicating whether or not the answer is correct.
- the learning data generation unit 62 generates learning data for learning the SVM 46 from these combinations and stores it in the learning data storage unit 64.
- the learning processing unit 66 learns the SVM 46 using the learning data stored in the learning data storage unit 64. As a result of this processing, when the SVM 46 receives the same type of feature combination as the learning data generation unit 62 generates, the combination of the original sentence combination (question sentence and answer candidate) from which the feature is obtained is obtained. It is possible to output a value indicating whether it is correct, that is, a measure of whether the answer candidate is correct as an answer to the question sentence.
- the corpus storage unit 48 stores a corpus including a large number of sentences. Each of these sentences is subjected to the same kind of analysis processing as that performed in advance by the learning data generation unit 62, and is attached with information for ranking of answer candidates similar to a part of the learning data.
- the candidate sentence search unit 82 performs an existing candidate sentence search process, and extracts a predetermined number of answer candidates for the question sentence from the corpus storage unit 48.
- the answer candidate sentences extracted by the candidate sentence search unit 82 are stored in the answer candidate sentence storage unit 84 together with information for ranking the answer candidates.
- the question sentence analysis unit 86 performs predetermined analysis processing on the question sentence, generates information necessary for generating a feature, and provides the feature vector generation unit 88 with the information.
- the feature vector generation unit 88 receives the information from the question sentence analysis unit 86, the feature vector generation unit 88 combines the answer candidate sentences stored in the answer candidate sentence storage unit 84 with the information for ranking the answer candidates, thereby learning data.
- a feature vector having the same configuration as the learning data generated by the generation unit 62 (except for a flag indicating whether or not the answer candidate is correct) is generated and provided to the response sentence ranking unit 90.
- the answer sentence ranking unit 90 gives the feature vector obtained from the combination of each answer candidate and the question sentence given from the feature vector generation unit 88 to the SVM 46. For each feature vector of each combination, the SVM 46 outputs a score indicating how appropriate the answer candidate included in the combination is as an answer to the question included in the combination.
- the answer sentence ranking unit 90 sorts the combinations of the question sentence and each answer candidate in descending order by their scores, and answers a predetermined number of answer candidates with higher scores to the question sentence given from the service use terminal 44 It returns to the service using terminal 44 as the sentence list 50.
- Non-Patent Document 1 It has been found that the system described in Non-Patent Document 1 cannot provide sufficient accuracy. In particular, the performance of the non-factoid type question answering system is considerably lower than the performance of the factual question answering system, and there is a need to improve the performance of the non-factoid type question answering system. In particular, in the future, it will be necessary not only to ask questions about facts but also to know the reason from some event and to infer its consequences from some event.
- an object of the present invention is to provide a question answering system capable of further improving accuracy in a question answering system concerning a question regarding a reason or a method.
- the question answering system is a non-factoid question answering system that receives an input of a non-factoid question and generates an answer to the question.
- This system is used by being connected to a corpus storage means for storing a corpus composed of a computer-readable document that is a document in a language to be processed.
- the system responds to the candidate search means for searching and retrieving a plurality of answer candidates for the question from the corpus storage means and receiving the input of the question.
- a feature generation means for generating a predetermined feature set and a feature set generated by the feature generation means are given.
- a score calculation means that is learned in advance so as to calculate a score indicating a degree that the answer candidate is a correct answer to the question, for the combination of the question and the answer candidate from which the set of features is generated,
- the answer may be a correct answer based on the score calculated by the score calculation means.
- Sex and an answer selection means for outputting as an answer to the question of the highest answer candidates.
- Information necessary for feature generation by the feature generation unit is attached to each of the answer candidates searched by the candidate search unit.
- the feature generation means performs a morphological analysis and a syntax analysis on the question, outputs a morpheme information and a syntactic information, and a phrase classified into the first class according to a certain evaluation criterion among the questions. And an evaluation means for specifying a phrase classified into the second class. The first class and the second class are assigned opposite polarities of evaluation.
- the question answering system further adds the analysis result by the syntax analysis means, the phrase location and the polarity of the evaluation class specified by the evaluation means, and the answer candidate for each combination of the question and the answer candidate.
- feature set generation means for generating a set of features based on the feature generation information.
- the feature generation unit further classifies each noun included in the output of the syntax analysis unit into one of a plurality of word semantic classes prepared in advance, and converts the noun into a corresponding semantic class.
- the feature set generation means performs the analysis by the syntax analysis means, the location of the phrase specified by the evaluation means, the polarity of the evaluation class, and the conversion by the semantic class conversion means for each combination of the question and the answer candidate.
- the first means for generating a set of features is included based on the output of the syntax analysis means after the processing and the information for generating the features attached to the answer candidates.
- the features generated by the first means based on the location of the phrase specified by the evaluation means and the polarity of the evaluation class are the polarity of the evaluation class of the phrase in the question and the polarity of the evaluation class of the phrase in the answer candidate. May include information indicating whether or not match.
- the feature generated by the first means based on the location of the phrase specified by the evaluation means and the polarity of the evaluation class further includes the evaluation class polarity of the phrase in the question and the evaluation of the phrase in the answer candidate. Contains information indicating the polarity of the class when it matches.
- the feature generated by the first means is obtained from the output of the syntax analysis means, and the feature obtained without using the output of the evaluation means or the output of the semantic class conversion means and the output of the evaluation means are used as the output of the evaluation means. And a feature obtained by combining the output of the semantic class conversion means with the output of the syntax analysis means.
- the feature generated by the first means may include a feature obtained by combining the output of the parsing means with the output of the evaluation means and further combining the output of the semantic class conversion means.
- the computer program according to the second aspect of the present invention is a computer program that implements a non-factoid type question answering system that receives an input of a non-factoid type question and generates an answer to the question.
- This computer is connected to a corpus storage means for storing a corpus consisting of a computer-readable document that is a document to be processed.
- a computer program comprises a candidate search means for searching and retrieving a plurality of answer candidates for a question from the corpus storage means in response to the computer receiving an input of the question, In response to receiving the input, a feature generation unit for generating a set of predetermined features for each combination of the question and each of the answer candidates stored in the candidate search unit, and a feature generation unit Given a set of features to be generated, for a combination of the question and answer candidate that generated the set of features, a score indicating the degree to which the answer candidate is a correct answer to the question is calculated.
- the score calculated by the score calculation means Based on, be a correct answer to the question is to function as the answer selection means for outputting as an answer to the question of the highest answer candidates.
- Information necessary for feature generation by the feature generation unit is attached to each of the answer candidates searched by the candidate search unit.
- a program part that causes a computer to function as a feature generation means is a computer that performs morphological analysis and syntactic analysis on a question, and outputs a morphological information and syntactic information. It is made to function as an evaluation means for specifying a phrase classified into the first evaluation class and a phrase classified into the second evaluation class.
- the computer program further attaches the computer to the combination of the question and each answer candidate, the analysis result by the syntax analysis means, the location of the phrase evaluated by the evaluation means and the polarity of the evaluation class, and the answer candidate. Based on the generated feature generation information, it is made to function as a feature set generation means for generating a feature set.
- FIG. 1 is a block diagram of a question answering system according to an embodiment of the present invention. It is a schematic diagram for demonstrating the structure of n-gram. It is a figure for demonstrating the structure of n-gram depending on a syntactic structure.
- the answer candidate A1-1 is more suitable as an answer to the question Q1.
- the direction of evaluation of “preferred” and “preferable” is referred to as “polarity”, and “phrase that describes a preferable event” and “phrase that describes a undesirable event”.
- a phrase to be evaluated with a polarity such as “is called an“ evaluation phrase ”.
- the phrase is classified into either a first evaluation class called “phrase that describes a preferable event” or a second evaluation class called “a phrase that describes an unfavorable event” according to a certain evaluation criterion. Identify the phrase to be played.
- These evaluation classes are assigned opposite polarities, such as “preferred” and “not preferable”.
- polarity can be considered variously according to evaluation criteria besides this. For example, “increase / decrease”, “activity / inactivity”, “delicious / tasteful”, “improvement / deterioration”, “adoption / non-adoption”, “happy / sad”, etc. can be considered.
- the second assumption of the present invention is that there is a fairly strong connection between the lexical semantic classes of the words in the question and those in the answer sentence.
- the answer to the question regarding the disease like the question Q1 shown in FIG. 2 is related to the chemical substance (like the answer candidate A1-1), or a virus or a part of the body, etc.
- the chemical substance like the answer candidate A1-1
- a virus or a part of the body etc.
- semantically close words are grouped into semantic classes and used as features.
- semantic classes can be constructed manually or a thesaurus based on existing meanings can be used, in this embodiment, words automatically constructed using the EM method from a large-scale corpus of web documents.
- Class groups were used. For this method, see Kazama and Torisawa, Inducing Gazetteers for Named Entity Details are described in “Recognition by Large-scale Clustering of Dependency Relations” (http: // http: //www.aclweb.org/anthology-new/P/P08/P08-1047.pdf).
- a further problem is that some answer candidates contain multiple phrases with different polarities.
- FIG. 3 shows an answer candidate A1-3 similar to the answer candidate A1-2 of FIG.
- the answer candidate A1-3 shown in FIG. 3 includes an unfavorable evaluation phrase 140 related to cancer and a preferable evaluation phrase 142 related to cancer.
- an answer to the question Q1 shown in FIG. 2 there remains a question as to which of the answer candidate A1-1 shown in FIG. 2 and the answer candidate A1-3 shown in FIG.
- a combination of the polarity of the evaluation and the contents of the evaluation phrase linked to the polarity of both the question sentence and the answer candidate is used.
- the non-factoid question answering system 160 acquires a document set 164 composed of a large number of documents on the Internet in advance, and uses the service use terminal 166 via the Internet 162.
- a list of items suitable as an answer to the question is generated from the accumulated document set 164 and returned to the service using terminal 166.
- a question from the service using terminal 166 is transmitted to the non-factoid type question answering system 160 as input data to a web form.
- this question answering system 160 is a document set 164 made up of a large number of documents acquired in advance from the Internet, and information necessary for feature formation for ranking answer candidates is attached to each sentence.
- An answer processing unit 170 for extracting and ranking several items and outputting a response sentence list 174 that lists a predetermined number of higher ranks to the service using terminal 166, and the answer processing unit 170 ranks answer candidates.
- the supervised machine learning for the SVM 176 is performed by using the SVM 176 used in the process of learning and the learning data prepared in advance.
- the SVM 176 When the SVM 176 receives a combination of predetermined features obtained from the combination of the question sentence and the answer sentence, whether the combination is correct, that is, for the question in the combination, the answer sentence in the combination is A learning processing unit 172 that indicates whether or not the answer is correct and outputs a score indicating how likely the answer is.
- the answer processing unit 170 searches the corpus storage unit 178 by a process similar to the conventional technique, and a predetermined number (for example, 300) of the many sentences.
- the question sentence analysis unit 226 performs analysis processing on the question sentence, performs morphological analysis and syntax analysis (syntactic analysis) used for the feature, and outputs morpheme information and syntactic information; Of the information output by the question sentence analysis unit 226, a semantic class is estimated by applying a predetermined statistical probability model to the word, and the information indicating the semantic class is attached and output.
- the evaluation processing unit 228 performs evaluation on the output of the semantic class conversion unit 230 and the question sentence analysis unit 226, and determines the evaluation phrase and its polarity as described above and outputs the
- each sentence of the document stored in the corpus storage unit 178 is also subjected to the same processing as that performed in advance by the question sentence analysis unit 226, the evaluation processing unit 228, and the semantic class conversion unit 230. deep. By doing so, it is possible to reduce the amount of processing for generating a feature vector from a set of questions and answer candidates later.
- the answer processing unit 170 further receives the output of the question sentence analysis unit 226, the output of the evaluation processing unit 228, and the output of the semantic class conversion unit 230, and further receives each answer candidate sentence and accompanying information from the answer candidate sentence storage unit 224. And a feature vector generation unit 232 for generating a feature vector to be applied to the SVM 176 based on information on both the question sentence and the answer candidate, and an output from the feature vector generation part 232 for each answer candidate sentence.
- the answer candidates are ranked based on the score obtained from the SVM 176, the answer sentence list 174 including a predetermined upper number of answer candidates is generated, and the answer for replying to the service using terminal 166
- the learning processing unit 172 includes a QA sentence storage unit 190 for storing a large number of QA sentences together with a flag indicating the suitability of the combination, and each combination of a question sentence and an answer candidate stored in the QA sentence storage unit 190
- the QA sentence analysis unit 192 for performing the same processing as the question sentence analysis unit 226, and the semantic class information is attached to each word in the output of the QA sentence analysis unit 192 using a statistical model.
- Semantic class conversion unit 196 and an evaluation processing unit 194 for outputting a tag indicating an evaluation phrase and its polarity by performing an evaluation process on each of the question sentence and the answer sentence of the QA sentence.
- the learning data generation unit 198 that outputs the learning data
- the learning data storage unit 200 that stores the learning data that the learning data generation unit 198 outputs
- the learning data stored in the learning data storage unit 200 are used for the SVM 176.
- An SVM learning unit 202 for performing supervised machine learning.
- the candidate sentence search unit 222 determines, for each question sentence, the top predetermined order in the descending order of possibility of including, for example, answers from among 600 million documents stored in the corpus storage unit 178. The number of documents (for example, 300) is adjusted to be extracted. Each candidate is further divided into a set of answer candidates consisting of five consecutive sentences. In order to reduce the possibility that a correct answer cannot be obtained due to an error in the document dividing method, the two divided sentences can be shared with each other.
- the candidate sentence search unit 222 further scores each of the answer candidates ac obtained for the question q in this way by a scoring function S (q, ac) shown in the following equation (1).
- the part includes words included in the question, and further includes three clue words (reason, cause, and factor) indicating the causal relationship. Search for.
- the candidate sentence search unit 222 selects 300 answer candidates for the question based on the ranking according to the formula (1), and provides the answer sentence ranking unit 234 via the answer candidate sentence storage unit 224.
- the scoring function S (q, ac) shown in Expression (1) assigns a score similar to tf (logarithmized index word frequency) -idf (reciprocal of document frequency) to each answer candidate.
- 1 / dist (t 1 , t 2 ) in the formula (1) plays a role like tf, and 1 / df (t 2 ) is given t which is shared by the question q and the answer candidate ac. a idf for 1 and t 2.
- T is a word set consisting of nouns, verbs, and adjectives in the question q that also appear in the answer candidate ac. Note that if clue words are present in the answer candidate ac, they are added to the word set T.
- N is the total number of documents (600 million)
- dist (t 1 , t 2 ) is the distance (number of characters) between the words t 1 and t 2 in the answer candidate ac.
- df (t) is the document frequency of the word t
- ⁇ ⁇ 0,1 ⁇ is an index
- ⁇ 1 if ts (t1, t2)> 1
- ⁇ 0 otherwise.
- the feature set includes (1) a feature (indicated by “MSA”) representing the result of morphological analysis and syntactic analysis between a question and an answer candidate, and (2) a feature representing a semantic word class appearing in the question and the answer candidate ( And (3) a feature (indicated by “SA”) representing the result of the evaluation analysis.
- MSA a feature representing the result of morphological analysis and syntactic analysis between a question and an answer candidate
- SA a feature representing a semantic word class appearing in the question and the answer candidate
- MSA is a feature that has been widely used for re-ranking of answer candidates. This feature is intended to clarify the connection between the question and the answer candidate at the morpheme, clause, and syntactic structure level.
- g-grams (n is 1 to 3) of morphemes, phrases, and syntactic structures are extracted from each question and answer candidate.
- the sentence includes a portion 240 composed of four consecutive clauses A, B, C, and D as shown in FIG.
- the phrase 3-gram obtained from these four phrases is a 3-gram 242 composed of the phrases A, B, and C shown in FIG. 6 (B1), and a 3-gram 244 composed of the phrases B, C, and D shown in (B2). It is.
- the syntactic structure n-gram means a path having three consecutive clauses in the syntactic structure network in this specification.
- a syntactic structure network 260 as shown in FIG.
- 3-gram 262 consisting of clauses A, B and C shown in FIG. 7 (B1), and 3 consisting of clauses A, C and D shown in FIG. 7 (B2).
- -gram 264 is obtained. The difference can be seen by comparing these with FIG. Note that the 1-gram of the syntactic structure n-gram is not used because it matches the phrase 1-gram.
- MSA As the MSA, four types of features of MSA1 to MSA4 shown in Table 270 of FIG. 8 can be used. The description is as shown in FIG. 8 and will not be repeated individually.
- the MSA 2 is an n-gram that includes the word in question, such as “type gun” in FIG.
- a word semantic class refers to a set of words that are semantically similar.
- such a word semantic class is constructed by using the noun clustering technique described in the above-mentioned Kazama and Torisawa documents. The algorithm described in this document is based on the assumption that semantically similar words tend to appear in similar contexts.
- a syntactic dependency relationship between words is treated as a context, and a probability model for a hidden class regarding a non-verb dependency relationship is defined by the following equation (2).
- n is a noun
- v is a verb or noun to which the noun n is subordinated by a grammatical relationship r (a postscript in Japanese)
- c is a hidden class.
- the frequency of the dependency relationship was obtained from a corpus consisting of 600 million sentences.
- SWC is for reflecting the relationship between the word in question and the word in answer candidate.
- the learning data there are a question having a word of a specific semantic class and an answer candidate having a word of a specific semantic class, and the mutual relationship is positive (the answer candidate is a correct answer to the question).
- the SVM determines that if a word of the same semantic class as the question of the learning data is in the question, the other semantics are the same, and the same semantic class as the word of the specific semantic class in the answer of the learning data would be selected as the answer to that question.
- the same processing can be considered at the word level instead of the word semantic level.
- the connection at a specific word level can be statistically modeled.
- the combination of words and words is too specific, and the problem of data sparseness arises. Therefore, it is difficult to generalize the model, and the reliability of the model cannot be increased.
- SWC1 and SWC2 are used. Their explanation is as described in FIG. 8.
- SWC1 To obtain SWC1, follow the procedure below. First, all nouns in the n-gram of MSA1 are converted into semantic classes to which they belong. Among the obtained n-grams, the one in which the semantic class of at least one word remains is adopted as SWC1.
- SWC2 is an n-gram among answer candidates, and is the same as the word existing in the question, but changed to their semantic class.
- ⁇ SA The feature SA is further classified into two types. The first is evaluation analysis at the word level (word polarity). This is represented by SA @ W1 to SA @ W4 in FIG. The second is evaluation analysis at the phrase level (phrase polarity). These are represented by SA @ P1 to SA @ P10 in FIG.
- Word polarity (SA @ W)
- the polarity of the word is determined by referring to a dictionary prepared in advance for the polarity tendency of the word.
- a dictionary provided by the tool program introduced by Non-Patent Document 3 is used as the dictionary. This feature is intended to clarify the relationship between the polarity of the word in the question and the polarity of the word in the correct answer.
- the polarity of the word in the question and the polarity of the word in the selected answer are likely to be equal to each other.
- SA @ W1 and SA @ W2 shown in FIG. 8 are evaluation analysis features obtained from a word polarity n-gram having at least one word having word polarity. These n-grams are obtained by converting all nouns in the MSA1 n-gram to corresponding word polarities by referring to the dictionary. SA @ W1 relates to all word polarity n-grams in questions and answer candidates. SA @ W2 is an n-gram obtained by SA @ W1, in which the original n-gram includes a question word.
- combining the word polarity with the semantic class of the word allows the classifier to identify those combinations that are particularly meaningful. For example, a word having a negative word polarity and a semantic class of “condition” would represent an “unfavorable condition”. As a result, the classifier can learn the correlation between the words representing these negative conditions and their relevance to the question about the disease.
- SA @ W3 and SA @ W4 are this type of feature. These are defined similarly to SA @ W1 and SA @ W2, respectively. However, the difference is that the polarity of words in SA @ W1 and SA @ W2 is replaced with the meaning class of words and the polarity of words.
- the SA @ W3 and SA @ W4 n-grams are referred to as (word) class-polarity coupled n-grams.
- Phrase polarity (SA @ P)
- an existing tool program (according to Non-Patent Document 3 described above) is used for extracting an evaluation phrase and determining its polarity.
- an evaluation phrase that does not include a word in the question is not useful for determining a correct answer.
- only the evaluation phrase extracted from the sentence including at least one word in the question is used for generating the phrase polarity feature.
- the first category includes SA @ P1 and SA @ P2 shown in FIG.
- the second category includes SA @ p3 to SA @ P5 in FIG.
- the third category includes SA @ P6 to SA @ P10 in FIG.
- the feature of the first category relates to the matching of the polarity of the evaluation phrase between the question and the answer candidate. Examine combinations of evaluation phrases between questions and answer candidates. If any of the polarities coincides with each other, the index indicating the polarity and the polarities are SA @ P1 and SA @ P2, respectively.
- the features of the second category are based on the assumption that evaluation phrases often constitute the center of correct answers. It is necessary to use features representing these.
- SA @ P3 to SA @ P5 in FIG. 8 are for that purpose. Refer to FIG. 8 for the contents of these features.
- the feature of the third category is for generalizing the feature related to the content of the question or the answer described above by using the semantic class of the word.
- these features are a combination of the semantic class n-gram and class-polarity combined n-gram of the word obtained from the evaluation phrase, with the phrase polarity. Refer to FIG. 8 for the contents of these features.
- a plurality of combinations (QA sentences) composed of questions and their answers prepared manually are prepared and stored in the QA sentence storage unit 190.
- a combination of these questions and answers a set consisting of a certain question and a correct answer to the question is prepared.
- the number of people participating in this preparation is often limited, and it is unavoidable that there may be restrictions on the system in that respect.
- the QA sentence analysis unit 192 performs morphological analysis, syntax analysis, and the like for each of these QA sentences, and attaches the obtained part-of-speech information, dependency relationship information, and the like to the QA sentence, and the evaluation processing unit 194 , Semantic class conversion unit 196, and learning data generation unit 198.
- the evaluation processing unit 194 searches for an evaluation phrase for each question and answer in each of the given QA sentences, and attaches the range and its polarity as tags to the corresponding part to the learning data generation unit 198. give.
- the semantic class conversion unit 196 generates a converted noun in each given QA sentence into a word semantic class, and provides the learning data generation unit 198 with it.
- the learning data generation unit 198 uses the morpheme and syntactic information from the QA sentence analysis unit 192, the information about the evaluation from the evaluation processing unit 194, and the information about the semantic class from the semantic class conversion unit 196, as shown in FIG. Are generated, and learning data with a flag attached to the QA sentence (a flag indicating whether the answer is appropriate for the question) is generated and stored in the learning data storage unit 200. .
- the SVM learning unit 202 learns the SVM 176 using the learning data stored in the learning data storage unit 200.
- the learned SVM 176 is made available by the answer sentence ranking unit 234.
- the reply processing unit 170 can perform reply processing.
- the service using terminal 166 transmits a question to the answer processing unit 170, the question sentence analyzing unit 226 and the candidate sentence searching unit 222 receive this question.
- the candidate sentence search unit 222 searches the top 300 answer candidate sentences that are highly likely to be answer candidates from among a large number of sentences stored in the corpus storage unit 178, It outputs to the answer candidate sentence storage unit 224.
- the scale used for creating the score of the answer candidate sentence is as shown in Expression (1).
- the question sentence analysis unit 226 performs morpheme analysis and syntax analysis on the received question, and outputs morpheme information and syntactic information.
- the evaluation processing unit 228 performs an evaluation process on the information output from the question sentence analysis unit 226, performs tagging on the question sentence indicating the range of the evaluation phrase and its polarity, and sends it to the feature vector generation unit 232. give.
- the semantic class conversion unit 230 applies the statistical model for estimating the semantic class represented by the expression (2) to the noun among the information output by the question sentence analysis unit 226, thereby obtaining the semantic class. Is converted into information representing a semantic class and provided to the feature vector generation unit 232.
- the feature vector generation unit 232 includes an evaluation tagging question sentence output from the evaluation processing unit 228, morpheme information and syntax analysis information output from the question sentence analysis unit 226, information output from the semantic class conversion unit 230, and Based on the same information previously attached to each of the answer candidate sentences stored in the answer candidate sentence storage unit 224, a feature as shown in FIG. 8 is obtained for each combination of the question and each answer candidate. , To generate a feature vector composed of them. The feature vector generation unit 232 gives each combination of the question and each answer candidate to the answer sentence ranking unit 234.
- the answer sentence ranking unit 234 obtains a score indicating whether or not the answer candidate is appropriate as an answer to the question for each combination by applying the SVM 176 to the given combination. Further, the answer sentence ranking unit 234 sorts the combinations in descending order according to this score, and generates an answer sentence list 174 by arranging a predetermined upper number in the order of the score. The reply sentence ranking unit 234 returns the reply sentence list 174 thus obtained to the service use terminal 166.
- test set was constructed manually. Building a test set consists of generating a question and verifying the answer to it. Although it is desirable to build this test set with as many people as possible, in practice it was a limited number of people (4 people) due to various restrictions. In practice, the scope of the question will be broader, so the results of the following experiment are likely to indicate the upper limit of performance when using the above embodiment for actual applications.
- the top 20 answer candidates were first searched for each question using the system of the above embodiment, and all combinations of the questions and the questions were manually checked.
- the scale (Fleiss' Kappa) indicating the variation in evaluation among the evaluators was 0.611, which was an almost consistent evaluation.
- the correct answer to each question was determined by majority vote.
- FIG. 9 shows a question and answer candidate set 280 and a question and answer candidate set 282.
- P @ 1 indicates how many of the questions are correct as the top answer.
- MAP evaluates the overall quality of the top n answer candidates using the following formula.
- Q is a set of non-factoid type questions
- a q is a set of correct answers to non-factoid type questions q ⁇ Q
- Prec (k) is the accuracy at the k-th place of the top n answer candidates
- rel (k) represents an index that is 1 when rank k is a correct answer in A q and 0 otherwise.
- FIG. 10 shows the results obtained by the experiment with the above settings for five different systems.
- B-QA is a system in which the inventor has implemented the system described in Non-Patent Document 1, and this is the basic system.
- B-Ranker is a B-QA answer search system with re-ranking using features (MSA) obtained from morpheme information and syntactic information.
- MSA features obtained from morpheme information and syntactic information.
- the system according to the present embodiment used in the present embodiment is indicated as “present invention” and utilizes all the features listed in FIG.
- What is described as “upper limit” refers to a system in which, when there is a correct answer among the top 20 answer candidates, all n correct answers are listed in the top n answer candidates. This shows the best result (upper limit) that can be expected in this experiment.
- the results of each system compared to the upper limit system are shown in parentheses.
- FIG. 10 shows that the system according to the present invention shows higher accuracy than the two systems to be compared.
- an improvement in accuracy of 11.6% is obtained with P @ 1 over B-QA.
- the accuracy of B-Ranker (6.4% at P @ 1)
- the set of features used in the present invention is superior to the features used in the prior art documents. I understand that there is. Further, since there is a difference in performance between B-Ranker and B-QA, it can be seen that re-ranking is effective.
- MSA indicates B-Ranker.
- SA was divided into two groups, SA @ W (features representing words and their polarities) and SA @ P (features representing phrases and their polarities). From the difference in performance between MSA + SWC + SA and other results, it can be seen that all of the features used contributed to the improvement in accuracy. From the viewpoint of P @ 1, it can be seen that the highest improvement is obtained by SA @ P, which is a feature related to the phrase and its polarity. That is, the performance difference between the present invention and MSA + SWC + SA @ W has reached 4.4%. According to this fact, when unfavorable things happen, the reasons are often unfavorable, and when desirable things happen, the reasons are often desirable. It has been shown that the assumptions underlying the idea are likely to be correct.
- the answer processing unit 170 and the learning processing unit 172 are each realized by computer hardware, a program executed by the computer hardware, and data stored in the computer hardware. . Both may be implemented in the same computer hardware.
- FIG. 12 shows the external appearance of the computer system 330
- FIG. 13 shows the internal configuration of the computer system 330.
- the computer system 330 includes a computer 340 having a memory port 352 and a DVD (Digital Versatile Disc) drive 350, a keyboard 346, a mouse 348, and a monitor 342.
- DVD Digital Versatile Disc
- the computer 340 boots up a CPU (Central Processing Unit) 356, a bus 366 connected to the CPU 356, the memory port 352, and the DVD drive 350, and A read only memory (ROM) 358 that stores programs and the like, and a random access memory (RAM) 360 that is connected to the bus 366 and stores program instructions, system programs, work data, and the like.
- the computer system 330 further includes a network interface (I / F) 344 that provides a connection to a network that enables communication with other terminals.
- I / F network interface
- a computer program for causing the computer system 330 to operate as a non-factoid type question answering system is stored in the DVD 362 or the removable memory 364 attached to the DVD drive 350 or the memory port 352, and further transferred to the hard disk 354.
- the program may be transmitted to the computer 340 through a network (not shown) and stored in the hard disk 354.
- the program is loaded into the RAM 360 when executed.
- the program may be loaded directly into the RAM 360 from the DVD 362, from the removable memory 364, or via the network IF 344.
- This program includes a plurality of instructions for causing the computer 340 to operate as the non-factoid type question answering system of this embodiment.
- Some of the basic functions necessary to perform this operation are provided by operating system (OS) or third party programs running on computer 340 or modules of various programming toolkits installed on computer 340. . Therefore, this program does not necessarily include all functions necessary for realizing the system and method of this embodiment.
- This program can be used as a non-factoid question answering system by calling an appropriate function or a suitable program tool in a programming tool kit in a controlled manner so as to obtain a desired result. It is only necessary to include an instruction for executing the operation.
- the operation of computer system 330 is well known. Therefore, it does not repeat here.
- the question is sent as text from the service using terminal, and the answer is also sent back as text.
- the present invention is not limited to such an embodiment, and can be applied to, for example, a question answering system using voice.
- the service use terminal 166 shown in FIG. 5 is a mobile phone, and voice recognition is performed on the voice signal of the question received from the service use terminal 166 before the candidate sentence search unit 222 and the question sentence analysis unit 226.
- the question may be converted into text data.
- the answer sentence list 174 may be returned to the service using terminal 166 in the form of a web page, or by performing speech synthesis based on the answer having the highest score, You may reply to the service utilization terminal 166 as an audio
- the above embodiment related to Japanese.
- the present invention is not applicable only to Japanese. Any language can be applied as long as it is capable of creating learning data and collecting a sufficiently large amount of computer-readable text.
- the sentence stored in the corpus storage unit 178 is processed in advance by performing the same processing as that performed by the question sentence analysis unit 226, the evaluation processing unit 228, and the semantic class conversion unit 230. Assume that they are stored in association. By doing so, the processing time for the feature generation in the feature vector generation unit 232 can be shortened.
- the present invention is not limited to such an embodiment. Specifically, for example, only the information necessary for the candidate sentence search by the candidate sentence search unit 222 is attached to the document stored in the corpus storage unit 178, and selected when the feature vector is generated by the feature vector generation unit 232. Each of the candidate sentences may be subjected to processing necessary for feature generation (processing similar to the processing by the question sentence analysis unit 226, the evaluation processing unit 228, and the semantic class conversion unit 230).
- the candidate sentence search unit 222 searches for a candidate answer sentence for a question.
- a sentence having a vocabulary similar to the question sentence is searched mainly based on the document frequency and the word frequency in the document. ing.
- the present invention is not limited to such an embodiment. Any standard may be used as long as it is considered possible to extract a sentence that can be an answer to the question.
- the evaluation class may be a plurality of classes that can be given a certain order, and the target may be classified into one of the plurality of evaluation classes.
- two or more evaluation criteria may be provided, and the objects may be classified into a plurality of evaluation classes based on the two or more criteria.
- the present invention is a question answering system related to a question regarding a reason or method using natural language processing, and can be used in an industry that manufactures, uses, and lends a question answering system or the like that can further improve accuracy. it can.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- Databases & Information Systems (AREA)
- Computational Linguistics (AREA)
- Computing Systems (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Human Computer Interaction (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Optimization (AREA)
- Mathematical Analysis (AREA)
- Computational Mathematics (AREA)
- Algebra (AREA)
- Health & Medical Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- General Health & Medical Sciences (AREA)
- Machine Translation (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
Description
本願発明では、ノン・ファクトイド型の質問とその回答との間には、以下のような関係があることに着目し、これを第1の仮定とした。すなわち、
・好ましくないことが起きると、その理由も何か好ましくないものであることが多く、
・好ましいことが起きると、その理由もまた好ましいものであることが多い、
という点に着目したものである。
Recognition by Large-scale Clustering of Dependency Relations」(http://http://www.aclweb.org/anthology-new/P/P08/P08-1047.pdf)に詳細が記載されている。
図4を参照して、本実施の形態に係るノン・ファクトイド型質問応答システム160は、予めインターネット上の大量の文書からなる文書集合164を獲得しておき、インターネット162を介してサービス利用端末166からノン・ファクトイド型の質問を受けたことに応答して、蓄積しておいた文書集合164からその質問に対する回答としてふさわしいもののリストを生成し、サービス利用端末166に返信するものである。本実施の形態では、サービス利用端末166からの質問はウェブのフォームへの入力データとしてノン・ファクトイド型質問応答システム160に送信されるものとする。
候補文検索部222としては、本実施の形態では、http://lucene.apache.org/solrで配布されているSolrを用いる。本実施の形態では、候補文検索部222は、質問文1つに対し、コーパス記憶部178に格納されている6億個の文書のうちから例えば回答を含む可能性の高い順に、トップの所定個数(例えば300個)の文書を抽出するように調整する。各候補はさらに、5つの連続する文からなる回答候補の集合に分割される。文書の分割方法の誤りにより正しい回答が得られない可能性を小さくするために、分割される文書については、互いに2つの文までは共有できるようにしている。
前記したとおり、ランキングには3種類の素性集合を用いる教師付分類器(SVM)を使用する。素性集合は、(1)質問と回答候補との形態素分析及び統語分析の結果を表す素性(「MSA」で示す。)、(2)質問と回答候補とに現れる意味的単語クラスを表す素性(「SWC」で示す。)、及び(3)評価分析の結果を表す素性(「SA」で示す。)、を含む。
MSAは、回答候補の再ランキングのために従来から広く用いられている素性である。この素性は、質問と回答候補との間の、形態素、文節、及び統語構造レベルでの結び付きを明らかにするためのものである。
phrases)、統語構造チェーンの集まりである。これらはそれぞれ、形態素分析プログラム(例えばhttp://nlp.ist.i.kyoto-u.ac.jp/index.php?JUMAN)及び構文解析プログラム(例えばhttp://nlp.ist.i.kyoto-u.ac.jp/index.php?KNP)を用いて得ることができる。
単語意味クラスとは、意味的に類似した単語の集合のことをいう。本実施の形態では、前記したKazama及びTorisawaの文献に記載された名詞のクラスタリング技術を用い、このような単語意味クラスを構築した。この文献に記載されたアルゴリズムは、意味的に類似した単語は類似した文脈に現れる傾向がある、という仮定に基づくものである。この方法は、単語間の統語的係り受け関係を文脈として扱うことにより、非動詞の係り受け関係に関する、隠れクラスについての確率モデルを以下の式(2)により定義する。
SAという素性はさらに2種類に分類される。第1は単語レベルでの評価分析である(単語の極性)。これを図8ではSA@W1~SA@W4で表す。第2は、フレーズレベルでの評価分析である(フレーズ極性)。これらは図8ではSA@P1~SA@P10で表す。
単語の極性は、予め準備された、単語の極性傾向のための辞書の参照により行なう。辞書としては、本実施の形態では、非特許文献3により紹介されたツールプログラムにより提供されるものを用いる。この素性は、質問中の単語の極性と、正しい回答中の単語の極性との関連を明らかにするためのものである。ここでは、前提にしたがい、学習の結果、質問中の単語の極性と、選択される回答中の単語の極性とは互いに等しいことが多くなる筈である。
本実施の形態では、評価フレーズの抽出及びその極性の判定は、既存のツールプログラム(前述の非特許文献3による。)を用いる。実験の結果、質問内の語を含まないような評価フレーズについては、正しい回答を定めるために有用ではないことがわかった。したがって、本実施の形態では、質問内の語を少なくとも1つ含むような文から抽出された評価フレーズのみを、フレーズ極性の素性生成のために用いることとした。
以上に説明した本実施の形態に係るノン・ファクトイド型の質問応答システム160の動作は、基本的に図1に示す従来のものと同様の手順を踏む。すなわち、SVM176の学習と、回答処理部170による質問の処理とである。これらに先立ち、評価処理部194で使用する評価ツール及び辞書、並びに、意味クラス変換部196で使用する統計的確率モデルを作成しておく必要がある。この確率モデルは、式(2)によるものである。さらに、多数の文を収集し、コーパス記憶部178に格納する。これらについては、既に述べたように、素性ベクトル生成部232が素性を生成するために必要な情報を作成しておき、各文に付しておく。一方、人手により作成した、質問とその回答とからなる複数の組合せ(QA文)を準備し、QA文記憶部190に格納する。これら質問と回答との組合せとしては、ある質問とその質問に対する正しい回答とからなる組とを準備する。これら質問と回答との組合せとしては、できるだけ多くの種類の質問を用いることが望ましい。そのためには、できるだけ多くの人にこのQA文の作成に参加してもらうことが理想的である。もっとも、現実的にはこの準備に参加する人の数は限定されることが多く、その点でシステムにある制約が生じ得ることはやむを得ない。
回答文ランキング部234にSVM176が組込まれることにより、回答処理部170による回答処理が可能になる。サービス利用端末166が質問を回答処理部170に送信すると、質問文解析部226及び候補文検索部222がこの質問を受ける。
上記した実施の形態により、ノン・ファクトイド型の質問に対する回答の精度としてどの向上が得られるかについて、実験により調べた。
上記実施の形態を評価するため、手作業によりテストセットを構築した。テストセットの構築は、質問の生成と、それに対する回答の検証とである。なるべく多くの人によりこのテストセットを構築することが望ましいが、実際には様々な制約から、限られた人数(4人)による作業となった。実際には質問の範囲はより広くなるであろうから、以下の実験の結果は、実際の応用に上記実施の形態を用いたときの性能の上限を示すものとなると思われる。
実験では、上記したテストセットを用い、10分割交差検定を行なうことでシステムの評価を行なった。学習には、線形カーネルを持つTinySVM(http://chasen.org/~taku/software/TinySVM/)を用いた。評価は、P@1(上位1個の精度)と、MAP(Mean Average Precision:平均適合率(AP)の平均)とにより行なった。
5つの異なるシステムについて、上記設定の実験により得られた結果を図10に示す。図10において、B-QAは、非特許文献1に記載されたシステムを発明者が実装したものであり、これが基本システムとなる。B-RankerはB-QAの回答検索システムに、形態素情報及び統語情報から得た素性(MSA)を用いた再ランキングを付加したものである。本実施の形態に係るシステムの、本実施の形態に用いたものは、「本発明」として示してあり、図8に挙げた全ての素性を利用したものである。「上限」と記載したものは、トップ20の回答候補中に正しい回答が存在するときには必ずその上位n個の回答候補の中にn個の正しい回答を全て挙げるようなシステムのことを指す。これは、本実験で期待できる最上の結果(上限)を示す。各システムによる結果を上限システムと比較したものをカッコ内に示す。
以上のようにこの実施の形態によれば、上記した仮定に基づき、質問中の評価フレーズと、回答候補中の評価フレーズとを特定し、それらの極性が同じである場合に回答候補が正しいものである、とする考えに基づいて素性を選択することにより、回答検索の精度の向上を得ることができた。さらに、そのような評価に関する統計的モデルを学習する際に、質問が広い範囲にわたる可能性があること、それに対して学習のためのデータの範囲及び数には限りがあることに鑑み、質問及び回答候補中の名詞について、意味クラスを導入し、それら名詞を意味クラスに置換えたものを素性として導入した。これら素性の導入により、回答を検索する対象となるコーパスが非常に多くの文を含み、かつ学習データに限りがある場合であっても、従来のノン・ファクトイド型の質問応答システムと比較して大きく精度が向上したノン・ファクトイド型の質問応答システムを得ることができた。
この実施の形態のシステムのうち、回答処理部170及び学習処理部172はそれぞれ、コンピュータハードウェアと、そのコンピュータハードウェアにより実行されるプログラムと、コンピュータハードウェアに格納されるデータとにより実現される。両者が同一のコンピュータハードウェア内に実装されてもよい。図12はこのコンピュータシステム330の外観を示し、図13はコンピュータシステム330の内部構成を示す。
40、170 回答処理部
42、172 学習処理部
44、166 サービス利用端末
46、176 SVM
48、178 コーパス記憶部
50、174 回答文リスト
60、190 QA文記憶部
62、198 学習データ生成部
64、200 学習データ記憶部
66 学習処理部
82、222 候補文検索部
84、224 回答候補文記憶部
86、226 質問文解析部
88、232 素性ベクトル生成部
90、234 回答文ランキング部
160 ノン・ファクトイド型の質問応答システム
192 QA文解析部
194、228 評価処理部
196、230 意味クラス変換部
Claims (6)
- ノン・ファクトイド型の質問の入力を受け、当該質問に対する回答を生成するノン・ファクトイド型の質問応答システムであって、
処理対象の言語の文書であって、コンピュータ読取可能な文書からなるコーパスを記憶するためのコーパス記憶手段に接続されて用いられ、
質問の入力を受けたことに応答して、前記コーパス記憶手段から、当該質問に対する複数個の回答候補を検索し取り出すための候補検索手段と、
質問の入力を受けたことに応答して、当該質問と、前記候補検索手段に記憶された回答候補の各々との組合せについて、所定の素性の集合を生成するための素性生成手段と、
前記素性生成手段により生成される前記素性の集合が与えられると、当該素性の集合を生成する元となった質問及び回答候補の組合せについて、当該回答候補が当該質問に対する正しい回答である度合いを示すスコアを算出するように予め学習されたスコア算出手段と、
前記質問と、前記回答候補に記憶された回答候補の各々との組合せについて、前記スコア算出手段により算出されたスコアに基づき、前記質問に対する正しい回答である可能性が最も高い回答候補を前記質問に対する回答として出力する回答選択手段とを含み、
前記候補検索手段により検索された回答候補の各々には、前記素性生成手段による素性の生成に必要な情報が付されており、
前記素性生成手段は、
前記質問について、形態素解析及び構文解析を行ない、形態素情報及び統語情報を出力するための構文解析手段と、
前記質問のうち、ある評価基準にしたがって第1の評価クラスに分類されるフレーズと第2の評価クラスに分類されるフレーズとを特定するための評価手段とを含み、前記第1の評価クラス及び第2の評価クラスには、それぞれ互いに相反する評価の極性が割当てられており、
前記質問応答システムはさらに、前記質問と、前記回答候補の各々との組合せについて、前記構文解析手段による解析結果と、前記評価手段により特定されたフレーズの箇所及びその評価クラスの極性と、前記回答候補に付されている素性生成用の情報とに基づいて、前記素性の集合を生成するための素性集合生成手段を含む、ノン・ファクトイド型の質問応答システム。 - 請求項1に記載のノン・ファクトイド型の質問応答システムであって、
前記素性生成手段はさらに、前記構文解析手段の出力に含まれる各名詞を、予め準備された、複数個の、単語の意味クラスのいずれかに分類し、当該名詞を対応する意味クラスに変換するための意味クラス変換手段を含み、
前記素性集合生成手段は、前記質問と、前記回答候補の各々との組合せについて、前記構文解析手段による解析結果と、前記評価手段により特定されたフレーズの箇所及びその評価クラスの極性と、前記意味クラス変換手段による変換がされた後の前記構文解析手段の出力と、前記回答候補に付されている素性生成用の情報とに基づいて、前記素性の集合を生成するための第1の手段を含む、請求項1に記載のノン・ファクトイド型の質問応答システム。 - 前記第1の手段が、前記評価手段により特定されたフレーズの箇所及び評価クラスの極性に基づいて生成する素性は、質問中のフレーズの評価クラスの極性と、回答候補中のフレーズの評価クラスの極性とが一致するか否かを示す情報を含む、請求項2に記載のノン・ファクトイド型の質問応答システム。
- 前記第1の手段が生成する素性は、前記構文解析手段の出力から得られ、前記評価手段の出力又は前記意味クラス変換手段の出力を用いずに得られる素性と、前記構文解析手段の出力に、前記評価手段の出力を組合せて得られる素性と、前記構文解析手段の出力に、前記意味クラス変換手段の出力を組合せて得られる素性とを含む、請求項2に記載のノン・ファクトイド型の質問応答システム。
- 前記第1の手段が生成する素性は、前記構文解析手段の出力に、前記評価手段の出力を組合せ、さらに前記意味クラス変換手段の出力を組合せて得られる素性を含む、請求項2に記載のノン・ファクトイド型の質問応答システム。
- ノン・ファクトイド型の質問の入力を受け、当該質問に対する回答を生成するノン・ファクトイド型の質問応答システムをコンピュータにより実現させるコンピュータプログラムであって、
前記コンピュータは、処理対象の言語の文書であって、コンピュータ読取可能な文書からなるコーパスを記憶するためのコーパス記憶手段に接続され、
前記コンピュータプログラムは、前記コンピュータを、
質問の入力を受けたことに応答して、前記コーパス記憶手段から、当該質問に対する複数個の回答候補を検索し取り出すための候補検索手段と、
質問の入力を受けたことに応答して、当該質問と、前記候補検索手段に記憶された回答候補の各々との組合せについて、所定の素性の集合を生成するための素性生成手段と、
前記素性生成手段により生成される前記素性の集合が与えられると、当該素性の集合を生成する元となった質問及び回答候補の組合せについて、当該回答候補が当該質問に対する正しい回答である度合いを示すスコアを算出するように予め学習されたスコア算出手段と、
前記質問と、前記回答候補に記憶された回答候補の各々との組合せについて、前記スコア算出手段により算出されたスコアに基づき、前記質問に対する正しい回答である可能性が最も高い回答候補を前記質問に対する回答として出力する回答選択手段として機能させ、
前記候補検索手段により検索された回答候補の各々には、前記素性生成手段による素性の生成に必要な情報が付されており、
前記コンピュータを前記素性生成手段として機能させるプログラム部分は、前記コンピュータを、
前記質問について、形態素解析及び構文解析を行ない、形態素情報及び統語情報を出力するための構文解析手段と、
前記質問のうち、ある評価基準にしたがって第1の評価クラスに分類されるフレーズと、第2の評価クラスに分類されるフレーズとを特定するための評価手段として機能させ、前記第1の評価クラス及び第2の評価クラスには、それぞれ互いに相反する評価の極性が割当てられており、
前記コンピュータプログラムはさらに、前記コンピュータを、前記質問と、前記回答候補の各々との組合せについて、前記構文解析手段による解析結果と、前記評価手段により特定されたフレーズの箇所及びその評価クラスの極性と、前記回答候補に付されている素性生成用の情報とに基づいて、前記素性の集合を生成するための素性集合生成手段として機能させる、コンピュータプログラム。
Priority Applications (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP13751215.8A EP2819031A4 (en) | 2012-02-23 | 2013-01-23 | SYSTEM FOR ANSWERING QUESTIONS WITHOUT FACTORY AND COMPUTER PROGRAM THEREFOR |
US14/377,999 US9697477B2 (en) | 2012-02-23 | 2013-01-23 | Non-factoid question-answering system and computer program |
KR1020147023478A KR101968102B1 (ko) | 2012-02-23 | 2013-01-23 | 논팩토이드형 질의 응답 시스템 및 컴퓨터 프로그램 |
CN201380010249.4A CN104137102B (zh) | 2012-02-23 | 2013-01-23 | 非事实型询问应答系统以及方法 |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2012036972A JP5825676B2 (ja) | 2012-02-23 | 2012-02-23 | ノン・ファクトイド型質問応答システム及びコンピュータプログラム |
JP2012-036972 | 2012-02-23 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2013125286A1 true WO2013125286A1 (ja) | 2013-08-29 |
Family
ID=49005480
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2013/051327 WO2013125286A1 (ja) | 2012-02-23 | 2013-01-23 | ノン・ファクトイド型質問応答システム及びコンピュータプログラム |
Country Status (6)
Country | Link |
---|---|
US (1) | US9697477B2 (ja) |
EP (1) | EP2819031A4 (ja) |
JP (1) | JP5825676B2 (ja) |
KR (1) | KR101968102B1 (ja) |
CN (1) | CN104137102B (ja) |
WO (1) | WO2013125286A1 (ja) |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR101559376B1 (ko) | 2014-02-24 | 2015-10-12 | 서울대학교산학협력단 | 엔트로피를 최대화하는 실시간 조합 자질 구축 방법 및 이를 수행하는 조합 자질 구축 장치 |
US9471601B2 (en) | 2014-03-25 | 2016-10-18 | International Business Machines Corporation | Images for a question answering system |
US9727552B2 (en) | 2015-05-27 | 2017-08-08 | International Business Machines Corporation | Utilizing a dialectical model in a question answering system |
US9842161B2 (en) | 2016-01-12 | 2017-12-12 | International Business Machines Corporation | Discrepancy curator for documents in a corpus of a cognitive computing system |
US9886479B2 (en) | 2014-07-29 | 2018-02-06 | International Business Machines Corporation | Managing credibility for a question answering system |
US10102275B2 (en) | 2015-05-27 | 2018-10-16 | International Business Machines Corporation | User interface for a query answering system |
US10146858B2 (en) | 2015-12-11 | 2018-12-04 | International Business Machines Corporation | Discrepancy handler for document ingestion into a corpus for a cognitive computing system |
US10176250B2 (en) | 2016-01-12 | 2019-01-08 | International Business Machines Corporation | Automated curation of documents in a corpus for a cognitive computing system |
JP2019160286A (ja) * | 2018-03-07 | 2019-09-19 | 株式会社哲英 | 質問応答システム、質問受付回答システム、第1次回答システム、及びそれらを用いた質問応答方法 |
JP2019197498A (ja) * | 2018-05-11 | 2019-11-14 | 国立研究開発法人情報通信研究機構 | 対話システム及びそのコンピュータプログラム |
CN110837738A (zh) * | 2019-09-24 | 2020-02-25 | 平安科技(深圳)有限公司 | 相似问识别方法、装置、计算机设备及存储介质 |
CN111309875A (zh) * | 2018-12-10 | 2020-06-19 | 百度在线网络技术(北京)有限公司 | 问题应答的方法、装置、设备和存储介质 |
Families Citing this family (63)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9634855B2 (en) | 2010-05-13 | 2017-04-25 | Alexander Poltorak | Electronic personal interactive device that determines topics of interest using a conversational agent |
JP5825676B2 (ja) * | 2012-02-23 | 2015-12-02 | 国立研究開発法人情報通信研究機構 | ノン・ファクトイド型質問応答システム及びコンピュータプログラム |
JP6150282B2 (ja) * | 2013-06-27 | 2017-06-21 | 国立研究開発法人情報通信研究機構 | ノン・ファクトイド型質問応答システム及びコンピュータプログラム |
US9471874B2 (en) * | 2013-12-07 | 2016-10-18 | International Business Machines Corporation | Mining forums for solutions to questions and scoring candidate answers |
US9276939B2 (en) * | 2013-12-17 | 2016-03-01 | International Business Machines Corporation | Managing user access to query results |
US9754207B2 (en) | 2014-07-28 | 2017-09-05 | International Business Machines Corporation | Corpus quality analysis |
JP6414956B2 (ja) | 2014-08-21 | 2018-10-31 | 国立研究開発法人情報通信研究機構 | 質問文生成装置及びコンピュータプログラム |
KR101581816B1 (ko) * | 2014-10-14 | 2016-01-05 | 서강대학교산학협력단 | 기계학습을 이용한 음성인식방법 |
US9946763B2 (en) * | 2014-11-05 | 2018-04-17 | International Business Machines Corporation | Evaluating passages in a question answering computer system |
US9501525B2 (en) * | 2014-11-05 | 2016-11-22 | International Business Machines Corporation | Answer sequence evaluation |
KR102094934B1 (ko) * | 2014-11-19 | 2020-03-31 | 한국전자통신연구원 | 자연어 질의 응답 시스템 및 방법 |
JP6345577B2 (ja) | 2014-11-26 | 2018-06-20 | 株式会社ユニバーサルエンターテインメント | 情報提供システム、情報提供方法、プログラム、及びデータ構造 |
JP6087333B2 (ja) | 2014-11-26 | 2017-03-01 | 株式会社ユニバーサルエンターテインメント | 情報提供システム、情報提供方法、及びプログラム |
US10061842B2 (en) | 2014-12-09 | 2018-08-28 | International Business Machines Corporation | Displaying answers in accordance with answer classifications |
US10127284B2 (en) * | 2014-12-18 | 2018-11-13 | International Business Machines Corporation | Scoring attributes in deep question answering systems based on algorithmic source code influences |
US9928269B2 (en) | 2015-01-03 | 2018-03-27 | International Business Machines Corporation | Apply corrections to an ingested corpus |
JP6551968B2 (ja) * | 2015-03-06 | 2019-07-31 | 国立研究開発法人情報通信研究機構 | 含意ペア拡張装置、そのためのコンピュータプログラム、及び質問応答システム |
CN104794109B (zh) * | 2015-04-09 | 2017-06-23 | 山西大学 | 一种应用于学习机的智能解答系统 |
FR3037843B1 (fr) * | 2015-06-24 | 2018-01-05 | Gaztransport Et Technigaz | Procede et dispositif de decoupe de matiere isolante fibreuse ou alveolaire |
US10223440B2 (en) * | 2015-06-29 | 2019-03-05 | International Business Machines Corporation | Question and answer system emulating people and clusters of blended people |
JP6544131B2 (ja) * | 2015-08-14 | 2019-07-17 | 富士ゼロックス株式会社 | 情報処理装置及び情報処理プログラム |
JP6618735B2 (ja) * | 2015-08-31 | 2019-12-11 | 国立研究開発法人情報通信研究機構 | 質問応答システムの訓練装置及びそのためのコンピュータプログラム |
CN105630917A (zh) * | 2015-12-22 | 2016-06-01 | 成都小多科技有限公司 | 智能应答方法及装置 |
US9720981B1 (en) | 2016-02-25 | 2017-08-01 | International Business Machines Corporation | Multiple instance machine learning for question answering systems |
JP6623852B2 (ja) * | 2016-03-09 | 2019-12-25 | 富士通株式会社 | 検索制御プログラム、検索制御装置及び検索制御方法 |
US10592519B2 (en) | 2016-03-29 | 2020-03-17 | Microsoft Technology Licensing, Llc | Computational-model operation using multiple subject representations |
US20170293621A1 (en) * | 2016-04-06 | 2017-10-12 | International Business Machines Corporation | Natural language processing based on textual polarity |
US20170293620A1 (en) * | 2016-04-06 | 2017-10-12 | International Business Machines Corporation | Natural language processing based on textual polarity |
US10706044B2 (en) | 2016-04-06 | 2020-07-07 | International Business Machines Corporation | Natural language processing based on textual polarity |
CN107402912B (zh) * | 2016-05-19 | 2019-12-31 | 北京京东尚科信息技术有限公司 | 解析语义的方法和装置 |
JP6767042B2 (ja) * | 2016-09-26 | 2020-10-14 | 国立研究開発法人情報通信研究機構 | シナリオパッセージ分類器、シナリオ分類器、及びそのためのコンピュータプログラム |
JP6721179B2 (ja) * | 2016-10-05 | 2020-07-08 | 国立研究開発法人情報通信研究機構 | 因果関係認識装置及びそのためのコンピュータプログラム |
JP6929539B2 (ja) * | 2016-10-07 | 2021-09-01 | 国立研究開発法人情報通信研究機構 | ノン・ファクトイド型質問応答システム及び方法並びにそのためのコンピュータプログラム |
WO2018066489A1 (ja) * | 2016-10-07 | 2018-04-12 | 国立研究開発法人情報通信研究機構 | ノン・ファクトイド型質問応答システム及び方法並びにそのためのコンピュータプログラム |
US10699215B2 (en) * | 2016-11-16 | 2020-06-30 | International Business Machines Corporation | Self-training of question answering system using question profiles |
US10303763B2 (en) * | 2017-01-06 | 2019-05-28 | International Business Machines Corporation | Process for identifying completion of domain adaptation dictionary activities |
JP2019020893A (ja) | 2017-07-13 | 2019-02-07 | 国立研究開発法人情報通信研究機構 | ノン・ファクトイド型質問応答装置 |
JP6218057B1 (ja) * | 2017-07-14 | 2017-10-25 | Jeインターナショナル株式会社 | 自動応答サーバー装置、端末装置、応答システム、応答方法、およびプログラム |
US10902738B2 (en) * | 2017-08-03 | 2021-01-26 | Microsoft Technology Licensing, Llc | Neural models for key phrase detection and question generation |
JP6972788B2 (ja) | 2017-08-31 | 2021-11-24 | 富士通株式会社 | 特定プログラム、特定方法および情報処理装置 |
US10387572B2 (en) * | 2017-09-15 | 2019-08-20 | International Business Machines Corporation | Training data update |
US11409749B2 (en) * | 2017-11-09 | 2022-08-09 | Microsoft Technology Licensing, Llc | Machine reading comprehension system for answering queries related to a document |
KR102005525B1 (ko) * | 2017-12-29 | 2019-07-30 | (주)헤르스 | 딥러닝 기술을 활용한 화학물질 안전관리의 질의응답 관리시스템 |
US11449762B2 (en) | 2018-02-20 | 2022-09-20 | Pearson Education, Inc. | Real time development of auto scoring essay models for custom created prompts |
US11741849B2 (en) | 2018-02-20 | 2023-08-29 | Pearson Education, Inc. | Systems and methods for interface-based machine learning model output customization |
JP6772213B2 (ja) * | 2018-04-23 | 2020-10-21 | 日本電信電話株式会社 | 質問応答装置、質問応答方法及びプログラム |
US11106664B2 (en) * | 2018-05-03 | 2021-08-31 | Thomson Reuters Enterprise Centre Gmbh | Systems and methods for generating a contextually and conversationally correct response to a query |
CN110489531B (zh) * | 2018-05-11 | 2023-05-30 | 阿里巴巴集团控股有限公司 | 高频问题的确定方法和装置 |
JP7081671B2 (ja) | 2018-08-06 | 2022-06-07 | 富士通株式会社 | 評価プログラム、評価方法および情報処理装置 |
RU2704738C1 (ru) * | 2018-10-15 | 2019-10-30 | Общество с ограниченной ответственностью "Уралинновация" | Автоматизированная информационно-голосовая вопросно-ответная система |
US11822588B2 (en) * | 2018-10-24 | 2023-11-21 | International Business Machines Corporation | Supporting passage ranking in question answering (QA) system |
CN109739956B (zh) * | 2018-11-08 | 2020-04-10 | 第四范式(北京)技术有限公司 | 语料清洗方法、装置、设备及介质 |
CN111382573A (zh) * | 2018-12-12 | 2020-07-07 | 百度在线网络技术(北京)有限公司 | 用于答案质量评估的方法、装置、设备和存储介质 |
DE112019006005T5 (de) * | 2019-01-08 | 2021-09-02 | Mitsubishi Electric Corporation | Semantische-Beziehungen-Lerneinrichtung, Semantische-Beziehungen-Lernverfahren und Semantische-Beziehungen-Lernprogramm |
US11093491B2 (en) | 2019-01-16 | 2021-08-17 | International Business Machines Corporation | Question answering system using topic-sensitive source reliability scoring |
US11262978B1 (en) * | 2019-06-19 | 2022-03-01 | Amazon Technologies, Inc. | Voice-adapted reformulation of web-based answers |
RU2722461C1 (ru) * | 2019-11-19 | 2020-06-01 | Общество с ограниченной ответственностью "Уралинновация" | Голосовая роботизированная вопросно-ответная система и способ ее автоматического взаимодействия с электронным устройством пользователя |
RU2724600C1 (ru) * | 2019-11-19 | 2020-06-25 | Общество с ограниченной ответственностью "Уралинновация" | Голосовая роботизированная вопросно-ответная система и способ ее автоматического взаимодействия с электронным устройством пользователя |
CN111339268B (zh) * | 2020-02-19 | 2023-08-15 | 北京百度网讯科技有限公司 | 实体词识别方法和装置 |
US20210279605A1 (en) * | 2020-03-06 | 2021-09-09 | International Business Machines Corporation | Efficiently generating accurate responses to a multi-facet question by a question answering system |
JP7568359B2 (ja) | 2020-06-04 | 2024-10-16 | 東京エレクトロン株式会社 | サーバ装置、顧客サポートサービス提供方法及び顧客サポートサービス提供プログラム |
US20220318230A1 (en) * | 2021-04-05 | 2022-10-06 | Vianai Systems, Inc. | Text to question-answer model system |
CN115455160B (zh) * | 2022-09-02 | 2024-08-06 | 腾讯科技(深圳)有限公司 | 一种多文档阅读理解方法、装置、设备及存储介质 |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2009116662A (ja) * | 2007-11-07 | 2009-05-28 | National Institute Of Information & Communication Technology | 質問応答装置、質問応答方法、及びプログラム |
JP2009157791A (ja) * | 2007-12-27 | 2009-07-16 | Nippon Telegr & Teleph Corp <Ntt> | 質問応答方法、装置、プログラム並びにそのプログラムを記録した記録媒体 |
Family Cites Families (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2002247029A (ja) * | 2000-02-02 | 2002-08-30 | Sony Corp | 認証装置、認証システムおよびその方法、処理装置、通信装置、通信制御装置、通信システムおよびその方法、情報記録方法およびその装置、情報復元方法およびその装置、その記録媒体 |
JP2004139553A (ja) * | 2002-08-19 | 2004-05-13 | Matsushita Electric Ind Co Ltd | 文書検索システムおよび質問応答システム |
CN101377777A (zh) * | 2007-09-03 | 2009-03-04 | 北京百问百答网络技术有限公司 | 一种自动问答方法和系统 |
US8543565B2 (en) * | 2007-09-07 | 2013-09-24 | At&T Intellectual Property Ii, L.P. | System and method using a discriminative learning approach for question answering |
CN101872349B (zh) * | 2009-04-23 | 2013-06-19 | 国际商业机器公司 | 处理自然语言问题的方法和装置 |
KR20130021944A (ko) * | 2011-08-24 | 2013-03-06 | 한국전자통신연구원 | 서술형 질의 응답 방법 및 장치 |
JP5825676B2 (ja) * | 2012-02-23 | 2015-12-02 | 国立研究開発法人情報通信研究機構 | ノン・ファクトイド型質問応答システム及びコンピュータプログラム |
JP5924666B2 (ja) * | 2012-02-27 | 2016-05-25 | 国立研究開発法人情報通信研究機構 | 述語テンプレート収集装置、特定フレーズペア収集装置、及びそれらのためのコンピュータプログラム |
JP6206840B2 (ja) * | 2013-06-19 | 2017-10-04 | 国立研究開発法人情報通信研究機構 | テキストマッチング装置、テキスト分類装置及びそれらのためのコンピュータプログラム |
JP6150282B2 (ja) * | 2013-06-27 | 2017-06-21 | 国立研究開発法人情報通信研究機構 | ノン・ファクトイド型質問応答システム及びコンピュータプログラム |
JP6150291B2 (ja) * | 2013-10-08 | 2017-06-21 | 国立研究開発法人情報通信研究機構 | 矛盾表現収集装置及びそのためのコンピュータプログラム |
JP6403382B2 (ja) * | 2013-12-20 | 2018-10-10 | 国立研究開発法人情報通信研究機構 | フレーズペア収集装置、及びそのためのコンピュータプログラム |
JP5907393B2 (ja) * | 2013-12-20 | 2016-04-26 | 国立研究開発法人情報通信研究機構 | 複雑述語テンプレート収集装置、及びそのためのコンピュータプログラム |
JP5904559B2 (ja) * | 2013-12-20 | 2016-04-13 | 国立研究開発法人情報通信研究機構 | シナリオ生成装置、及びそのためのコンピュータプログラム |
-
2012
- 2012-02-23 JP JP2012036972A patent/JP5825676B2/ja active Active
-
2013
- 2013-01-23 KR KR1020147023478A patent/KR101968102B1/ko active IP Right Grant
- 2013-01-23 EP EP13751215.8A patent/EP2819031A4/en not_active Withdrawn
- 2013-01-23 CN CN201380010249.4A patent/CN104137102B/zh not_active Expired - Fee Related
- 2013-01-23 WO PCT/JP2013/051327 patent/WO2013125286A1/ja active Application Filing
- 2013-01-23 US US14/377,999 patent/US9697477B2/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2009116662A (ja) * | 2007-11-07 | 2009-05-28 | National Institute Of Information & Communication Technology | 質問応答装置、質問応答方法、及びプログラム |
JP2009157791A (ja) * | 2007-12-27 | 2009-07-16 | Nippon Telegr & Teleph Corp <Ntt> | 質問応答方法、装置、プログラム並びにそのプログラムを記録した記録媒体 |
Non-Patent Citations (5)
Title |
---|
KYOSUKE YOSHIDA: "Introduction of a Probabilistic Language Model to Non-Factoid Question-Answering Using Examples of Q&A Pairs", IPSJ SIG NOTES 2011 (HEISEI 23) NENDO 1, 15 June 2011 (2011-06-15), pages 1 - 10, XP055159624 * |
MASAKI MURATA; SACHIYO TSUKAWAKI; TOSHIYUKI KANAMARU; QING MA; HITOSHI ISAHARA: "A System for Answering Non-factoid Japanese Questions by Using Passage Retrieval Weighted Based on Type of Answer", PROC. OF NTCIR-6 |
RYUICHIRO HIGASHINAKA; HIDEKI ISOZAKI: "Corpus-based Question Answering for Why-Questions", PROC. OFIJCNLP, pages 418 - 425 |
See also references of EP2819031A4 |
TETSUJI NAKAGAWA; KENTARO INUI; SADAO KUROHASHI: "Proc. of Human Language Technologies: The 2010 Annual Conference of Computational Linguistics", June 2010, ASSOCIATION FOR COMPUTATIONAL LINGUISTICS, article "Dependency Tree-based Sentiment Classification Using CRFs with Hidden Variables", pages: 786 - 794 |
Cited By (26)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR101559376B1 (ko) | 2014-02-24 | 2015-10-12 | 서울대학교산학협력단 | 엔트로피를 최대화하는 실시간 조합 자질 구축 방법 및 이를 수행하는 조합 자질 구축 장치 |
US9471601B2 (en) | 2014-03-25 | 2016-10-18 | International Business Machines Corporation | Images for a question answering system |
US9495387B2 (en) | 2014-03-25 | 2016-11-15 | International Business Machines Corporation | Images for a question answering system |
US9886479B2 (en) | 2014-07-29 | 2018-02-06 | International Business Machines Corporation | Managing credibility for a question answering system |
US9886480B2 (en) | 2014-07-29 | 2018-02-06 | International Business Machines Corporation | Managing credibility for a question answering system |
US10157174B2 (en) | 2015-05-27 | 2018-12-18 | International Business Machines Corporation | Utilizing a dialectical model in a question answering system |
US9727552B2 (en) | 2015-05-27 | 2017-08-08 | International Business Machines Corporation | Utilizing a dialectical model in a question answering system |
US10102275B2 (en) | 2015-05-27 | 2018-10-16 | International Business Machines Corporation | User interface for a query answering system |
US10942958B2 (en) | 2015-05-27 | 2021-03-09 | International Business Machines Corporation | User interface for a query answering system |
US11030227B2 (en) | 2015-12-11 | 2021-06-08 | International Business Machines Corporation | Discrepancy handler for document ingestion into a corpus for a cognitive computing system |
US10146858B2 (en) | 2015-12-11 | 2018-12-04 | International Business Machines Corporation | Discrepancy handler for document ingestion into a corpus for a cognitive computing system |
US10366116B2 (en) | 2016-01-12 | 2019-07-30 | International Business Machines Corporation | Discrepancy curator for documents in a corpus of a cognitive computing system |
US11308143B2 (en) | 2016-01-12 | 2022-04-19 | International Business Machines Corporation | Discrepancy curator for documents in a corpus of a cognitive computing system |
US11074286B2 (en) | 2016-01-12 | 2021-07-27 | International Business Machines Corporation | Automated curation of documents in a corpus for a cognitive computing system |
US10176250B2 (en) | 2016-01-12 | 2019-01-08 | International Business Machines Corporation | Automated curation of documents in a corpus for a cognitive computing system |
US9842161B2 (en) | 2016-01-12 | 2017-12-12 | International Business Machines Corporation | Discrepancy curator for documents in a corpus of a cognitive computing system |
JP2020123371A (ja) * | 2018-03-07 | 2020-08-13 | 株式会社哲英 | 質問応答システム、質問受付回答システム、第1次回答システム、及びそれらを用いた質問応答方法 |
JP2019160286A (ja) * | 2018-03-07 | 2019-09-19 | 株式会社哲英 | 質問応答システム、質問受付回答システム、第1次回答システム、及びそれらを用いた質問応答方法 |
JP7534823B2 (ja) | 2018-03-07 | 2024-08-15 | 株式会社哲英 | 質問応答システム、質問受付回答システム、第1次回答システム、及びそれらを用いた質問応答方法 |
JP7311899B2 (ja) | 2018-03-07 | 2023-07-20 | 株式会社哲英 | 質問応答システム、質問受付回答システム、第1次回答システム、及びそれらを用いた質問応答方法 |
JP2019197498A (ja) * | 2018-05-11 | 2019-11-14 | 国立研究開発法人情報通信研究機構 | 対話システム及びそのコンピュータプログラム |
JP7126682B2 (ja) | 2018-05-11 | 2022-08-29 | 国立研究開発法人情報通信研究機構 | 対話システム及びそのコンピュータプログラム |
CN111309875B (zh) * | 2018-12-10 | 2023-08-04 | 百度在线网络技术(北京)有限公司 | 问题应答的方法、装置、设备和存储介质 |
CN111309875A (zh) * | 2018-12-10 | 2020-06-19 | 百度在线网络技术(北京)有限公司 | 问题应答的方法、装置、设备和存储介质 |
CN110837738A (zh) * | 2019-09-24 | 2020-02-25 | 平安科技(深圳)有限公司 | 相似问识别方法、装置、计算机设备及存储介质 |
CN110837738B (zh) * | 2019-09-24 | 2023-06-30 | 平安科技(深圳)有限公司 | 相似问识别方法、装置、计算机设备及存储介质 |
Also Published As
Publication number | Publication date |
---|---|
US20150026106A1 (en) | 2015-01-22 |
EP2819031A4 (en) | 2016-05-18 |
KR20140128346A (ko) | 2014-11-05 |
JP2013171550A (ja) | 2013-09-02 |
US9697477B2 (en) | 2017-07-04 |
CN104137102B (zh) | 2017-06-13 |
EP2819031A1 (en) | 2014-12-31 |
CN104137102A (zh) | 2014-11-05 |
KR101968102B1 (ko) | 2019-04-11 |
JP5825676B2 (ja) | 2015-12-02 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
JP5825676B2 (ja) | ノン・ファクトイド型質問応答システム及びコンピュータプログラム | |
KR102256240B1 (ko) | 논팩토이드형 질의 응답 시스템 및 방법 | |
KR102640564B1 (ko) | 질의 응답 시스템의 훈련 장치 및 그것을 위한 컴퓨터 프로그램 | |
Zhou et al. | Predicting discourse connectives for implicit discourse relation recognition | |
Cimiano et al. | Learning taxonomic relations from heterogeneous sources of evidence | |
CA2484410C (en) | System for identifying paraphrases using machine translation techniques | |
Turney et al. | Experiments with three approaches to recognizing lexical entailment | |
Mills et al. | Graph-based methods for natural language processing and understanding—A survey and analysis | |
Alami et al. | Automatic texts summarization: Current state of the art | |
Alsallal et al. | Intrinsic plagiarism detection using latent semantic indexing and stylometry | |
Sardinha | An assessment of metaphor retrieval methods | |
Liebeskind et al. | Semiautomatic construction of cross-period thesaurus | |
Dung | Natural language understanding | |
CN109992651A (zh) | 一种问题目标特征自动识别和抽取方法 | |
Zhang et al. | The identification of the emotionality of metaphorical expressions based on a manually annotated chinese corpus | |
CN111680493B (zh) | 英语文本分析方法、装置、可读存储介质及计算机设备 | |
Quan et al. | Combine sentiment lexicon and dependency parsing for sentiment classification | |
Dunn et al. | Language-independent ensemble approaches to metaphor identification | |
CN102346777A (zh) | 一种对例句检索结果进行排序的方法和装置 | |
Zouaoui et al. | Ontological Approach Based on Multi-Agent System for Indexing and Filtering Arabic Docu-ments | |
Tanaka et al. | Acquiring and generalizing causal inference rules from deverbal noun constructions | |
Chitra et al. | Genetic algorithm based feature selection for paraphrase recognition | |
Schwartz et al. | Minimally supervised classification to semantic categories using automatically acquired symmetric patterns | |
Zhang et al. | An ATE system based on probabilistic relations between terms and syntactic functions | |
CN117725153B (zh) | 文本匹配方法、装置、电子设备和存储介质 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 13751215 Country of ref document: EP Kind code of ref document: A1 |
|
REEP | Request for entry into the european phase |
Ref document number: 2013751215 Country of ref document: EP |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2013751215 Country of ref document: EP Ref document number: 14377999 Country of ref document: US |
|
ENP | Entry into the national phase |
Ref document number: 20147023478 Country of ref document: KR Kind code of ref document: A |
|
NENP | Non-entry into the national phase |
Ref country code: DE |