JP6674411B2 - Utterance generation device, utterance generation method, and utterance generation program - Google Patents

Description

  The present invention relates to an utterance generation device, an utterance generation method, and an utterance generation program.

  There are roughly two types of dialog systems, which are divided into a task-oriented dialog system and a non-task-oriented dialog system (chat dialog system). In recent years, the entertainment property, daily conversation with a robot, and the like have attracted attention, and a chat dialogue system has been actively studied.

  Answer utterances that convey the understanding of the system to the user are important for the conversation to proceed. For example, in the case of a task-oriented dialogue system for making a hotel reservation, the user can confirm his / her understanding of the system by transmitting the result of understanding the user's utterance as "I want to stay in a day." With a chat dialogue system, by repeating the content uttered by the user, it can be shown that the user understands the story.

  In a chat dialogue system, a response utterance is often generated by extracting a keyword from a user utterance and applying the extracted keyword to a template. For example, a keyword "Mt. Fuji" is extracted from the utterance "Mt. Mt. Fuji" and applied to the pattern "Is ..", so that a response utterance "Mt. Fuji" can be generated.

  A specific description of this technique is disclosed in Non-Patent Document 1.

  In recent years, as disclosed in Non-Patent Document 2, a semantic structure (a structure including a predicate and its terms) called a predicate-term structure is extracted from a user utterance, and a response utterance is generated based on the extracted predicate-term structure. Techniques for generating are also being studied. For example, by extracting the predicate-argument structure of "User-Ga Fuji-Ni-Go" from the user utterance "I went to Mt.Fuji", by changing the expression at the end, a response utterance of "I went to Mt.Fuji" is generated. be able to.

J. Weizenbaum, "ELIZA-a computer program for the study of natural language communication between man and machine", Communications of the ACM, vol. 9, pp. 36-45, 1966. Ryuichiro Higashinaka, Kenji Imamura, Toyomi Meguro, Chiaki Miyazaki, Nozomi Kobayashi, Hiroaki Sugiyama, Toru Hirano, Hoshiro Makino, Yoshihiro Matsuo, "Towards an open domain conversational system fully based on natural language processing, In Proc. COLING, pp. 928- 939, 2014.

  Since the response utterance that can be generated by the existing chat dialogue system is generated from the words, the predicate term structure, and the like in the user's utterance, only the response utterance that is superficially expressed in the user's utterance can be generated. For this reason, the user is given an impression that the user's utterance is simply generated repeatedly, and it is not possible to indicate that the user is listening to the talk more deeply.

  An object of the present invention is to provide an utterance generation device, an utterance generation method, and an utterance generation program that can obtain information outside a word that is not directly included in a user utterance in a chat dialogue system.

  In order to achieve the above object, an utterance generation device according to the present invention co-occurs in an input unit for inputting a user utterance and example data which is a set of a set of a user utterance and extraordinary information, or in a corpus document. Based on a set of predicate-argument structures, a search that generates candidates for the non-verbal information from the unspoken information corresponding to the user utterance input by the input unit or the predicate-arrangement structure corresponding to the user utterance And the user utterance input by the input unit, and a ranking model for calculating a score representing the likelihood of information outside of the user utterance with respect to the user utterance. And an information ranking section for calculating the score for each of the information candidates.

  In addition, based on a type identification model for identifying the type of the implicit information, a type identification unit that identifies and assigns the type to each of the implicit information, and the type is assigned. A type filter unit that outputs, from the unspoken information, the unspoken information satisfying a predetermined condition regarding the type may be further included.

  The apparatus further includes a type identification model creation unit that creates the type identification model based on data that includes a set of the wording information to which the type is added in advance, and the type identification unit includes the type identification model. Based on the type identification model created by the creating unit, the type may be identified and assigned to each of the unspoken information.

  In addition, an output unit that outputs, as a response utterance, the unspoken information whose score satisfies a predetermined condition among the unspoken information candidates for which the score has been calculated may be further included.

  Further, the information processing apparatus may further include an expression conversion unit that converts each of the unspoken information output by the output unit into an utterance sentence.

  Further, based on a combination of the user utterance and the non-verbal information of the positive example and a pair of the user utterance and the non-verbal information of the negative example, each of the user utterance and the non-verbal information is subjected to morphological analysis. Then, the ranking model is created based on whether the combination of the stems of the obtained morphemes and the combination of the stems are obtained from a positive example set or a negative example set. Further comprising a ranking model creating unit, wherein the unspoken information ranking unit performs a morphological analysis on each of the user utterance input by the input unit and the unspoken information candidates generated by the search unit. And calculating the score for the candidate of the unspoken information based on the combination of the stems of the obtained morphemes and the ranking model created by the ranking model creating unit. Unishi and may be.

  In order to achieve the above object, an utterance generation method according to the present invention is an utterance generation method in an utterance generation device including an input unit, a search unit, and an unranked information ranking unit, wherein the input unit generates a user utterance. Inputting, the search unit is input by the input unit based on example data, which is a set of a set of user utterances and extraordinary information, or a set of predicate term structures co-occurring in a corpus document. Generating a candidate for the unspoken information from the unspoken information corresponding to the user utterance or the predicate item structure corresponding to the user utterance; and The unspoken information generated by the search unit based on the input user utterance and a ranking model for calculating a score indicating the likelihood of the unspoken information for the user utterance For each candidate, including a step of calculating the score.

  In order to achieve the above object, an utterance generation program of the present invention is a program for causing a computer to function as each unit of the utterance generation device.

  According to the present invention, it is possible to obtain extraordinary information that is not directly included in a user utterance in a chat dialogue system.

It is a block diagram showing the whole utterance generation device composition concerning an embodiment. FIG. 3 is a block diagram illustrating a functional configuration of a learning unit according to the embodiment. FIG. 3 is a block diagram illustrating a functional configuration of an utterance generation unit according to the embodiment. 9 is a flowchart illustrating a flow of an utterance generation process according to the embodiment.

  Hereinafter, the present embodiment will be described with reference to the drawings.

  The utterance generation device according to the present embodiment generates an utterance that is not limited to the user utterance by generating a response utterance based on information that can be estimated from the user utterance. In the present embodiment, information that can be estimated from the user's utterance is referred to as “unspoken information”, and even if the “unspoken information” is not explicitly expressed in the utterance, it is generally understood by humans. Information that can be ".

  Specifically, the utterance generation device according to the present embodiment receives a user's utterance, estimates extraordinary information, and generates a response utterance using the extraordinary information. In addition, the utterance generation device according to the present embodiment generates a response utterance, particularly, by generating an utterance that confirms unsolicited information or asks a question. In estimating the word-to-word information, candidates of word-to-word information are listed by (method 1) and (method 2) shown below, and a score representing the likelihood of a candidate pair of a user utterance and each word-to-word information is listed. Is calculated, and the information candidates are estimated by ranking the information candidates.

(Method 1) Data containing a set of pairs of user utterances and spontaneous information is prepared, and user utterances in the data are searched for using the input user utterances. Are listed.

(Method 2) Prepare data of a set of sets of predicate-argument structures that co-occur in a document (eg, <raining, wet laundry>) and search for a predicate-argument structure similar to the user's utterance. Among them, the predicate-argument structure combined with the searched predicate-arrangement structure is regarded as implicit information, thereby listing candidates of implicit information.

  In the methods (method 1) and (method 2), candidates of unspoken information for the user utterance are enumerated, and the enumerated candidates are ranked to estimate the unspoken information. Specifically, a set of a user utterance and correct non-word information (information of a positive example) and a set of a user utterance and incorrect word information (negative example of a word) are recorded. Based on the data, a regression model for estimating the likelihood of information outside of the user utterance is learned. Using the learned regression model, the likelihood of candidate information outside the word for the utterance is estimated as a score, and ranking is performed.The information outside the word whose score is equal to or greater than a predetermined threshold is calculated as the word outside the word for the user utterance. Estimate as information.

  Since various types of information can be included in the extraneous information estimated as the extraneous information for the user utterance, if all the extraneous information estimated is output as the response utterance, May be uncomfortable. For example, in response to a user's utterance "I was ranked first in the test", when estimating extra information that "the user wants to be praised," the user responds "I want to be praised." May be uncomfortable. Therefore, after estimating the type of the estimated extraordinary information based on the predefined type of the extraordinary information, only the specific type of extraneous information is output as a response utterance.

  With respect to the estimated extraordinary information, the extraordinary information is used as it is (for example, declarative sentence), or the end is confirmed (for example, "~ What is"), and the question format (for example, "~ What is it? ") And output the response utterance by using it as the utterance in the converted form.

  FIG. 1 is a block diagram illustrating an overall configuration of an utterance generation device 10 according to the present embodiment. As illustrated in FIG. 1, the utterance generation device 10 according to the present embodiment includes a learning unit 12 and an utterance generation unit 14. The learning unit 12 creates data and a model required by the utterance generation unit 14. Further, the utterance generation unit 14 generates a set of a plurality of response utterances for the user utterance based on the data and the model created by the learning unit 12 in a state where the likelihood score is given.

  FIG. 2 is a block diagram illustrating a functional configuration of the learning unit 12 of the utterance generation device 10 according to the present embodiment. The learning unit 12 includes a corpus document co-occurrence data creation unit 20, a corpus document co-occurrence data storage unit 22, a ranking model creation unit 24, a ranking model storage unit 26, a type identification model creation unit 28, and a type identification model storage unit 30. It has.

  FIG. 3 is a block diagram illustrating a functional configuration of the utterance generation unit 14 of the utterance generation device 10 according to the present embodiment. The utterance generation unit 14 includes an input unit 40, an example data storage unit 42, a search unit 44, an information ranking unit 46, a type identification unit 48, a type filter unit 50, an expression conversion unit 52, and an output unit 54. ing. The search unit 44 includes an example search unit 44a and a co-occurrence search unit 44b.

  The arrows in FIGS. 1 and 2 indicate the input / output relationship of each unit. A broken line indicates that each unit uses the model or data.

  The corpus-document co-occurrence data creating unit 20 is configured to select candidates of the unspoken information from the unspoken information corresponding to the user utterance input by the input unit 40 based on a set of predicate-term structures co-occurring in the corpus document. Is generated and stored in the corpus document co-occurrence data storage unit 22.

  Specifically, first, a predicate term structure analysis is performed on each document in order to extract an event from the input document corpus. For the predicate-term structure analysis, it is preferable to use a tool that can extract predicates and terms. As a tool from which predicates and terms can be extracted, for example, JDEP developed by the applicant disclosed in Non-Patent Document 3 below can be mentioned.

[Non-Patent Document 3] Kenji Imamura, Genichiro Kikui, and Norihito Yasuda, "Japanese dependency parsing using scquential labeling for semi-spoken language.", In Proc ACL, 2007.

For example, as described below, a blog document corpus including three documents is received as input.
Document A: "I drove a car and went to Mt. Fuji. I still liked the mountains."
Document B: "I went to Mt. Fuji and climbed the mountain. The scenery was beautiful."
Document C: "I like the mountains, so I went to Mt. Fuji. I drove a car to move."

On the other hand, the result of performing predicate term structure analysis is as follows. Note that "I" (symbol representing the first person) indicates the author of the blog. Predicates are represented by verbs, and terms are represented by nouns. The particles "ga", "$", and "two" represent the type of the term, and represent the subject, the direct object, and the indirect object, respectively.
Document A: "Iga car @ drive", "Iga Fujisan go", "Igayamaga lover", "Iga think"
Document B: "I go to Mt. Fuji", "Climb to Iga", "Beautiful scenery"
Document C: “Iga Yamaga I like”, “Iga Fujisan Go”, “Iga move”, “Iga car ヲ drive”

  Next, all combinations of predicate-argument structures extracted from the same document are extracted as a set, and the number of co-occurrences in which the extracted set co-occurs in the document is counted. Can be

  Finally, based on the number of co-occurrences given to each set of predicate-argument structures, a set of predicate-argument structures whose co-occurrence count is equal to or greater than a predetermined threshold is used as co-occurrence data as a corpus document co-occurrence data storage 22.

  In the above-described example, for example, when the threshold is set to 2, a set of predicate item structures having a co-occurrence count of 2 or more is extracted, and a set of predicate item structures shown in Table 2 below is output as co-occurrence data.

  The corpus document co-occurrence data storage unit 22 stores the corpus document co-occurrence data created by the corpus document co-occurrence data creation unit 20.

  The ranking model creation unit 24 morphologically converts each of the user utterance and the non-verbal information based on a pair of the user utterance and the positive verbal information and a pair of the user utterance and the negative verbal information. Analyze and create a ranking model based on whether the combination of stems among the obtained morphemes and the combination of stems are obtained from a positive set or a negative set, It is stored in the ranking model storage unit 26. The ranking model is a model for calculating a score indicating the likelihood of information other than words in response to a user utterance.

  To create the ranking model, as shown in Table 3 below, a set of a user's utterance and information outside the positive example and a set of a user's utterance and information outside the negative example, which are manually created in advance. Use the data of In the following Table 3, 1 represents a positive example and 0 represents a negative example. “I” represents a speaker of a user utterance to which information other than words is added.

  In order to perform learning of a ranking model based on such data, a combination of a morpheme included in a user utterance and a morpheme included in information other than words is used as a feature.

  Here, as an example, the user's utterance and information other than the words shown in Table 4 below will be described as an example. Here, S is a symbol representing a speaker.

  Using a morphological analyzer such as JTAG developed by the applicant disclosed in Non-Patent Document 4 below, a morphological analysis is performed between the user's utterance and unspoken information.

[Non-Patent Document 4] Takeshi Fuchi, and Shinichiro Takagi, "Japanese morphological analyzer using word co-occurrence: JTAG", Proceedings of the 17th international conference on Computational linguistics-Volume 1, Association for Computational Linguistics, 1998.

  By performing the morphological analysis, the results shown in Table 5 below are obtained. In the following Table 5, the left column indicates the notation, the center column indicates the part of speech, and the right column indicates the stem.

  Next, using the results of the morphological analysis, content words (words having no grammatical meaning and representing meaning) are extracted, and content words extracted from user utterances and content words extracted from non-word information Create a pair with Specifically, from each of the user's utterance and the non-verbal information, morphemes whose part of speech is a noun, a verb stem, or an adjective stem were extracted, and each of the morphemes extracted from the user utterance and the extra verbal information was extracted. By combining each of the morphemes, a set of content words as shown in Table 6 below is created. In Table 6 below, each set of content words is shown as (morpheme extracted from user utterance, morpheme extracted from non-word information).

  A set of content words is created in this manner, and the user utterance is determined based on the created set of content words and whether the set of content words is obtained from a positive example set or a negative example set. Learning a ranking model for calculating a score that represents the likelihood of implicit information with respect to.

  For example, logistic regression may be used as the learning algorithm. Details of logistic regression are disclosed in Non-Patent Document 5 below. As an algorithm other than the logistic regression, for example, a ranking SVM (Support Vector Machine) or the like is cited.

[Non-Patent Document 5] Daiya Takamura, Introduction to Machine Learning for Language Processing, Corona, 2010.

  Finally, the ranking model storage unit 26 stores, as a ranking model, the weight information indicating the likelihood of information other than the user utterance for each set of content words obtained as a result of learning.

  The ranking model storage unit 26 stores the ranking model created by the ranking model creation unit 24.

  The type identification model creation unit 28 creates a type identification model that identifies the type of the information that is implicit, based on data that includes a set of information that is implicitly assigned in advance, and stores the type identification model storage unit 30 To memorize.

  There are nine types shown in Table 7 below.

  These nine types are created from the viewpoint of what information the implicit information represents. The method of making each type and a detailed description are disclosed in Non-Patent Document 6 below.

[Non-Patent Document 6] Wataru Mitsuda, Ryuichiro Higashinaka, Yoshihiro Matsuo, "Classification of unspoken information in dialogue using multiple worker groups", IEICE Technical Report Vol. 116 No. 379 Language Understanding and Communication, pp. 13-18, 2016.

  In the present embodiment, a type identification model for identifying a type with respect to unspoken information is learned using data in which a type is previously assigned to each unspoken information.

  A feature extraction method will be described with reference to the information shown in Table 8 below as an example. Note that the extraordinary information exemplified in Table 8 below is extraordinary information classified into the type of “belief 1”.

  When a morphological analysis is performed on the information that is implicit, the results shown in Table 9 below are obtained. The results in Table 9 below are output results when morphological analysis was performed using a morphological analyzer such as JTAG developed by the applicant. In Table 9 below, the left column represents notation, the left column in the middle represents part of speech, the right column in the middle represents stems, and the right column represents category numbers in the Japanese vocabulary system. The category numbers [] [] [] of the Japanese vocabulary system represent a general noun semantic attribute, a proper noun semantic attribute, and a verbal semantic attribute in order from the left. Each of these attributes is disclosed in Non-Patent Document 7. Here, S is a symbol representing a speaker.

[Non Patent Literature 7] Satoru Ikehara, Masahiro Miyazaki, Satoshi Shirai, Akio Yokoo, Hiromi Nakaiwa, Kentaro Ogura, Yoshifumi Oyama, Yoshihiko Hayashi, Japanese Vocabulary, Iwanami Shoten, 1997.

  The following features are obtained from the results of such morphological analysis. Note that a unigram represents one element of a sequence, and a bigram represents a set of two adjacent elements in a sequence ordered. The morpheme indicates the stem of each morpheme obtained as a result of the morphological analysis.

Morphological unigram: "S", "ha", "Mt. Fuji", "ga", "like", "da"
Morphological bigram: "S-ha", "ha-Mt. Fuji", "Mt. Fuji-ga", "ga-like", "like-da"
General noun semantic attribute unigram: "471", "1300"
General noun semantic attribute bigram: "471-1300"
Word semantic attribute unigram: "11"
Word meaning semantic attribute bigram: none

  When there are a plurality of category numbers of the general noun semantic attribute and the verbal semantic attribute, considering that the leftmost category number is the category number determined to be most appropriate by the JTAG, the leftmost category number is considered. Use only

  Using these features, a model is learned that inputs information that is not the word and outputs that type. As a learning algorithm, an SVM capable of performing multi-class classification may be used. Non-Patent Document 5 discloses the multi-class classification in detail.

  The type identification model storage unit 30 stores the type identification model created by the type identification model creation unit 28.

  The input unit 40 inputs one user utterance and outputs the input user utterance to the example search unit 44a, the co-occurrence search unit 44b, and the non-word information ranking unit 46. For example, user utterances shown in Table 10 below are input as user utterances, and the input user utterances are output to the example search unit 44a, the co-occurrence search unit 44b, and the extraordinary information ranking unit 46.

  Note that, in addition to the user utterance, the input unit 40 may receive, as an input, a conversation context (a series of utterances) up to the user utterance.

  The example data storage unit 42 stores example data, which is a set of sets of user utterances and unspoken information.

  The example search unit 44a generates a candidate for unspoken information from unspoken information corresponding to the user utterance input by the input unit 40, based on the example data stored in the example data storage unit 42.

  Specifically, first, the user utterance input by the input unit 40 and the user utterance in the example data are converted into vectors using word2vec. word2vec is a general method of converting an arbitrary text into a fixed-length vector by performing learning using a text corpus.

  For learning word2vec, a corpus document such as Wikipedia (registered trademark) is used. At this time, the same corpus document as that used in the corpus document co-occurrence data creation unit may be used. A specific learning method is disclosed in Non-Patent Document 8 below.

[Non-Patent Document 8] Tomas Mikolov, Kai Chen, and Jeffrey Dean, "Efficient estimation of word representation in vector space", CoRR, Vol. Abs / 1301.3781, 2013.

The cosine similarity may be used to calculate the similarity between the user utterance converted into a vector using word2vec and input by the input unit 40 and the user utterance in the example data. The cosine similarity is a general measure used to measure the similarity between vectors, and is represented by the following equation (1). Note that, in the following equation (1):

When

Represents a vector,

Represents cosine similarity.

... (1)

  Calculates the similarity between the user utterance input by the input unit 40 and the user utterance in the example data, and the information outside the word that is paired with the user utterance whose cosine similarity is equal to or greater than the threshold in the example data. Is output as an implicit information candidate.

  An example of a process for a user utterance “I went to Mt. Fuji to see autumn leaves” as a user utterance is shown below. Table 11 below shows an example of the example data.

  Table 12 below shows the result of vectorizing the user utterance input by the input unit 40 and the user utterance in the example data and calculating the similarity.

  In Table 12, when the similarity threshold is set to 0.8, as shown in Table 13 below, the unspoken information paired with the user utterance whose similarity is equal to or greater than the threshold is replaced with the unspoken information of the unspoken information. The information is output to the information ranking section 46 as a candidate.

  The co-occurrence search unit 44b is configured to execute a predicate term corresponding to the user's utterance based on a set of co-occurrence predicate term structures stored in the corpus document co-occurrence data storage unit 22 as corpus document co-occurrence data. From the structure, an implicit information candidate is generated.

  When searching for information candidates, the corpus document co-occurrence data stored in the corpus document co-occurrence data storage unit 22 is read, and the read corpus document co-occurrence data is input by the input unit 40. A predicate-argument structure similar to the user utterance is searched. Then, in the set of predicate item structures, the predicate item structure combined with the retrieved predicate item structure is output to the implicit information ranking unit 46 as an implicit information candidate.

  In addition, as a method of searching for a predicate term structure similar to the user utterance input by the input unit 40, a method using the similarity between vectors as in the above-described example search unit 44a is exemplified.

  The word ranking section 46 is generated by the example search section 44a or the co-occurrence search section 44b based on the user utterance input by the input section 40 and the ranking model stored in the ranking model storage section 26. A score representing the likelihood of the extraordinary information for the user utterance is calculated for each of the extraordinary information candidates.

  Table 14 below shows user utterances input by the input unit 40, and Table 15 below shows candidates for unspoken information.

  Table 16 below shows scores calculated in such an example and indicating the likelihood of extraordinary information for the user's utterance.

  As shown in Table 17 below, information that has a calculated score equal to or greater than a threshold (for example, 0.2) is output to the type identification unit 48 in a ranking format in which the scores are arranged in ascending order.

  The type identification unit 48 identifies and assigns a type to each of the word information based on the type identification model stored in the type identification model storage unit 30.

  For example, assume that information other than the words shown in Table 18 below is input.

  In such a case, as shown in Table 19 below, a type is assigned to each implicit information.

  The type filter unit 50 outputs, to the expression conversion unit 52, the unspoken information that satisfies the predetermined condition regarding the type from the unspoken information to which the type is assigned.

  As a method of selecting the unspoken information that satisfies the predetermined condition, for example, a setting file in which the types of the unspoken information to be output are listed in advance is stored, and the unspoken information of the type included in the setting file is stored. There is a method of selecting only.

  For example, assume that only the type of “Fact 1” and the type of “Fact 2” are listed in the setting file. In this case, when information other than the type given in Table 20 below is input, the information outside the word shown in Table 21 below is selected using the setting file.

  In this way, by filtering the information with the type added out of the word based on the type, the information out of the word that may not have to be uttered to the user, for example, “I want to praise S” Can be excluded from the unintended information to be output. In addition, as a method of selecting non-word information that satisfies a predetermined condition, for example, a setting file in which types of non-word information that should not be output may be stored in advance may be stored.

  The expression conversion unit 52 converts each of the unspoken information output by the type filter unit 50 into an utterance sentence, and outputs the utterance sentence to the output unit 54.

  When converting the expression of the implicit information, the end of the implicit information is converted to “-isu” to thereby convert it into a confirmation format. The conversion method is not limited to this, and the end of the implicit information may be converted to “-What?” To form a question. Further, without performing the conversion, the unspoken information may be directly output as the response utterance.

  When information other than words shown in Table 22 below is input, each information of words is converted to an expression shown in Table 23 below. In Table 22 below, only the information that is implicit is shown.

  The output unit 54 outputs, as a response utterance, information of words whose score satisfies a predetermined condition among information of words whose score has been calculated, in a state where a score is assigned. As an output method, a method of displaying data indicating a response utterance on a display unit such as a display, transmitting data indicating a response utterance to an external device, and outputting a sound indicating a response utterance by a sound output unit is provided. No.

  When information other than the words to which a score is assigned as shown in Table 24 below is input, the score is determined to be equal to or more than a predetermined threshold (for example, 0.5) as shown in Table 25 below based on the score. Is output as a response utterance in a state where a score is given.

  The utterance generation device 10 according to the present embodiment is configured by, for example, a computer device including a CPU (Central Processing Unit), a RAM (Random Access Memory), and a ROM (Read Only Memory) for storing various programs. Further, the computer configuring the utterance generation device 10 may include a storage unit such as a hard disk drive and a nonvolatile memory. In the present embodiment, the CPU reads out and executes a program stored in a storage unit such as a ROM and a hard disk, so that the hardware resources and the program cooperate to realize the above-described functions.

  Next, a flow of an utterance generation process performed by the utterance generation device 10 according to the present embodiment will be described with reference to a flowchart illustrated in FIG. In the present embodiment, the utterance generation process is started at a timing at which predetermined data for starting the execution of the utterance generation process is input to the utterance generation device 10. However, the present invention is not limited to this. For example, the utterance generation processing may be started at a timing when data indicating the interaction rule is input.

  In step S101, the input unit 40 inputs a user utterance.

  In step S103, the example search unit 44a extracts the word from the word information corresponding to the user utterance input by the input unit 40, based on the example data which is a set of the user utterance and the word utterance. Generate information candidates.

  In step S105, the co-occurrence search unit 44b generates a predicate term structure corresponding to the user utterance input by the input unit 40 based on a set of predicate term structures co-occurring in the corpus document, which is corpus document co-occurrence data. From the above, an information candidate is generated.

  In step S107, the non-verbal information ranking unit 46 uses the example based on the user utterance input by the input unit 40 and a ranking model for calculating a score indicating the likelihood of the non-verbal information for the user utterance. A score is calculated for each of the unspoken information candidates generated by the search unit 44a or the co-occurrence search unit 44b.

  In step S109, the type identification unit 48 identifies and assigns a type to each of the unspoken information based on a type identification model for identifying the type of the unspoken information.

  In step S111, the type filter unit 50 performs filtering based on the type from the unspoken information to which the type is assigned, and outputs only unspoken information that satisfies a predetermined condition regarding the type.

  In step S113, the expression conversion unit 52 converts each of the unspoken information into an utterance sentence by converting the expression.

  In step S115, the output unit 54 outputs, as the response utterance, the information of the unspoken information whose score satisfies the predetermined condition among the candidates of the unspoken information whose score is calculated, and executes the program of the utterance generation processing. To end.

  As described above, in the present embodiment, based on the example data that is a set of the set of the input user utterance and the information that is not the word or the set of the predicate term structure that co-occurs in the corpus document, A candidate for unspoken information is generated from the unspoken information corresponding to the utterance or the predicate item structure corresponding to the user utterance. Further, based on the input user utterance and a ranking model for calculating a score indicating the likelihood of the information outside the word with respect to the user utterance, a score for each of the generated candidates of the information outside the word is calculated. Is calculated. In addition, among the candidates of the unspoken information whose score has been calculated, the unspoken information whose score satisfies a predetermined condition is output as a response utterance.

  Thus, by generating a response utterance using information other than words, the chat dialogue system can respond to various contents that are not limited to the contents of the user's utterance. Since the user can be told that he / she understands the story well, a dialogue system that can be used for a longer time is realized.

  In the present embodiment, the operations of the constituent elements of the functions shown in FIGS. 1 to 3 are constructed as programs and installed and executed on a computer used as the utterance generation device 10. However, the present invention is not limited to this. May be distributed through the Internet.

  Further, the constructed program may be stored in a portable storage medium such as a hard disk or a CD-ROM, and may be installed on a computer or distributed.

Reference Signs List 10 utterance generation device 12 learning unit 14 utterance generation unit 20 corpus document co-occurrence data creation unit 22 corpus document co-occurrence data storage unit 24 ranking model creation unit 26 ranking model storage unit 28 type identification model creation unit 30 type identification model storage unit 40 Input unit 42 Example data storage unit 44 Search unit 44a Example search unit 44b Co-occurrence search unit 46 Implicit information ranking unit 48 Type identification unit 50 Type filter unit 52 Expression conversion unit 54 Output unit

Claims (8)

An input unit for inputting a user utterance,
Based on example data, which is a set of a set of a user utterance and information outside of the word, or a set of predicate item structures co-occurring in the corpus document, the word outside the word corresponding to the user utterance input by the input unit Information, or a predicate-argument structure corresponding to the user utterance, a search unit that generates candidates for the unspoken information,
Based on the user utterance input by the input unit, and a ranking model for calculating a score indicating the likelihood of information outside the word for the user utterance, the information of the word outside the word generated by the search unit For each of the candidates, an implicit information ranking unit that calculates the score,
An utterance generation device including: Based on a type identification model for identifying the type of the word information, a type identification unit that identifies and assigns the type to each of the word information,
The utterance generation device according to claim 1, further comprising: a type filter unit that outputs, from the unspoken information to which the type is assigned, the unspoken information that satisfies a predetermined condition regarding the type. A type identification model creation unit that creates the type identification model based on data that includes a set of information that is not the word to which the type is given in advance,
The utterance generation device according to claim 2, wherein the type identification unit identifies and assigns the type to each of the unspoken information based on the type identification model created by the type identification model creation unit. . The output unit that outputs, as a response utterance, the unspoken information whose score satisfies a predetermined condition, among the candidates of the unspoken information for which the score is calculated, further comprising an output unit. The utterance generation device described in the section. The utterance generation device according to claim 4, further comprising an expression conversion unit configured to convert each of the unspoken information output from the output unit into an utterance sentence. Based on a set of the user utterance and the non-verbal information of the positive example, and a set of the user utterance and the non-verbal information of the negative example, morphologically analyze each of the user utterance and the non-verbal information, A ranking for creating the ranking model based on a combination of stems among the obtained morphemes and a combination of the stems obtained from a positive set or a negative set. It further comprises a model creation unit,
The verbal information ranking unit performs a morphological analysis of each of the user utterance input by the input unit and the verbal information candidate generated by the search unit, and obtains a stem of the obtained morphemes. The utterance generation device according to any one of claims 1 to 5, wherein the score is calculated for the candidate of the unspoken information based on a combination of the candidates and the ranking model created by the ranking model creation unit. . An utterance generation method in an utterance generation device including an input unit, a search unit, and an unmarked information ranking unit,
The input unit inputs a user utterance,
The search unit is based on example data, which is a set of a set of a user utterance and extra-word information, or a set of predicate-term structures co-occurring in a corpus document, based on the user utterance input by the input unit. Generating a candidate for the unspoken information from the corresponding unspoken information, or a predicate-argument structure corresponding to the user utterance,
The non-verbal information ranking unit is based on the user utterance input by the input unit, and a ranking model for calculating a score indicating the likelihood of the non-verbal information with respect to the user utterance. For each of the generated implicit information candidates, calculating the score,
An utterance generation method including:   An utterance generation program for causing a computer to function as each unit of the utterance generation device according to any one of claims 1 to 6.