JP2013250925A - Answer type estimation device, method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To estimate an answer type to a question sentence in detail.SOLUTION: A feature quantity extraction unit 13, from a question sentence to be estimated, extracts a feature quantity regarding a character N-gram, a numeral classifier and a suffix, a semantic attribute of a content word that appears at the end of the question sentence and the last character of the content word that appears at the end of the question sentence respectively; and generates a feature quantity vector having the extracted feature quantity as elements. An answer type estimation unit 15 estimates an answer type of the question sentence to be estimated based on an answer type model 20 which is learned by using correct answer data which comprises a question sentence and an answer type of the correct answer, and the feature quantity vector of the question sentence to be estimated.

Description

  The present invention relates to an answer type estimation apparatus, method, and program, and more particularly, to an answer type estimation apparatus, method, and program for estimating an answer type for a question sentence in a natural language.

  A system that outputs a random answer for a question sentence in a natural language is called a question answering system. The question answering system is a “factoid type” answering questions that are mainly answered by words, such as “The highest mountain in the world”, “What kind of person is Akira Kurosawa”, and “Why do fireflies shine?” Like this, it is difficult to answer in words, and it is classified as “non-factoid type” that answers mainly in sentences. In order to output an appropriate response to a question, it is important to appropriately estimate an answer type indicating the type of answer to the question.

  Conventionally, in order to estimate an answer type, many techniques for describing and estimating a rule by hand are used. For example, if the word “why” is present in the question sentence, a rule that presumes that the answer type is “reason” is described. Similarly, when the word “where” is present in the question sentence, the answer type is “place name”, and when the word “who” is present, the answer type is assumed to be “person name”. Write the rules.

  However, there are a variety of aspects of the question text entered by humans. For example, if the question is “Where is the laboratory that developed XX?”, The “organization name” will be the answer type. However, just because the word “where” exists in the question sentence does not necessarily mean that the answer type is “place name”. In order to estimate the answer type with high accuracy for such various questions, it is necessary to describe a very large number of rules. In this way, a method using machine learning has been proposed in order to eliminate the need for describing rules by labor. For example, a method has been proposed in which learning data provided with a correct answer type is learned using a support vector machine (SVM) to estimate the answer type (see, for example, Non-Patent Document 1). ).

  By the way, conventionally, in the estimation of answer types, in particular, the estimation of factoid type answer types, eight answer types are often used as the types of answer types to be estimated. This is a classification of the answer type used in a question answering contest (TREC: Text Retrieval Conference), and has been used as a standard.

Jun Suzuki, Hirotoshi Taira, Yutaka Sasaki, and Eisaku Maeda, "Question Classification using HDAG Kernel", Proc. COLING, 2002.

  However, when an attempt is made to construct a question answering system that answers questions in a pinpoint manner, eight answer types are not sufficient. For example, in the conventional estimation of the eight types of answer types for the question sentences such as “What is the tallest mountain in the world?” And “What is the longest river in the world?”, Both answer types are “place names”. This could result in answering the name of the river despite listening to the mountain, or answering the name of the municipality despite listening to the name of the river.

  In order to avoid such erroneous answers, an approach of increasing the types of answer types is effective. However, increasing the number of answer types means that the meaning type of the question must be understood in more detail than before, and the answer type must be estimated, but only the information that has been used to estimate the eight answer types in the past. Then, there is a possibility that it is not possible to classify with high accuracy for the detailed answer types with increased types.

  The present invention has been made in view of the above facts, and an object thereof is to provide an answer type estimation device, method, and program capable of estimating in detail an answer type for a question sentence.

  In order to achieve the above object, the answer type estimation device of the present invention includes, from a question sentence in natural language, a letter N-gram, a classifier, a suffix, a semantic attribute of a content word appearing at the end of the question sentence, and the question A plurality of learning question sentences to which feature quantity extraction means for extracting feature quantities relating to at least one of the last characters of the content word appearing at the end of the sentence and a correct answer type indicating the type of answer to the question sentence are given From the answer type model that learned the correspondence between the feature amount extracted from each of the answer types and each of the answer types assigned to each of the plurality of learning question sentences, and the estimation target question sentence whose answer type is unknown Answer type estimating means for estimating the answer type of the estimation target question sentence based on the feature quantity extracted by the feature quantity extracting means.

  According to the answer type estimation device of the present invention, the feature amount extraction means includes a character N-gram, a classifier, a suffix, a semantic attribute of a content word appearing at the end of a question sentence, and a question sentence A feature amount relating to at least one of the last characters of the content word that appears last is extracted. Then, the answer type estimating means assigns the feature amount extracted from each of the plurality of learning question sentences to which the answer type correct answer indicating the type of the answer to the question sentence is given to each of the plurality of learning question sentences The answer type of the estimation target question sentence based on the answer type model that learned the association with each of the answered type and the feature quantity extracted by the feature quantity extraction means from the estimation target question sentence whose answer type is unknown Is estimated.

  As described above, as the feature amount, the feature relating to at least one of the character N gram, the classifier, the suffix, the semantic attribute of the content word appearing at the end of the question sentence, and the last character of the content word appearing at the end of the question sentence. By using detailed information using quantities, it is possible to classify the answer types with high accuracy even when there are many types of answer types, and therefore it is possible to estimate the answer types for the question sentences in detail.

  The answer type model includes a first answer type model and a second answer type model in which the types of estimated answer types are larger than the first answer type model, and the answer type estimating means includes the first answer type model. Based on the answer type model and the feature quantity extracted from the estimation target question sentence, an upper answer type of the estimation target question sentence is estimated, and extracted from the second answer type model and the estimation target question sentence. Based on the characteristic amount, the answer type of the lower hierarchy restricted by the estimation result of the answer type of the upper hierarchy can be estimated as the answer type of the estimation target question sentence. Thus, the estimation accuracy of the answer type can be improved by estimating the answer type hierarchically.

  In addition, the answer type estimation device of the present invention can be configured to include learning means for learning the answer type model using the plurality of learning question sentences. Thereby, it is possible to have a function of learning the answer type model.

  Further, in the answer type estimation method of the present invention, the feature quantity extracting means includes a character N gram, a classifier, a suffix, a semantic attribute of a content word appearing at the end of the question sentence, and the question A plurality of learning questions in which a feature amount related to at least one of the last characters of the content word appearing at the end of the sentence is extracted, and the answer type estimation means is assigned an answer type correct answer indicating the type of answer to the question sentence An answer type model that learns correspondence between the feature amount extracted from each sentence and each answer type assigned to each of the plurality of learning question sentences, and an estimation target question sentence whose answer type is unknown From the feature amount extracted by the feature amount extraction means, the answer type of the estimation target question sentence is estimated.

  The answer type estimation program of the present invention is a program for causing a computer to function as each means constituting the answer type estimation apparatus.

  According to the answer type estimation device, method, and program of the present invention, the character N-gram, the classifier, the suffix, the semantic attribute of the content word that appears at the end of the question sentence, and the end of the question sentence as the feature amount Answers to question sentences because it is possible to classify answer types with high accuracy even if there are many types of answer types by using detailed information using features for at least one of the last words of the content word The type can be estimated in detail.

It is a block diagram which shows the functional structure of the answer type estimation apparatus which concerns on 1st Embodiment. It is a figure which shows an example of the morphological analysis result of a question sentence. It is a flowchart which shows the learning process routine in 1st Embodiment. It is a flowchart which shows the estimation process routine in 1st Embodiment. It is a block diagram which shows the functional structure of the answer type estimation apparatus which concerns on 2nd Embodiment. It is a flowchart which shows the estimation process routine in 2nd Embodiment.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

<About answer types>
First, the answer types handled in this embodiment will be described.

  The first is to classify into 13 types of answers. The 13 answer types are “name: person name”, “name: other”, “quantity: other”, “quantity: date”, “quantity: amount”, “quantity: period”, “meaning”, “cause” , “Association”, “reputation”, “method”, “true / false”, and “other”. “Name: Person”, “Name: Other”, “Quantity: Other”, “Quantity: Date”, “Quantity: Amount”, and “Quantity: Period” are the answer types for factoid questions, “Cause”, “association”, “reputation”, “method”, “true / false”, and “other” are the answer types for the non-factoid question.

  The second is to classify into 200 types of answers. The 200 answer types use the extended named entity system, which is a proposed named entity classification system by Sekine of New York University. The extended specific expression system is a classification of specific expressions into all 200 classes. For example, classes of extended specific expressions are assigned to rivers, bridges, railways, and the like. Traditionally, rivers, bridges, railways, and so on have been grouped together as “artifacts”, but they are handled differently in the extended named system.

  The answer types applicable to the present invention are not limited to the above example.

<First Embodiment>
Next, a first embodiment will be described. In the first embodiment, since any of the 13 types of answers and the 200 types of answers can be handled in the same manner, a case where 200 types of answers are applied will be described below.

  The answer type estimation apparatus 10 according to the first embodiment includes a CPU, a RAM, a program for executing an answer type estimation processing routine including a learning process and an estimation process described later, and a ROM storing various data. It consists of a computer equipped with. Further, an HDD may be provided as a storage means.

  As shown in FIG. 1, this computer can be functionally represented by a configuration including a learning unit 11 and an estimation unit 12. The learning unit 11 further includes a feature amount extraction unit 13, a model learning unit, and the like. 14, and the estimation unit 12 can be further represented by a configuration including a feature amount extraction unit 13 and an answer type estimation unit 15. Note that the feature amount extraction unit 13 is a functional unit common to the learning unit 11 and the estimation unit 12.

  The learning unit 11 learns an answer type model 20 for estimating an answer type for an estimation target question sentence whose answer type is unknown, using a plurality of input correct answer data.

  The correct answer data is a pair of a natural language question sentence (text data) and a correct answer type, and the answer type (200 types) is a meaning content that the question sentence asks in a large amount of question sentences collected. The correct answer label indicating is given. An example of the question sentence and the assigned correct answer label is shown below.

・ What is the sun god of Indian mythology, which is regarded as a heaven in Buddhism? (Answer type: God)
・ Which female duo sang songs like “I like you Captain”? (Answer type: Show_Organization)
・ Which clan has inherited the copyrighted work of the Kamakura shogunate? (Answer type: Family)
・ What day is it recommended to spend time quietly avoiding urgent business on six days? (Answer type: Day_Of_Week)
・ When was the Japanese shogunate in Kamakura? (Answer type: Era)
・ What is the average daily calorie intake for Japanese women? (Answer type: Calorie)

  The feature quantity extraction unit 13 extracts a plurality of feature quantities from each correct answer data, and generates a feature quantity vector having each feature quantity as an element. Features include interrogative, interrogative for numerical values, letter N-gram, word N-gram, part-of-speech N-gram, semantic attribute N-gram, classifier / suffix, last classifier / suffix, last noun, last noun The feature amount related to the character, noun, proper noun, and predicate attribute is used. Hereinafter, each feature amount will be described.

  The feature amount related to “interrogative word” is a binary feature amount indicating whether or not a predefined question word is included in the question sentence. When the question word is present in the question sentence, the feature amount value is set to 1. To do. For example, the following question words are defined in advance.

Why, why, why, who, where, when, how much, how, which,
How much, how much, how much, how much, how
How, what, what, how much, how much, what, what, what,
What, what, what, what, what, words that start with "what"

  The feature value related to the “question word about a numerical value” is a binary feature value indicating whether or not a question word about a predefined numerical value is included in the question sentence, and when the question word about the numerical value exists in the question sentence. The feature value is set to 1. For example, questionables for the following numerical values are defined in advance.

How much, how much, how much, how much, how much,
How much, how much, how much

  The feature amount related to “character N-gram” is a binary feature amount that represents the presence of each N-gram of unigram, bigram, and trigram extracted from a sequence obtained by decomposing a question sentence in character units. Since information that is important for answer type estimation may be included in a long noun or the like, information obtained by dividing a question sentence in character units more detailed than word units is used as a feature amount.

  Unigram, for example, for the question sentence “Yamada ’s birthday is”, first, it is broken down into character units of “mountain”, “field”, “no”, “birth”, “raw”, “day”, and “ha”. Then, the feature quantity indicating the presence of each character is extracted as a binary feature quantity. A bigram is a feature value representing the presence of two consecutive characters, and a trigram is a characteristic value representing the presence of three consecutive characters. In the case of a trigram, “no”-“birth”-“raw” and the like are extracted as feature quantities, and the value is 1. When obtaining a bigram and trigram, a symbol (S) representing the beginning of a sentence and a symbol (E) representing the end of a sentence are added to the beginning and end of the series, and then the feature amount is extracted.

  The feature amount related to “word N-gram” is a binary feature amount representing the existence of each N-gram of unigram, bigram, and trigram extracted from a series obtained by dividing the question sentence into words by morphological analysis. The extraction method of each feature amount is the same as that of the character N-gram described above, except that the character unit or the word unit is different. For morphological analysis, JTAG, a morphological analysis engine developed by NTT, can be used.

  The feature amount related to “part of speech N-gram” is a binary feature amount representing the existence of each N-gram of unigram, bigram, and trigram extracted from the series of parts of speech information obtained from the morphological analysis result of the question sentence. It is. For example, FIG. 2 shows the result of a morphological analysis of a question sentence “President Tanaka is how old”. In the example of FIG. 2, each line corresponds to one word. Word (surface), part of speech, standard form, final form (if any), semantic attribute based on Japanese vocabulary system (noun semantic attribute, proper noun attribute, Word attribute: Unless otherwise specified, the meaning attribute is hereinafter referred to as a noun semantic attribute). EOS is a symbol representing the end of a sentence. From this morphological analysis result, a series composed only of part-of-speech information is obtained. However, the interrogative and noun suffixes: nouns and classifiers are important information for estimating the answer type, so the standard form is used instead of the part of speech. In the case of the example in FIG. 2, a series of “noun: unique: surname: organization”, “president”, “continuous particle”, “Number”, and “age” is obtained. From this part-of-speech series, in the same manner as the above-described extraction of the character N-gram, a feature amount representing the presence of each N-gram of unigram, bigram, and trigram is extracted.

  The feature amount related to “semantic attribute N-gram” is a binary feature amount representing the existence of each N-gram of unigram, bigram, and trigram extracted from the sequence of semantic attributes given to each word included in the question sentence. . However, for a word whose part of speech is an interrogative word, the standard form is used instead of the semantic attribute. In addition, if the part of speech has a noun suffix: a noun or a classifier, if a semantic attribute is given, the hierarchy of the semantic attribute is traced to the fourth level from the root (hereinafter referred to as “four levels” Replace it with the semantic attribute of the eye. When a plurality of semantic attributes are assigned to one word, the first one is used. The reason for following the top of the semantic attribute is to generalize the semantic attribute and improve the efficiency of later learning.

  In the example of “How old is President Tanaka” in FIG. 2, words having no semantic attribute are skipped, and “5”, “4 (the fourth layer of 323)”, “2585 (the fourth layer of 2595) are skipped. ) "Series is obtained. A feature quantity representing the presence of a unigram, bigram, and trigram is extracted from the semantic attribute series in the same manner as the character N-gram.

  The feature amount related to “a classifier / suffix” is a binary feature amount indicating the presence of the part of speech for all the classifiers and suffixes in the question sentence obtained by morphological analysis of the question sentence. Also, a feature amount indicating the presence of a standard form of a word whose part of speech is a classifier or a suffix is added. Furthermore, when a semantic attribute is given to a word whose part of speech is a classifier or a suffix, the feature quantity indicating the existence of each semantic attribute and the presence of the semantic attribute in the fourth layer of the given semantic attribute are displayed. The feature quantity to represent is also added. Classifiers (“people”, “sheets”, “books”, etc.) and suffixes (“president”, “director”, “sama”, etc.) are not content words such as nouns and verbs. However, these are effective as a clue to distinguish whether the question is asking for the number of people, asking for the number of sheets, asking for the number of people, or asking for the name of the person. It is used as

  The feature quantity related to the “last classifier / suffix” is the classifier or suffix that appears at the end of the question sentence in the part-of-speech series obtained by morphological analysis of the question sentence (hereinafter referred to as the “last classifier / suffix”). Is a feature amount extracted for the last classifier / suffix based on the fact that there is a high possibility of being directly connected to the answer type. Specifically, the same feature quantity as the feature quantity related to “a classifier / suffix” is extracted, but a prefix indicating that the feature quantity is related to “the last classifier / suffix” is added to the feature quantity name. The feature value.

  As for the feature value related to “last noun”, the last noun appearing in the question sentence (hereinafter referred to as “last noun”) may be directly linked to the answer type, similar to the last classifier / suffix above. This is a feature amount extracted for the last noun based on its high nature. For example, the nouns “侍” and “Geisha” appearing at the end of questions such as “Who is famous for XX?” And “What is the Geisha in XX case?” Are subordinate concepts of people. If it can be determined, it is effective for estimating the answer type of the person name. Therefore, the semantic attribute of the last noun is used as a feature amount. Specifically, for each semantic attribute assigned to the last noun, a feature quantity representing the presence of all the semantic attributes in the middle of the path following the hierarchy of the Japanese vocabulary system to the root is extracted.

  The feature value for “Last Noun Character” is a feature that represents the presence of the last character of the last noun that is extracted based on the fact that the last character of the last noun is also important information for estimating the answer type. Amount. For example, in the example of “How old is President Tanaka” in FIG. 2, the last letter “age” of the last noun “what age” is extracted, and the value becomes 1.

  In the present embodiment, the case where attention is paid to “noun” as the feature amount related to “last noun” and “letter of last noun” has been described. However, the content word may be a word representing lexical meaning. Can be treated as similar feature quantities.

  Features related to “nouns / proprietary nouns / property attributes” are feature values extracted to make use of various contents contained in the question sentence as information, and all of the semantic attributes given to the words included in the question sentence Is a feature amount representing the presence of the. In the example of “How old is President Tanaka” in FIG. 2, feature quantities N-5, N-362, P-67, N-323, N-269, N-2595, N-2589 are extracted, Each value is 1. Here, N- is a prefix added to the noun semantic attribute, and P- is a prefix added to the semantic attribute of the proper noun.

  The model learning unit 14 learns the answer type model 20 that associates the feature amount vector generated from the question sentence of each correct answer data with the correct answer label indicating the correct answer type given to each correct answer data. Specifically, when there are L answer types in total, the score for the correct answer type of the question sentence is higher than the score for the other answer types for the feature amount vector generated from the question sentence. Learn such weights. When L weight vectors corresponding to L answer types are learned and a score is obtained by multiplying the feature vector by the weight vector corresponding to each answer type, the weight vector corresponding to the correct answer type The weight vector is learned so that the score using is the highest.

  Such learning is possible by using a multi-class SVM that is an extension of the SVM to a multi-class problem. In the present embodiment, multi-class SVM is used to learn response type association. For multi-class SVMs, a linear kernel is used. Here, since there are 200 types of answer types, 200 weight vectors are learned, and a set of weight vectors is generated as the answer type model 20.

  The estimation unit 12 estimates the answer type based on the answer type model 20 for the input estimation target question sentence.

  As described in the learning unit 11, the feature amount extraction unit 13 extracts the feature amount from the estimation target question sentence and generates a feature amount vector.

  The answer type estimation unit 15 multiplies each weight vector included in the answer type model 20 generated by the model learning unit 14 and the feature amount vector generated by the feature amount extraction unit 13 to obtain 200 types of responses. Find the score for each type. The answer types are rearranged in order from the answer type corresponding to the weight vector having the highest obtained score, and the top M answer types are output as estimation results in the order of score.

  Next, the operation of the answer type estimation device 10 according to the first embodiment will be described. When a plurality of correct answer data is input to the answer type estimation device 10 in the learning stage, the learning processing routine shown in FIG. Further, when the estimation target question sentence is input to the answer type estimation device 10 in the estimation stage, the estimation processing unit 12 executes a learning process routine shown in FIG. Hereinafter, each process is explained in full detail.

  First, in the learning processing routine, in step 100, the feature amount extraction unit 13 acquires a plurality of input correct answer data, and then in step 102, extracts the feature amount from the question sentence of each correct answer data. Thus, a feature quantity vector having each feature quantity as an element is generated.

  Next, in step 104, the model learning unit 14 learns the correspondence between each of the correct answer labels included in the correct answer data and each of the feature quantity vectors generated in step 102, so that the answer type model 20 is obtained. Generate. The generated answer type model 20 is stored in a predetermined storage area, and the learning processing routine is terminated.

  Next, in the estimation processing routine, in step 110, the feature amount extraction unit 13 acquires the input estimation target question sentence. Next, in step 112, as in step 102, the feature of the estimation target question sentence is acquired. Generate a quantity vector.

  Next, in step 114, the answer type estimation unit 15 obtains a score for each answer type based on the answer type model 20 and the feature vector generated in step 112, and the top M items in order of score. The answer type is output as the estimation result, and the estimation processing routine is terminated.

  As described above, according to the answer type estimation device 10 according to the first embodiment, the character N-gram, the classifier / suffix, the semantic attribute of the last content word, and the last content word By using detailed information using the feature value related to the last character, it is possible to classify the answer types with high accuracy even when there are many types of answer types. Can do. As a result, the answer to the question can be output to the pinpoint, and the efficiency of the user's intellectual activity can be improved.

<Second Embodiment>
Next, a second embodiment will be described. In addition, about the structure similar to the answer type estimation apparatus 10 which concerns on 1st Embodiment, the same code | symbol is attached | subjected and detailed description is abbreviate | omitted.

  As shown in FIG. 5, the computer constituting the answer type estimation device 210 according to the second embodiment can be functionally represented by a configuration including a learning unit 211 and an estimation unit 212. The unit 211 can be further represented by a configuration including the feature amount extraction unit 13, the first model learning unit 14a, and the second model learning unit 14b, and the estimation unit 212 further includes the feature amount extraction unit 13, It can be expressed by a configuration including a first answer type estimation unit 15a and a second answer type estimation unit 15b.

  The 1st model learning part 14a produces | generates the 1st answer type model 20a for estimating which of the above-mentioned 13 types of answer types is using the some correct answer data. The learning method is the same as that of the model learning unit 14 in the first embodiment.

  The second model learning unit 14b generates a second answer type model 20b for estimating which of the 200 types of answer types described above using the plurality of correct answer data. That is, the second model learning unit 14b is the same as the model learning unit 14 in the first embodiment, and the second answer type model 20b is the same as the answer type model 20 in the first embodiment.

  The correct answer data used in the first model learning unit 14a includes correct answer labels corresponding to 13 types of answer types, and the correct answer data used in the second model learning unit 14b is correct answers corresponding to 200 types of answer types. Contains a label. Correspondences between these 13 types of responses and 200 types of responses are predetermined.

  The first answer type estimation unit 15a estimates the answer type of the estimation target question sentence using the first answer type model 20a and the feature quantity vector of the estimation target question sentence. The estimation method is the same as that of the answer type estimation unit 15 in the first embodiment. The answer type estimated here is one of the above-mentioned 13 answer types, and this is referred to as an upper-layer answer type.

  Further, the first answer type estimation unit 15a delivers the estimation result and the feature quantity vector to the second answer type estimation unit 15b when the estimated answer type of the upper hierarchy is the answer type of the factoid type question. If the estimated upper-layer answer type is a non-factoid type answer type, the upper-layer answer types up to the top M are output as estimation results in the order of scores for each upper-layer answer type. Whether the type is a factoid type or a non-factoid type can be determined based on which of the upper-layer answer types has the highest score for each upper-layer answer type.

  The second answer type estimation unit 15b estimates the answer type of the estimation target question sentence using the second answer type model 20b and the feature amount vector of the estimation target question sentence. The estimation method is the same as that of the answer type estimation unit 15 in the first embodiment. The answer type estimated here is one of the above-mentioned 200 answer types, and this is referred to as a lower-layer answer type.

  Further, the second answer type estimation unit 15b uses the estimation result of the upper hierarchy answer type as a constraint when estimating the lower hierarchy answer type. For example, when the estimation result of the answer type of the upper hierarchy is “quantity: date”, the estimation result of the answer type of the lower hierarchy is selected only from those related to the date.

  Next, the operation of the answer type estimation device 210 according to the second embodiment will be described. The learning process routine is the same as the learning process routine in the first embodiment in each of the first model learning unit 14a and the second model learning unit 14b, and thus the description thereof is omitted. Hereinafter, the estimation processing routine will be described with reference to FIG. In addition, about the process same as the estimation process routine in 1st Embodiment, the same code | symbol is attached | subjected and detailed description is abbreviate | omitted.

  In steps 110 and 112, the feature amount extraction unit 13 generates a feature amount vector of the estimation target question sentence.

  Next, in step 200, the first answer type estimation unit 15a estimates the answer type of the upper hierarchy of the estimation target question sentence using the first answer type model 20a and the feature quantity vector of the estimation target question sentence. .

  Next, in Step 202, the first answer type estimation unit 15a determines whether the answer type of the higher hierarchy estimated in Step 200 is a factoid type or a non-factoid type. In the case of the non-factoid type, the process proceeds to step 204, where the first answer type estimation unit 15a outputs the upper M answer types in the higher order in the score order as the estimation result, and ends the estimation processing routine.

  On the other hand, if the factoid type is determined in step 202, the process proceeds to step 206, where the second answer type estimation unit 15b uses the second answer type model 20b and the feature quantity vector of the estimation target question sentence. Under the restriction of the estimation result in step 200, the answer type in the lower hierarchy of the estimation target question sentence is estimated, the answer types in the upper M lower hierarchy are output as the estimation result in order of score, and the estimation processing routine is terminated. To do.

  As described above, according to the answer type estimation device 210 according to the second embodiment, in addition to the effects of the first embodiment, the estimation result of the upper hierarchy answer type is estimated as the lower hierarchy answer type. As a constraint when doing so, it is possible to improve the accuracy of estimating the answer type by hierarchically estimating the answer type.

  Here, an experiment for confirming the effectiveness of the feature amount used in the present embodiment will be described. In this experiment, a five-fold cross-validation was used from data in which 200 types of answer types were assigned to a question sentence on hand. The evaluation scale used was the percentage of questions with the correct answer type estimated at the top. As a result, it was found that when the feature amount was not used, the accuracy was 81.80%, whereas when the feature amount was used, the accuracy was improved to 85.22%. Therefore, the effectiveness of the feature amount used in the present embodiment was confirmed.

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  For example, although the case where the learning unit and the estimation unit are configured by one computer has been described in the above embodiment, the learning unit and the estimation unit may be configured by separate computers.

  Moreover, although the above-mentioned answer type estimation apparatus has a computer system inside, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used. .

  In the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium. The present invention can also be realized by installing a program on a known computer via a medium or a communication line.

10, 210 Answer type estimation device 11, 211 Learning unit 12, 212 Estimation unit 13 Feature amount extraction unit 14 Model learning unit 15 Answer type estimation unit

Claims (5)

From a question sentence in natural language, at least one of a letter N-gram, a classifier, a suffix, a semantic attribute of a content word appearing at the end of the question sentence, and a last character of the content word appearing at the end of the question sentence A feature quantity extracting means for extracting a feature quantity;
The feature amount extracted from each of the plurality of learning question sentences to which the correct answer of the answer type indicating the type of answer to the question sentence, and each of the answer types provided to each of the plurality of learning question sentences An answer that estimates the answer type of the estimation target question sentence based on the answer type model that learned the association with the feature quantity and the feature quantity extracted by the feature quantity extraction means from the estimation target question sentence whose answer type is unknown Type estimation means;
Answer type estimation device. The answer type model includes a first answer type model and a second answer type model in which the estimated answer type is greater than the first answer type model,
The answer type estimating means estimates an answer type of an upper hierarchy of the estimation target question sentence based on the first answer type model and a feature amount extracted from the estimation target question sentence, and the second answer type A lower layer answer type constrained by an estimation result of the upper layer answer type is estimated as an answer type of the estimation target question sentence based on a model and a feature amount extracted from the estimation target question sentence Item 1. An answer type estimation device according to item 1.   The answer type estimation device according to claim 1, further comprising learning means for learning the answer type model using the plurality of learning question sentences. The feature quantity extracting means extracts a character N-gram, a classifier, a suffix, a semantic attribute of a content word appearing at the end of the question sentence, and a last of the content word appearing at the end of the question sentence from a question sentence in a natural language. Extracting features for at least one of the characters,
The answer type estimating means gives the feature amount extracted from each of the plurality of learning question sentences to which the correct answer of the answer type indicating the type of answer to the question sentence is given, and is given to each of the plurality of learning question sentences Of the estimation target question sentence based on the answer type model that has learned the correspondence with each of the answer types that have been determined and the feature quantity extracted by the feature quantity extraction means from the estimation target question sentence whose answer type is unknown Answer type estimation method.   An answer type estimation program for causing a computer to function as each means constituting the answer type estimation device according to any one of claims 1 to 3.