JP2013250926A - Question answering device, method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To output an appropriate response by extracting an object and a response accurately from a question sentence which asks an attribute of an object.SOLUTION: In a learning unit 11, an object model generation unit 12 imparts an object label to show an object part and learns an object model 21 by using a plurality of learning data in which proper nouns and the like are replaced with specific symbols; and an attribute model generation unit 13 learns an attribute model 22 by using the attribute included in each of the learning data and feature quantity extracted from the part other than a surface layer of the object of the learning data. In a response unit 31, an object extraction unit 32 replaces proper nouns and the like of an unknown question sentence with specific symbols and extracts the object part from the unknown question sentence by using the object model 21; an attribute determination unit 33 determines the attribute of the unknown question sentence by using the feature quantity extracted from the part other than the surface layer of the object part of the unknown question sentence, and the attribute model 22; a response output unit 35 retrieves a response to the extracted object part and the determined attribute from the database and outputs it.

Description

  The present invention relates to a question answering apparatus, method, and program, and more particularly, to a question answering apparatus, method, and program for outputting a response to a question sentence for asking a target attribute.

  There are relatively many types of question sentences that ask the target attribute among the question sentences in natural language. For example, “What is Yamada's birthday?” Or “How tall is the Sky Tree?”. Each of them asks the attribute “Birthday” of the subject “Mr. Yamada” and the attribute “Height” of the subject “Sky Tree”. If the target and attributes can be correctly extracted from such a question sentence, for example, by searching the “Birthday” field of “Yamada-san” from the database of person names and birthdays, Mr. Yamada ’s birthday can be obtained. It is possible to answer, or by searching the “height” field of “sky tree” from the database of buildings and their heights, the height of the sky tree can be answered.

  Conventionally, manual rules and machine learning methods have been used as methods for extracting objects and attributes from sentences in natural language. For example, a method of creating a rule for extracting a continuation of nouns or a method of extracting a detected specific expression as a target by using a specific expression extraction technique has been used for extraction of objects. In addition, attribute extraction is a major topic in natural language processing, and research has been conducted on extracting attributes related to specific objects from many documents. For example, a method for extracting attributes of a specific object from a document on the web has been proposed (see, for example, Non-Patent Document 1).

Naoki Tani, Danelle de Bolegala, Mitsuru Ishizuka, “Attribute Extraction from the Web based on Correlation between Attributes”, Japanese Society for Artificial Intelligence, 2010.

  However, the prior art is intended to extract attributes from a large collection of documents, and does not estimate which attribute is being questioned for an object in a single question sentence.

  When extracting a target from a question sentence, conventionally, as described above, noun phrases, proper nouns, proper expressions, etc. in the question sentence are detected by using noun continuation or proper expression extraction technology, and these are detected. The targeted approach is taken. For example, if the question is “What is Yamada's birthday?”, The only proper noun “Yamada” is selected as the target.

  However, there are various ways to ask questions in natural language. For example, in the case of the question “What was the team born in XX team?”, The target is “Yamada”, but in the conventional method, “XX team” and “Yamada” are acquired as targets. In this case, it is not easy for the computer to determine which target attribute the question is about. For this reason, it is necessary to deal with ad hoc rules such as determining a word having a longer word length as a target or giving priority to a word after “no”. In addition, when an unknown word is included in a question, the word cannot be divided well by morphological analysis, and extraction of the target may fail. In addition, when questions are written by the inversion method etc. (for example, “When was it born? Yamada, XX team?”), It is difficult to predetermine rules corresponding to all aspects. Therefore, there are cases where the target cannot be extracted appropriately.

  As for attribute extraction, there are many studies to extract target attributes from a set of documents, but no method has been proposed for identifying attributes included in a question sentence that queries the target attributes.

  The present invention has been made in consideration of the above facts, a question answering apparatus, a method, and a method capable of accurately extracting a target and an attribute from a question sentence for asking a target attribute and outputting an appropriate response, and The purpose is to provide a program.

  In order to achieve the above object, the question answering apparatus of the present invention learns using a plurality of learning question sentences in which a target label indicating that a target is attached to a target part of a question sentence for asking a target attribute. A target extraction unit that extracts a target portion from an unknown question sentence whose target and attribute are unknown based on the target model that has been determined; each of the attributes that each of the plurality of learning question sentences asks; and the plurality of learning questions Attribute determination for determining an attribute included in the unknown question sentence based on the attribute model learned to associate with each feature quantity extracted from each sentence and the feature quantity extracted from the unknown question sentence A response for retrieving information corresponding to the unknown question sentence from a predetermined database using the means, the target portion extracted by the target extraction means, and the attribute determined by the attribute determination means It is configured to include a power means.

  According to the question answering device of the present invention, the target extraction unit is learned using a plurality of learning questions with a target label indicating that the target is included in a target part of a question that asks for a target attribute. The target portion is extracted from the unknown question sentence whose target and attribute are unknown based on the target model, and the attribute determination means includes each of the attributes asked by the plurality of learning question sentences and the plurality of learning questions. The attribute included in the unknown question sentence is determined based on the attribute model learned to be associated with each of the feature quantities extracted from each sentence and the feature quantity extracted from the unknown question sentence. And a response output means searches the information corresponding to an unknown question sentence from a predetermined | prescribed database using the object part extracted by the object extraction means, and the attribute determined by the attribute determination means, and outputs a response.

  Thus, based on the target model and attribute model learned using the learning question sentence, sufficient variation of the learning question sentence to extract the target and determine the attribute for the unknown question sentence By using, it is possible to accurately extract a target and an attribute from a question sentence for asking about the target attribute and output an appropriate response.

  In addition, the target extracting unit replaces a surface layer of a word whose part of speech of the unknown question sentence is a predetermined part of speech as a target candidate with a specific symbol, and then the part of speech is a target candidate in each of the learning question sentences. The target portion can be extracted from the unknown question sentence based on the target model learned by replacing the surface layer of a word having a predetermined part of speech with the specific symbol. As a result, the target model is generalized, and unknown words can be appropriately extracted as target portions.

  Further, the attribute determination means includes an attribute model learned using a feature quantity extracted from a portion other than the surface layer of each target portion of the plurality of learning question sentences, and a feature extracted from the unknown question sentence Based on the quantity, an attribute included in the unknown question sentence can be determined. As a result, it is possible to suppress the influence of surface layer bias or the like in the combination of the target and the attribute appearing in the learning question sentence, and improve the attribute determination accuracy.

  The question answering device of the present invention includes a feature word list for each attribute generated by extracting words characteristically appearing in each of the learning question sentences according to the attributes, and the unknown question sentence And an attribute verification unit that verifies the likelihood of the attribute determined by the attribute determination unit based on the word. Thereby, the reliability of the attribute determination result can be improved.

  Further, the question answering apparatus of the present invention uses the plurality of learning question sentences to generate the target model and the attribute model, and the feature words using the plurality of learning question sentences. A list generation means for generating a list can be included. Thus, it is possible to have a function of generating the above-described target model, attribute model, and feature word list.

  In the question answering method of the present invention, the object extracting means is learned by using a plurality of learning question sentences in which a target label indicating that a target is included in a target part of a question sentence for asking a target attribute. A target portion is extracted from an unknown question sentence whose target and attribute are unknown based on the target model, and an attribute determination unit includes each of the attributes asked by each of the plurality of learning question sentences and the plurality of learning questions. Determining an attribute included in the unknown question sentence based on the attribute model learned to associate with each of the feature quantities extracted from each sentence and the feature quantity extracted from the unknown question sentence; A method in which an output unit searches for information corresponding to the unknown question sentence from a predetermined database using the target portion extracted by the target extraction unit and the attribute determined by the attribute determination unit, and outputs a response In .

  The question answering program of the present invention is a program for causing a computer to function as each means constituting the question answering apparatus.

  According to the question answering apparatus, method, and program of the present invention, based on the target model and the attribute model learned using the learning question sentence, the target extraction and attribute determination are performed on the unknown question sentence. Therefore, by using a sufficiently-variable learning question sentence, it is possible to accurately extract a target and an attribute from a question sentence for asking a target attribute and to output an appropriate response.

It is a block diagram which shows the functional structure of the question answering apparatus which concerns on this Embodiment. It is a figure which shows an example of learning data. It is a figure which shows an example of the learning data after provision of a target label and replacement to a specific symbol. It is a figure which shows an example of the template for producing | generating a feature from learning data. It is a figure which shows an example of an attribute number and the extracted feature-value. It is a figure which shows an example of the morphological analysis result of an unknown question sentence. It is a figure which shows an example of the unknown question sentence after the replacement to a specific symbol. It is a figure which shows an example which provided the object label to the unknown question sentence. It is a flowchart which shows the learning process routine in this Embodiment. It is a flowchart which shows the response process routine in this Embodiment.

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

  A question answering apparatus 10 according to the present embodiment is a computer including a CPU, a RAM, a program for executing a question answering process routine including a learning process and a response process described later, and a ROM storing various data. It is configured. 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 a response unit 31. The learning unit 11 further includes a target model generation unit 12 and an attribute model generation. The response unit 31 further includes a target extraction unit 32, an attribute determination unit 33, an attribute verification unit 34, and a response output unit 35. Can be expressed in a configuration including

  First, each part of the learning unit 11 will be described in detail.

  The target model generation unit 12 acquires a plurality of question sentences asking for questions about the target attribute. Objects and attributes are already known in the learning data. An example of learning data is shown in FIG. The example of FIG. 2 is a question sentence for asking an attribute “url” of a target “ΔTV”. Each of such learning data is subjected to morphological analysis to obtain information on the surface layer and part of speech of each word. In the learning data after the morphological analysis, a target label indicating that the target word is a target is given to the target word. In the example of FIG. 2, the question “△△ TV” is included in the question without fail, so that “△△ TV” in the learning data for the target word “△△ TV”. The target label is given to the appearance position of.

  Further, the target model generation unit 12 replaces a surface layer of a word that can be a target of each of the learning data, for example, a word whose part of speech is a part of speech predetermined as a target candidate such as a proper noun with a specific symbol. FIG. 3 shows an example of learning data in which a target label is assigned to a target word and the surface layer of the target word is replaced with a specific symbol. In the example of FIG. 3, the first field is the surface layer, the second field is the part of speech, and the third field is the target label. As target labels, “B-Tagged” and “I-Tagged” representing target words and “O” representing non-target words are used. B means Begin (beginning of object), I means Inside (in object), and O means Other. In addition, an asterisk (*) is used as a specific symbol to replace a surface layer of a word that can be a target. Here, JTAG by NTT is used for the morphological analysis, and the part of speech is replaced with * for the noun: name: proper, noun, noun suffix: noun, number, noun: proper, noun: adjective, and undef word. Yes. By replacing proper nouns with specific symbols in this way, for example, parts where the part of speech is a noun word followed by the part of speech is a noun: proper: local word, etc. "B-Tagged" and "I-Tagged").

  Further, the target model generation unit 12 uses the learning data that has been subjected to the above-described processing to generate a target model 21 for assigning a target label to an unknown question sentence whose target and attribute are unknown by a sequence labeling technique. learn. Specifically, learning is performed using conditional random fields (CRF). CRF is a common technique for sequence labeling such as proper expression extraction. At the time of learning, the CRF generates a feature from the learning data as described above and performs learning. In order to generate this feature, for example, a template as shown in FIG. 4 can be used as a template describing a feature generation method. This template is described in accordance with the software format of CRF ++. When a label is estimated for a certain word, it indicates that a feature is generated using information of three words before and after three words. The target model 21 is learned by CRF using the feature generated by the template from the learning data.

  The attribute model generation unit 13 assigns an attribute number indicating which of the N attributes the word representing the attribute included in each learning data belongs to each of the learning data. The number N of attributes is determined by the number of attributes included in all the learning data used. As attributes, for example, birthDate, Category, MVP, RadioStation, actors, address, affiliation, age, airdate, airtime, etc. can be used, and the respective attribute numbers are, for example, 1, 2,. be able to.

  Moreover, the attribute model production | generation part 13 extracts a feature-value for every learning data from the part which excluded the surface layer of the word used as object part among each surface layer and part of speech of learning data. As the feature amount, a word unigram, a part of speech, and a word bigram are used. All of these feature quantities have a binary feature of 1 if the feature exists and 0 if the feature does not exist.

  For example, in the case of learning data “Taro Kono's birthday” (target: Taro Kono, attribute: birthDate), first, the surface layer of the target portion is excluded from the feature extraction target. By doing so, if there are many questions in the learning data that ask about “Birthday” of “Taro Kono”, the attribute is determined to be “Birthday” only by the character string “Taro Kono”. Can be prevented. Specifically, the word series “Kono”, “Taro”, “No”, “Birthday”, and “Ha” is obtained from the morphological analysis result of the learning data “Taro Kono's birthday”. In addition, “noun: proper: surname: organization”, “noun: first name: proper”, “case particle: union”, “noun”, and “continuous particle” are obtained as part of speech of each word. From this, as the word unigram, the presence of each of the words “no”, “birthday”, and “ha” excluding the target “Taro Kono” is extracted as a binary feature amount. Also, the presence of “no-birthday”, “birthday-ha”, and “ha-E” as word bigrams is extracted as a binary feature amount. “E” is a symbol representing the end of a sentence. In addition, the presence of each of “noun: proper: surname: organization”, “noun: first: proper”, “case particle: union”, “noun”, and “continuous particle” is extracted as a binary feature amount. Note that the existence of a specific expression type may be added as a feature quantity using a specific expression extractor.

  The specific feature amount obtained from the learning data “Taro Kono's birthday” is, for example, the feature amount as shown in FIG. In the example of FIG. 5, the first field is an attribute number. For example, when the number N of attributes is 119, the attribute number is any one of 1 to 119, and in the example of FIG. The field after the attribute number is the feature quantity, and the place where the feature quantity value is 0 is omitted. For example, 4580 corresponds to the word bigram “no-birthday” and 4707 corresponds to the word bigram “ha-E”. Further, 10784 corresponds to the existence of “noun”, and 10788 corresponds to the existence of “noun: name: unique”.

  Further, the attribute model generation unit 13 uses the feature amount extracted as described above to determine the attribute to which the word representing the attribute included in the unknown question sentence belongs by Multiclass SVM (support vector machine). The model 22 is learned. Multiclass SVM is an extension of SVM, which has high classification performance, to a multiclass problem. Specifically, when there are N attributes in total, a weight that makes a score for a correct attribute higher than scores for other attributes is learned for a certain feature amount. N weight vectors corresponding to N attributes are learned, and when a certain feature vector is multiplied by each weight vector, the score when the weight vector corresponding to the correct attribute is used is the highest. Thus, the weight vector is determined. In the case of the learning data example “Taro Kono's birthday”, among the attributes 1 to 119, 119 weight vectors are learned so that the score for the attribute number 18 is the highest. . This weight is determined so that the correct attribute of the entire learning data has a higher score for each learning data.

  Note that the attribute model 22 is a model for determining which of the N attributes the word representing the attribute included in the unknown question sentence belongs to. Therefore, which of the words representing the attribute included in the unknown question sentence is which If the attribute does not belong to the attribute, the attribute determination unit 33 described later determines that the attribute is incorrect. For example, even if the question is “Taro Kono's debut date”, if there is no attribute corresponding to the word “debut date” representing the attribute, the attribute may be determined as birthDate.

  The feature word list generation unit 14 generates a feature word list 23 used in an attribute verification unit 34 described later in order to prevent the above-described malfunction. The feature word list 23 is a list of words that appear characteristically in a part other than the target part of the learning data according to the attribute, for each attribute. For example, if the determined attribute is “birthDate”, if it is a question sentence containing words such as “birthday” or “born”, the attribute is recognized (accepted) as birthDate, and so on. For example, a feature word list 23 that can obtain a verification result that is not birthDate (rejection) is generated by a method using a statistical test.

  Specifically, first, the appearance frequency of each word in the learning data is acquired for each attribute. Then, each word that appears in the attribute is compared with the appearance frequency in the attribute other than the attribute, and it is confirmed by the likelihood ratio test whether or not it appears significantly more in the attribute. As a result of the test, if it is found that the number is significant (for example, p <0.001), the word is added to the feature word list 23 as a feature word of the attribute. Here, the appearance frequency of the word is calculated by excluding the target word as before.

  For example, as described above, when the number N of attributes is 119, 119 feature word lists are generated. An example of a part of the feature word list (attributes “birthDate” and “nearestStation”) is shown below.

birthDate: birthday, date, birth, date of birth, birthday
nearestStation: Dome, nearest, nearest, stadium, line, way, near, get off, train, station

  Next, each part of the response unit 31 will be described in detail.

  Similar to the processing for the learning data in the target model generation unit 12, the target extraction unit 32 performs morphological analysis on the input unknown question sentence, assigns a target label, and replaces a target word with a specific symbol. The target model 21 is used to extract the target from the unknown question sentence.

  For example, FIG. 6 shows an example of a morphological analysis result when an unknown question sentence “Taro Kono's birthday is” is input, and FIG. 7 shows an example of data after replacement processing with a specific symbol (*). . The difference from the processing for learning data is that the target word is unknown at this point, so the B-Tagged and I-Tagged target labels are not assigned, and all the O target labels are assigned. It is. Also, the third field before the target label field contains a word surface layer. The third field is a field that is not used in the estimation of the target label and the attribute, and is included as additional information indicating which word each line is for.

  The unknown question sentence data processed in this way is processed by the target model 21 (also referred to as a decoding process), thereby assigning B-Tagged and I-Tagged target labels to the target word. The result of assigning the target label is shown in FIG. The first row is the likelihood of the estimated sequence. The last field of each row is the marginal probability of each target label. This is a value indicating the certainty (reliability) of the estimated target label.

  Further, the target extraction unit 32 extracts a target portion by extracting a portion obtained by combining the surface layers of words to which B-Tagged and I-Tagged are assigned as a character string. In the example of FIG. 8, “Taro Kono” can be extracted. Note that there is a possibility that B-Tagged and I-Tagged target labels are given to a plurality of locations across the O target label due to the algorithm. In that case, it suffices to extract a portion having the highest reliability average. In the example of FIG. 8, the reliability of “Taro Kono” can be calculated as (log (0.999666) + log (0.999650)) / 2.

  Similar to the feature amount extraction process in the attribute model generation unit 13, the attribute determination unit 33 extracts feature amounts from a portion excluding the surface layer of the target portion of the input unknown question sentence, and uses the attribute model 22. The attribute included in the unknown question sentence is determined from the attribute number corresponding to the weight vector with the highest score.

  The attribute verification unit 34 refers to the feature word list 23, and verifies the determination result by determining whether or not the feature word corresponding to the attribute determined by the attribute determination unit 33 is included in the unknown question sentence. Do. For example, if an attribute called birthDate is determined for an unknown question sentence that says “Taro Kono's birthday is”, as in the above example, the feature word is “Tanjobi, Sun, Born, Date of birth” Since each feature word “birthday” is defined in the feature word list 23, it is verified whether any of these is included in the unknown question sentence. It can be determined by character string matching using regular expressions.

  Note that the attribute verification unit 34 may use one-class SVM used for singular value detection. This can be done by modeling the feature space of each class with one-class SVM using only the learning data corresponding to each attribute among the learning data used in Multiclass SVM.

  In attribute determination and verification, accuracy can be improved by using the relationship with the extracted object. For example, when the target is a person's name, there is no problem in determining that the date is an attribute, but if the target is a place, the birthDate cannot be taken as an attribute. In order to realize this, it is necessary to determine the target class. Prepare a correspondence table of the target, its target class, and the attribute that the target class can take in advance, and verify whether the determined attribute is an attribute that can be taken for the extracted target Can do.

  The response output unit 35 searches for an answer to the question from the database by using the target extracted by the target extraction unit 32 and the attribute determined by the attribute determination unit 33 and verified by the attribute verification unit 34. Answer is output as a response. For example, when an RDB (Relational Database) is used as a database, an inquiry sentence such as “select birthDate from database where name like Taro Kono” is created as an inquiry sentence by SQL. This is to obtain the “birthday” of “Taro Kono”. The RDB is searched based on this inquiry sentence, and if the answer exists in the RDB, the answer is output.

  Next, the operation of the question answering apparatus 10 of the present embodiment will be described. When a plurality of learning data is input to the question answering apparatus 10 in the learning stage, the learning processing routine shown in FIG. Further, when an unknown question sentence is input to the question answering device 10 in the response stage, the learning process routine shown in FIG. Hereinafter, each process is explained in full detail.

  First, in the learning processing routine, in step 100, the target model generation unit 12 acquires a plurality of question sentences as questions for the attributes of the target and the target whose attributes are known, and performs morphological analysis on each of the learning data. , Get information on the surface and part of speech of each word. Next, in step 102, the target model generation unit 12 assigns a target label indicating a target word in the learning data after morphological analysis. Next, in step 104, the target model generation unit 12 replaces the surface layer of words that can be the targets of the learning data with specific symbols. Next, in step 106, the target model generation unit 12 assigns a target label to an unknown question sentence whose target and attribute are unknown by a sequence labeling method using the learning data subjected to the processing of steps 100 to 104. The target model 21 for learning is learned. The target model 21 generated by learning is stored in a predetermined storage area.

  Next, in step 108, the attribute model generation unit 13 assigns an attribute number indicating an attribute included in each learning data to each learning data. Next, in step 110, the attribute model generation unit 13 extracts a feature amount for each learning data from the surface layer and the part of speech of the learning data excluding the surface layer of the word that is the target portion. Next, in step 112, the attribute model generation unit 13 learns the attribute model 22 for determining the attribute of the unknown question sentence by Multiclass SVM using the feature amount extracted as described above. The attribute model 22 generated by learning is stored in a predetermined storage area.

  Next, in step 114, the feature word list generation unit 14 acquires the appearance frequency of each word other than the target word in the learning data for each attribute, and compares it with the appearance frequency in other attributes. Words frequently appearing in the attribute are extracted, and a feature word list 23 for each attribute is generated. The generated feature word list 23 is stored in a predetermined storage area, and the learning processing routine is terminated.

  Next, in the response processing routine, in step 120, the target extraction unit 32 performs a morphological analysis on the input unknown question sentence, assigns a target label, and replaces the surface layer of the target word with a specific symbol. Then, the target part is extracted from the unknown question sentence using the target model 21.

  Next, in step 122, the attribute determination unit 33 extracts a feature amount from the portion of the input unknown question sentence excluding the surface layer of the target word extracted in step 120, and uses the attribute model 22. The attribute of the unknown question sentence is determined from the attribute number corresponding to the weight vector with the highest score.

  Next, in step 124, the attribute verification unit 34 refers to the feature word list 23 and determines whether or not the feature word corresponding to the attribute determined in step 122 is included in the unknown question sentence. The verification result is verified.

  Next, in step 126, the attribute verification unit 34 determines whether the verification result in step 124 is accepted or rejected. In the case of rejection, the process proceeds to step 128, where it is determined whether or not there is a next score in the attribute determination result. If there is, the process returns to step 124, and the score is the next score. The verification process is repeated for the determination result.

  If it is determined in step 126 that the request is accepted, the process proceeds to step 130 where the response output unit 35 uses the object extracted in step 120 and the attribute accepted in step 126. Search the database for answers to questions.

  Next, in step 132, the response output unit 35 determines whether an answer corresponding to the target and the attribute has been searched from the database. If a search is made, the process proceeds to step 134, the searched answer is output as a response, and the response processing routine is terminated. If no search is made and if a negative determination is made in step 128, the process proceeds to step 136, an error is output, and the response processing routine is terminated.

  As described above, according to the question answering apparatus of the present embodiment, since the target is extracted from the unknown question sentence based on the target model learned using the learning data to which the target label is assigned, If there are enough variations, it will be possible to respond to various question descriptions such as inverted sentences. In addition, an unknown word can be recognized as a target by learning a sequence after replacing a word that can be a target such as a proper noun in the learning data with a specific symbol and generalizing it. In addition, even if there are a plurality of target candidates, it is not necessary to describe an ad hoc rule for selection by selecting the one that is probabilistically the target.

  In addition, based on the attribute model that learned the correspondence between the learning data to which the attribute number indicating the attribute is assigned and the feature quantity extracted from the portion excluding the target surface layer of the learning data, the attribute of the unknown question sentence is Since the determination is made, the attribute determination accuracy is improved. Furthermore, since the determination result is verified based on the feature word list for each attribute, the reliability of the determination result can be improved.

  As a result, it is possible to accurately extract the target and the attribute from the question sentence asking the target attribute and output an appropriate response. As a result, the efficiency of the user's intellectual activities can be increased.

  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, in the above embodiment, when extracting a feature value from an unknown question sentence in order to determine an attribute of the unknown question sentence, the case where the feature quantity is extracted from a part other than the surface layer of the target part has been described. The feature amount may be extracted without excluding the surface layer of the target portion.

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

  In addition, the above-described question answering apparatus has a computer system therein, but 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.

DESCRIPTION OF SYMBOLS 10 Question response apparatus 11 Learning part 12 Target model generation part 13 Attribute model generation part 14 Feature word list generation part 21 Target model 22 Attribute model 23 Feature word list 31 Response part 32 Target extraction part 33 Attribute determination part 34 Attribute verification part 35 Response Output section

Claims (8)

An unknown question whose target and attribute are unknown based on a target model that has been learned using a plurality of learning questions with a target label attached to the target part of the question that asks the target attribute Target extraction means for extracting a target part from a sentence;
An attribute model that learns the correspondence between each of the attributes asked by each of the plurality of learning question sentences and each of the feature quantities extracted from each of the plurality of learning question sentences, and extracted from the unknown question sentence Attribute determination means for determining an attribute included in the unknown question sentence, based on the feature amount
A response output means for searching for information corresponding to the unknown question sentence from a predetermined database and outputting a response using the target portion extracted by the target extraction means and the attribute determined by the attribute determination means;
Including question answering device.   The target extracting means replaces a surface layer of a word whose part of speech is a part of speech predetermined as a target candidate with a specific symbol and the part of speech of the unknown question sentence as a target candidate in each of the learning question sentences. The question answering apparatus according to claim 1, wherein the target part is extracted from the unknown question sentence based on a target model learned by replacing a surface layer of a word having a defined part of speech with the specific symbol.   The attribute determination means includes an attribute model learned using a feature amount extracted from a portion other than the surface layer of each target portion of the plurality of learning question sentences, a feature amount extracted from the unknown question sentence, The question answering device according to claim 1 or 2 which judges an attribute contained in said unknown question sentence based on.   The question answering device according to claim 1, further comprising learning means for generating the target model and the attribute model using the plurality of learning question sentences.   The attribute determination based on a feature word list for each attribute generated by extracting words characteristically appearing in each of the learning question sentences according to the attribute, and words included in the unknown question sentence 5. The question answering apparatus according to claim 1, further comprising attribute verification means for verifying the likelihood of the attribute determined by the means.   The question answering apparatus according to claim 5, further comprising a list generation unit configured to generate the feature word list using the plurality of learning question sentences. Based on a target model learned by using a plurality of learning question sentences in which a target label indicating that the target extraction means is a target in a target part of a question sentence that asks a target attribute Extract the target part from the unknown unknown question sentence,
An attribute determination unit that learns an association between each of the attributes that each of the plurality of learning question sentences asks and each of the feature values extracted from each of the plurality of learning question sentences; Based on the feature amount extracted from the unknown question sentence, determine an attribute included in the unknown question sentence,
A response output unit searches for information corresponding to the unknown question sentence from a predetermined database using the target portion extracted by the target extraction unit and the attribute determined by the attribute determination unit, and outputs a response. Question answering method.   A question answering program for causing a computer to function as each means constituting the question answering apparatus according to any one of claims 1 to 6.