JP2018190170A - Speech generation apparatus, speech generation method, and speech generation program - Google Patents

Abstract

An utterance generation device, an utterance generation method, and an utterance generation program capable of obtaining extra information that is not directly included in a user utterance in a chat dialogue system.
Corresponding to a user utterance input by an input unit based on example data that is a set of a set of user utterances and extra information, or a set of predicate term structures that co-occur in a corpus document. A score indicating the likelihood of out-of-word information generated from the out-of-word information or the predicate term structure corresponding to the user utterance, and the likelihood of the out-of-word information with respect to the user utterance input by the input unit 40 A score is calculated for each of the generated non-verbal information candidates based on the ranking model for calculating.
[Selection] Figure 3

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 dialogue systems, which are divided into task-oriented dialogue systems and non-task-oriented dialogue systems (chat dialogue dialogue systems). In recent years, attention has been paid to entertainment, daily conversation with robots, etc., and research on chat dialogue systems has been actively conducted.

  In order to advance the conversation in the conversation, the response utterance that conveys the understanding of the system to the user is important. For example, in the case of a task-oriented dialogue system for making a hotel reservation, the user can confirm the understanding of the system by telling the user that he / she wants to stay overnight because he / she wants to stay overnight. If it is a chat dialogue system, it can show that the user's story is understood by repeating the content which the user spoke.

  In a chat dialogue system, a response utterance is often generated by extracting a keyword from a user utterance and applying it to a template. For example, by extracting the keyword “Mt. Fuji” from the utterance “I went to Mt. Fuji” and applying it to the pattern “I am”, a response utterance “I am Mt. Fuji” can be generated.

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

  In recent years, as disclosed in Non-Patent Document 2, a semantic structure called a predicate term structure (a structure consisting of a predicate and its terms) is extracted from a user utterance, and a response utterance is based on the extracted predicate term structure. The method of generating is also studied. For example, by extracting the predicate term structure of “User Ga Fujisan ni go” from the user utterance “I went to Mt. Fuji” and changing the expression at the end, a response utterance “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 utterances that can be generated by the existing chat dialogue system are generated from the words, predicate term structures, etc. that are being uttered by the user, only those that appear superficially in the user utterance can be generated. For this reason, the user is given the impression that the user utterance is simply generated repeatedly, and cannot be shown as listening deeper.

  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 extra information that is not directly included in a user utterance in a chat dialogue system.

  In order to achieve the above object, the utterance generation device of the present invention co-occurs in an example data or a corpus document that is a set of a combination of an input unit for inputting a user utterance and user utterance and extra information. Retrieval that generates candidates for out-of-word information from out-of-word information corresponding to the user utterance input by the input unit or predicate term structure corresponding to the user utterance based on a set of predicate term structures Unit, the user utterance input by the input unit, and the ranking model for calculating a score representing the likelihood of the information outside the speech with respect to the user utterance. A non-verbal information ranking unit that calculates the score for each of the information candidates.

  In addition, based on a type identification model for identifying the type of out-of-word information, a type identifying unit that identifies and gives the type to each of out-of-word information, and the type is given You may make it further include the type filter part which outputs the said extraordinary information which satisfy | fills the predetermined condition regarding the said type from the said extraordinary information.

  In addition, a type identification model creating unit that creates the type identification model based on data that includes a set of information other than the above-described information to which the type has been assigned in advance is further provided, and the type identification unit includes the type identification model Based on the type identification model created by the creation unit, the type may be identified and assigned to each of the extra information.

  Moreover, you may make it further include the output part which outputs the said extraordinary information that the said conditions satisfy | fill the conditions predetermined among the candidates of the extraordinary information from which the said score was calculated as a response utterance.

  Moreover, you may make it further include the expression conversion part which converts each of the said non-verbal information output by the said output part into an utterance sentence.

  Further, 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, each of the user utterance and the non-verbal information is morphologically analyzed. The ranking model is created based on whether the combination of stems among the obtained morphemes and the combination of stems is obtained from a positive example set or a negative example set. A non-verbal information ranking unit that performs a morphological analysis on each of the user utterances input by the input unit and the non-verbal information candidates generated by the search unit. Based on the combination of stems among the obtained morphemes and the ranking model created by the ranking model creation unit, the score is calculated for the candidate information outside the word. Unishi and may be.

  In order to achieve the above object, an utterance generation method of the present invention is an utterance generation method in an utterance generation apparatus including an input unit, a search unit, and an extra information ranking unit, wherein the input unit performs user utterance. The input step and the search unit are input by the input unit based on example data that is a set of user utterances and extra information or a set of predicate term structures that co-occur in a corpus document. Generating a candidate for the verbal information from the verbal information corresponding to the user utterance or the predicate term structure corresponding to the user utterance; and the verbal information ranking unit by the input unit Based on the input user utterance and a ranking model for calculating a score representing the likelihood of extra information regarding the user utterance, the extra information generated by the search unit For each candidate, including a step of calculating the score.

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

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

It is a block diagram which shows the whole structure of the speech production | generation apparatus which concerns on embodiment. It is a block diagram which shows the functional structure of the learning part which concerns on embodiment. It is a block diagram which shows the functional structure of the speech production | generation part which concerns on embodiment. It is a flowchart which shows the flow of the speech production | generation process which concerns on 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 a user utterance by generating a response utterance based on information that can be estimated from the user utterance. In this embodiment, information that can be estimated from user utterances is called “outside speech information”, and even if the “outside speech information” is not explicitly expressed in the utterance, it is generally understood by humans. Is defined as “information that can be used”.

  Specifically, the utterance generation device according to the present embodiment receives user utterances as input, estimates extra information, and generates response utterances using the extra information. In addition, the utterance generation device according to the present embodiment generates a response utterance by generating an utterance for confirming or asking a question in particular. In the estimation of out-of-word information, candidates for out-of-word information are listed in the following (Method 1) and (Method 2), and a score representing the likelihood of a candidate pair of user utterance and each out-of-word information Is calculated, and the information of words outside the speech is ranked to estimate the information outside the speech.

(Method 1) Candidates for out-of-language information by preparing data containing a set of user utterances and out-of-word information and searching for user utterances in the data using the input user utterance Is enumerated.

(Method 2) Data of a set of predicate term structure sets co-occurring in a document (e.g., <raining, wet laundry>) is prepared, a predicate term structure similar to a user utterance is searched, and data Among them, the predicate term structure combined with the searched predicate term structure is regarded as out-of-verb information, thereby enumerating candidates for out-of-verb information.

  By the methods (method 1) and (method 2), candidates for information outside the speech with respect to the user utterance are listed, and by ranking the candidates listed, the information outside the speech is estimated. Specifically, a set of user utterances and correct out-of-word information (positive out-of-word information) and a set of user utterances and out-of-word information (in negative out-of-word information) were recorded. Based on the data, a regression model is learned that estimates the likelihood of information other than words for user utterances. Using the learned regression model, ranking is performed after estimating the likelihood of the candidate for information outside the speech as a score, and the information outside the speech with a score equal to or greater than a predetermined threshold Estimate as information.

  Since various types of information can be included in the out-of-language information estimated as out-of-language information for the user utterance, if all of the estimated out-of-language information is output as response utterances, the user May be uncomfortable. For example, in response to a user utterance that says “I got the first test”, when I guessed the extra information that “the user wants to be praised” May be uncomfortable. For this reason, after estimating the estimated information type outside the speech based on the predefined information type outside the speech, only the specific type speech outside the information is output as a response utterance.

  In contrast to the estimated out-of-word information, the out-of-word information is used as it is (for example, a plain text), the confirmation form at the end (for example, "What is it?"), The form of the question (for example, "~" What is it? ”) Etc.), etc. are used as utterances to output response utterances.

  FIG. 1 is a block diagram showing an overall configuration of an utterance generation device 10 according to the present embodiment. As shown 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 with respect to the user utterance with the likelihood score assigned thereto based on the data and model created by the learning unit 12.

  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 apparatus 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 extraordinary 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.

  In addition, the arrow in FIG. 1 and FIG. 2 represents the input-output relationship of each part. A broken line indicates that each part uses the model or data.

  The corpus document co-occurrence data creation unit 20 generates a candidate for the non-verbal information from the non-verbal information corresponding to the user utterance input by the input unit 40 based on the set of predicate term structures that co-occur in the corpus document. Is generated and stored in the corpus document co-occurrence data storage unit 22.

  Specifically, first, predicate term structure analysis is performed on each document in order to extract events from the input document corpus. For the predicate term structure analysis, a tool capable of extracting predicates and terms may be used. As a tool capable of extracting predicates and terms, for example, JDEP developed by the applicant disclosed in Non-Patent Document 3 below can be cited.

[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, a blog document corpus including three documents is accepted as input as follows.
Document A: “I drove a car and went to Mt. Fuji.
Document B: “I went to Mt. Fuji and climbed the mountain. The scenery was beautiful.”
Document C: “I went to Mt. Fuji because I like mountains. I drove to drive.”

On the other hand, the result of predicate term structure analysis is as follows. “I” (symbol representing the first person) represents the author of the blog. Predicates are expressed as verbs, and terms are expressed as nouns. The particles “ga”, “wo”, and “two” represent the types of terms, and represent the subject, direct object, and indirect object, respectively.
Document A: “I drive my car”, “I go to Mt. Fuji”, “I love Iga”, “I think”
Document B: “I Go to Mt. Fuji”, “Climb I” and “Beautiful scenery”
Document C: “I love Gayamaga”, “I go to Iga Fujisan”, “I move”, “I drive Iga”

  Next, all combinations of predicate term structures extracted from the same document are extracted as a pair, and the number of co-occurrences in which the extracted pair co-occurs in the document is obtained, thereby obtaining the result shown in Table 1 below. It is done.

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

  In the above-described example, for example, when the threshold is set to 2, a set of predicate term structures having a co-occurrence count of 2 or more is extracted, and a set of predicate term 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 converts each of the user utterance and the non-verbal information into a morpheme based on the set of the user utterance and the non-verbal information of the positive example, and the combination of the user utterance and the non-verbal information of the negative example. Analyzing and creating a ranking model based on the combination of stems among the obtained morphemes, and whether the combination of stems is obtained from a positive example set or a negative example set, It is stored in the ranking model storage unit 26. The ranking model is a model for calculating a score representing the likelihood of information other than words for user utterances.

  As shown in Table 3 below, the ranking model is created by manually creating a set of user utterances and positive non-verbal information, and a set of user utterances and negative non-verbal information. The data is used. Note that, in the flags in Table 3 below, 1 represents a positive example and 0 represents a negative example. “I” represents a speaker of a user utterance given extra information.

  In order to perform the learning of the ranking model based on such data, a set of morphemes included in user utterances and morphemes included in extraordinary information is used as features.

  Here, as an example, user utterances and extra information shown in Table 4 below will be described as examples. 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, morphological analysis of user utterances and extra information is performed.

[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 morphological analysis, the results shown in Table 5 below are obtained. In Table 5 below, the left column represents the notation, the middle column represents the part of speech, and the right column represents the stem.

  Next, using the results of morphological analysis, we extract content words (words that have no grammatical meaning and represent meanings), extract content words from user utterances, and content words extracted from non-verbal information Create a pair with Specifically, the morpheme whose part of speech is a noun, verb stem, or adjective stem is extracted from each of the user utterance and verbal information, and extracted from each morpheme extracted from the user utterance and verbal information. By combining each morpheme, 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 extra information).

  A set of content words is created in this way, and based on whether the created set of content words and the set of content words are obtained from a positive example set or a negative example set, the user utterance A ranking model for calculating a score representing the likelihood of information other than the word is learned.

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

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

  Finally, the weight information representing the likelihood of out-of-word information with respect to the user utterance for each set of content words obtained as a result of learning is stored in the ranking model storage unit 26 as a ranking model.

  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 for identifying the type of extraordinary information based on data that includes a set of extraordinary information to which a type has been assigned in advance, and a type identification model storage unit 30. Remember me.

  There are nine types of types shown in Table 7 below.

  These nine types are created from the viewpoint of what kind of information the extra information represents. Each type of creation method and detailed description are disclosed in Non-Patent Document 6 below.

[Non-Patent Document 6] Mitsuda Mitsuo, Higashinaka Ryuichiro, Matsuo Yoshihiro, “Classification of Words 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 for out-of-word information is learned using data in which the type is previously assigned to the out-of-word information.

  The feature extraction method will be described by taking information outside the table shown in Table 8 as an example. Note that the non-verbal information exemplified in Table 8 below is non-verbal information classified into the type “belief 1”.

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

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

  The following features are obtained from the result of such morphological analysis. A unigram represents one element of a sequence, and a bigram represents a group of two adjacent elements in a sequence in order. The morpheme represents the stem of each morpheme obtained as a result of the morpheme analysis.

Morphological Unigrams: “S”, “Ha”, “Mount Fuji”, “Ga”, “Like”, “Da”
Morphological bigram: “S-ha”, “Ha-Fuji”, “Mt. Fuji-ga”, “Ga-I like”, “I like-”
General noun meaning attribute unigram: “471”, “1300”
General noun meaning attribute bigram: "471-1300"
Prescriptive meaning attribute unigram: "11"
Prescriptive meaning attribute bigram: None

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

  Using these features, learn the model that takes the information outside the word as input and outputs the type. As the learning algorithm, an SVM capable of multi-class classification may be used. The multi-class classification is disclosed in detail in Non-Patent Document 5 above.

  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 extra information ranking unit 46. For example, the 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 extra information ranking unit 46.

  In addition to the user utterance, the input unit 40 may accept as an input a conversation context (a series of utterances) before the user utterance.

  The example data storage unit 42 stores example data that is a set of user utterances and extra information.

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

  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 technique for converting an arbitrary text into a fixed-length vector by performing learning using a text corpus.

  A corpus document such as Wikipedia (registered trademark) is used for learning word2vec. 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 input by the input unit 40 converted into a vector using word2vec and the user utterance in the example data. The cosine similarity is a general scale used for measuring the similarity between vectors, and is expressed by the following equation (1). In the following formula (1)

When

Represents a vector,

Represents the cosine similarity.

... (1)

  The similarity between the user utterance input by the input unit 40 and the user utterance in the example data is calculated, and the extra information 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 a candidate for extra information.

  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 example data.

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

  In Table 12 above, if the similarity threshold is 0.8, as shown in Table 13 below, the verbal information paired with the user utterance whose similarity is equal to or greater than the threshold will be As a candidate, the information is output to the information ranking unit 46.

  The co-occurrence search unit 44b predicate terms corresponding to user utterances based on a set of predicate term structures co-occurring in the corpus document stored as corpus document co-occurrence data in the corpus document co-occurrence data storage unit 22. From the structure, a candidate for extra information is generated.

  When searching for candidate information outside the word, the corpus document co-occurrence data stored in the corpus document co-occurrence data storage unit 22 is read out, and the read corpus document co-occurrence data is input by the input unit 40. A predicate term structure similar to the user utterance is retrieved. Then, in the predicate term structure set, the predicate term structure combined with the searched predicate term structure is output to the out-of-word information ranking unit 46 as a candidate for out-of-word information.

  As a method for 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 example search unit 44a described above can be used.

  The extraordinary information ranking unit 46 is generated by the example search unit 44a or the co-occurrence search unit 44b based on the user utterance input by the input unit 40 and the ranking model stored in the ranking model storage unit 26. For each candidate for non-verbal information, a score representing the likelihood of the non-verbal information for the user utterance is calculated.

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

  Table 16 below shows a score representing the likelihood of out-of-word information for the user utterance calculated in such an example.

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

  Based on the type identification model stored in the type identification model storage unit 30, the type identification unit 48 identifies and assigns the type to each piece of extra information.

  For example, it is assumed 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 extra information.

  The type filter unit 50 outputs, to the expression conversion unit 52, extra information that satisfies a predetermined condition related to the type from the extra information provided with the type.

  As a method for selecting out-of-language information that satisfies a predetermined condition, for example, a configuration file in which types of out-of-language information to be output are stored in advance is stored, and the out-of-language information included in this configuration file is stored. The method of selecting only is mentioned.

  For example, assume that only the type of “Fact 1” and the type of “Fact 2” are listed in the configuration file. In this case, when the information other than the word given the type shown 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 out information that is given a type based on the type, information that is not likely to be spoken to the user, such as “S want to give up”, for example. Can be excluded from non-verbal information to be output. In addition, as a method for selecting out-of-language information that satisfies a predetermined condition, for example, a setting file in which types of out-of-word information that should not be output may be stored in advance.

  The expression conversion unit 52 converts each non-verbal 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 non-verbal information, the end of the non-verbal information is converted to a confirmation format by converting it to "~". The conversion method is not limited to this, and the question may be in the form of a question by converting the end of the extra information as “? Further, information other than speech may be output as a response utterance without conversion.

  When extra information as shown in Table 22 below is input, each extra information is converted into an expression as shown in Table 23 below. In Table 22 below, only extra information is shown.

  The output unit 54 outputs out-of-word information satisfying a predetermined score among the out-of-word information for which the score is calculated, as a response utterance in a state where the score is given. As an output method, there is 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, or outputting a voice indicating a response utterance by a voice output unit. Can be mentioned.

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

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

  Next, the flow of utterance generation processing by the utterance generation apparatus 10 according to the present embodiment will be described with reference to the flowchart shown in FIG. In the present embodiment, the utterance generation process is started at a timing when predetermined data for starting execution of the utterance generation process is input to the utterance generation apparatus 10. For example, the utterance generation process may be started at a timing when data indicating a dialogue rule is input.

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

  In step S103, the example search unit 44a uses the example data, which is a set of user utterances and extra information, from the extra information corresponding to the user utterances input by the input unit 40. Generate information candidates.

  In step S105, the co-occurrence search unit 44b has a predicate term structure corresponding to the user utterance input by the input unit 40 based on a set of predicate term structures that co-occur in the corpus document, which is corpus document co-occurrence data. From this, a candidate for out-of-word information is generated.

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

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

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

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

  In step S115, the output unit 54 outputs out-of-word information satisfying a predetermined condition as a response utterance out of the out-of-word information candidates for which the score is calculated, and executes the program for the present utterance generation processing. Exit.

  As described above, in the present embodiment, the input user is based on the example data that is a set of the set of the input user utterance and the extra information, or the set of predicate term structures that co-occur in the corpus document. Out-of-word information candidates are generated from out-of-word information corresponding to utterances or predicate term structures corresponding to user utterances. In addition, based on the input user utterance and a ranking model for calculating a score representing the likelihood of the non-verbal information with respect to the user utterance, a score is generated for each of the generated non-verbal information candidates. Is calculated. Further, out-of-word information whose score is calculated, out-of-word information satisfying a predetermined condition is output as a response utterance.

  Thus, by generating a response utterance using out-of-language information, the chat dialogue system can respond to various contents that are not limited to the user's utterance contents. Since it is possible to tell the user that the story is well understood, a dialogue system that can be used for a longer time is realized.

  In the present embodiment, the operation of the components of the functions shown in FIGS. 1 to 3 is constructed as a program and installed and executed on a computer used as the utterance generation device 10. However, the present invention is not limited to this. It may be distributed through.

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

DESCRIPTION OF SYMBOLS 10 Utterance production | generation apparatus 12 Learning part 14 Utterance production | generation part 20 Corpus document co-occurrence data creation part 22 Corpus document co-occurrence data storage part 24 Ranking model creation part 26 Ranking model memory | storage part 28 Type identification model creation part 30 Type identification model memory | storage part 40 Input unit 42 Example data storage unit 44 Search unit 44a Example search unit 44b Co-occurrence search unit 46 Extra 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 that is a set of user utterances and non-verbal information, or a predicate term structure co-occurring in a corpus document, an extra word corresponding to the user utterance input by the input unit A search unit that generates a candidate for the extra information from the information or the predicate term structure corresponding to the user utterance;
Based on the user utterance input by the input unit and a ranking model for calculating a score representing the likelihood of the out-of-language information with respect to the user utterance, the out-of-language information generated by the search unit For each candidate, a non-verbal information ranking unit that calculates the score;
An utterance generating device including Based on a type identification model for identifying the type of the outside information, a type identifying unit that identifies and gives the type to each of the outside information,
The utterance generation device according to claim 1, further comprising: a type filter unit that outputs, from the extra information provided with the type, the extra information satisfying a predetermined condition relating to the type. A type identification model creating unit that creates the type identification model based on data that includes a set of information other than the word to which the type is assigned in advance,
The utterance generation device according to claim 2, wherein the type identification unit identifies and assigns the type to each of the extraordinary information based on the type identification model created by the type identification model creation unit. . The output part which outputs as the response utterance the said extraordinary information which the said score satisfy | fills the conditions predetermined among the candidates of the extraordinary information for which the said score was calculated. The utterance generation device described in the paragraph. The utterance generation device according to claim 4, further comprising an expression conversion unit that converts each of the non-verbal information output by the output unit into an utterance sentence. Based on a set of user utterances and positive examples of non-verbal information and a set of user utterances and negative examples of non-verbal information, each of the user utterances and non-verbal information is morphologically analyzed, Ranking that creates the ranking model based on whether the combination of stems among the obtained morphemes and the combination of stems is obtained from a positive example set or a negative example set A model creation unit,
The non-verbal information ranking unit morphologically analyzes each of the user utterance input by the input unit and the non-verbal information candidates generated by the search unit, and 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 a candidate for the extra information based on a combination of each other 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 extra information ranking unit,
The input unit inputs a user utterance;
Based on the example data that is a set of user utterances and extra information, or a set of predicate term structures that co-occur in a corpus document, the search unit inputs the user utterances input by the input unit. Generating the verbal information candidates from the corresponding verbal information or the predicate term structure corresponding to the user utterance;
Based on the user utterance input by the input unit and a ranking model for calculating a score representing the likelihood of the non-verbal information with respect to the user utterance, Calculating the score for each of the generated non-verbal information candidates;
Utterance generation method including   The utterance production | generation program for functioning a computer as each part of the utterance production | generation apparatus in any one of Claims 1-6.