JP6126965B2 - Utterance generation apparatus, method, and program - Google Patents

Description

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

  There are roughly two types of dialogue systems: task-oriented dialogue systems and non-task-oriented dialogue systems. The former achieves a specific task by interaction with the system, and is used, for example, in a flight reservation system or a weather information retrieval system. Since these systems can assume what is spoken in advance, the system maintains a database of utterances created manually, or applies weather information extracted from the database to manual templates, and the system utters utterances. (Non-Patent Document 1).

  A non-task-oriented dialogue system handles a dialogue with no purpose, and the content of the dialogue is a so-called chat. Since various topics are spoken in the chat, it is impossible to assume the contents to be spoken in advance. Therefore, utterance generation is a very difficult task. In order to deal with a wide range of user input, in recent years, there is a technique in which sentences such as web and twitter are stored in a database, and a system utterance is selected by selecting an item similar to the user utterance (Non-Patent Document) 2).

Ryuichiro Higashinaka, Katsuhito Sudoh, Mikio Nakano, `` Incorpo-rating Discourse Features into Con_dence Scoring of Intention Recognition Results in Spoken Dialogue Systems '', Speech Communication, 2006, Volume 48, Issues 3-4, p.417-436

Shibata, M., Nishiguchi, T., and Tomiura, Y, "Dialog system for open-ended conversation using web documents.", Infomatica, (2009), 33 (3), p.277-284

  However, if the utterance is generated by selecting the utterance from the utterance database constructed from the web or Twitter in the chat conversation, the utterance may not be able to be performed according to the utterance intention of the system. For example, in a situation where the system determines that a user should ask a question about a certain topic, a question cannot be made if there is no question sentence on the topic on the web or Twitter. Since the system needs to ask questions and ask questions according to the situation, the inability to generate utterances according to the situation lowers the quality of the dialogue.

  As a countermeasure, it is conceivable that an utterance corresponding to the utterance intention is generated on the spot using an utterance generation technique in which the system generates an utterance by itself. For example, a sentence component (information such as what happens) is stored as a database, the component corresponding to the current topic is extracted, and then the component is converted into a sentence according to the utterance intention of the system. It is conceivable to speak after conversion. However, the constituent elements corresponding to the current topic are not always found, and the sentence conversion is not always successful.

  The present invention has been made in view of the above circumstances, and an object thereof is to provide an utterance generation apparatus, method, and program capable of stably generating natural utterance candidates for a user's utterance. .

  In order to achieve the above object, an utterance generation device according to the present invention is based on an input unit that accepts a word indicating a conversation topic, a user's utterance, and a word that indicates the conversation topic received by the input unit. The case element corresponding to the word indicating the topic of the dialogue is included from the predicate item structure database storing a plurality of predicate item structures that are combinations of case elements that are case elements corresponding to the predicate and the previous description word. A predicate term structure retrieval unit that retrieves a predescription term structure, a sentence generation unit that generates an utterance sentence based on each of the previous description term structure searched by the previous description term structure search unit, and the sentence Based on a predetermined rule for determining whether or not the sentence is a natural sentence from each of the utterance sentences generated by the generating unit, the utterance sentence that is not a natural sentence is removed and the utterance that has not been removed Sentence A sentence determination unit that outputs the utterance candidates for the user's utterance received by the input unit, and when the utterance candidate is not output by the sentence determination unit, based on the user's utterance, a plurality of When the utterance corresponding to the user's utterance is searched from the utterance database storing the utterance, and when the utterance is searched, the searched utterance is output as the utterance candidate, and the utterance is not searched And an utterance search unit that outputs information indicating that utterance is impossible.

  An utterance generation method according to the present invention is an utterance generation method in an utterance generation apparatus including an input unit, a predicate term structure search unit, a sentence generation unit, a sentence determination unit, and an utterance search unit, wherein the input unit A step of receiving a word indicating a topic and a user's utterance, and a case corresponding to the predicate and the previous description word based on the word indicating the topic of the conversation received by the input unit by the previous description word term structure search unit. Before searching for the case element corresponding to the word indicating the topic of the conversation from the predicate item structure database storing a plurality of predicate item structures that are combinations with the case elements that are the elements of A step of extracting a description term structure, and a step of generating each utterance sentence by the sentence generation unit based on each of the previous description term term structures extracted by the previous description word term structure search unit; The sentence determination unit removes an utterance sentence that is not a natural sentence from each of the utterance sentences generated by the sentence generation unit, based on a predetermined rule for determining whether or not the sentence is a natural sentence. And outputting each of the uttered sentences that have not been removed as utterance candidates for the user's utterance received by the input unit, and the utterance search unit does not output the utterance candidates by the sentence determining unit. The utterance corresponding to the user's utterance is searched from an utterance database storing a plurality of utterances based on the user's utterance, and the searched utterance is retrieved when the utterance is searched. Outputting as an utterance candidate, and outputting information indicating that the utterance is impossible when the utterance is not searched.

  In addition, the input unit further accepts a dialogue act representing the intention of utterance, and the sentence generation unit, according to a predetermined order for the predescription word, the case element, and the case element A pretext of the predescription term structure searched by the structure search unit, the case element, and a plain text in which the case elements are arranged are generated and predetermined for the dialogue act received by the input unit. Based on a conversion rule for converting into a sentence expressing the dialogue action, the generated plain text can be converted into a sentence expressing the dialog action to generate the utterance sentence. .

  In addition, the sentence determination unit calculates a perplexity value indicating difficulty in generating a sentence based on an N-gram language model obtained in advance for each of the uttered sentences generated by the sentence generation unit. And, if the perplexity value is less than or equal to a threshold value, based on the rule that determines that the sentence is a natural sentence, the utterance sentence that is not a natural sentence is removed from each of the utterance sentences, and the utterance that has not been removed Each sentence can be output as the utterance candidate.

  The program according to 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.

  As described above, according to the utterance generation device, method, and program of the present invention, from the predicate term structure database storing a plurality of predicate term structures, the predicate term structure including a case element corresponding to a word indicating a conversation topic And the natural sentence of each of the utterance sentences generated based on each of the predicate term structures is output as an utterance candidate for the user's utterance, and when the utterance candidate is not output, the user's utterance Based on the utterance, the utterance corresponding to the user's utterance is searched from the utterance database storing a plurality of utterances, and when the utterance is searched, the utterance is output as the utterance candidate, so that the user's utterance is stably It is possible to generate a natural utterance candidate for.

It is a block diagram which shows the example of 1 structure of the speech database construction apparatus which concerns on embodiment of this invention. It is explanatory drawing which shows the example of the list of dialogue acts. It is a block diagram which shows one structural example of the predicate term structure database construction apparatus which concerns on embodiment of this invention. It is explanatory drawing which shows an example of the table data of predicate term structure. It is a block diagram which shows one structural example of the speech production | generation apparatus which concerns on embodiment of this invention. It is a figure which shows an example by which the predicate term structure was searched. It is a figure for demonstrating the perplexity value for determining whether a sentence is a natural sentence. It is a figure which shows an example by which the utterance was searched. It is a flowchart which shows the content of the speech database construction process routine which concerns on embodiment of this invention. It is a flowchart which shows the content of the predicate term structure database construction process routine which concerns on embodiment of this invention. It is a flowchart which shows the content of the speech production | generation process routine which concerns on embodiment of this invention.

<Overview>
First, an outline of an embodiment of the present invention will be described.

  In the embodiment of the present invention, two types of databases are used in combination. One is an utterance database constructed from utterances in microblogs such as Twitter (registered trademark), and the other is a predicate term structure database obtained by analyzing large-scale text data. The utterance database is a database in which a dialogue action indicating the intention of the utterance automatically estimated is assigned to a large number of utterances in the microblog, and the utterance and the dialogue action are stored as a pair. The predicate term structure database stores a predicate term structure obtained as a result of performing a predicate term structure analysis on a large amount of text data. The predicate term structure is a structure that represents a combination of a predicate and a case element corresponding to the predicate, which is a component of a sentence. Reference documents (Jun Hirahiro, Masaaki Nagata, “Predicate using structure learning” An example of predicate term structure analysis is described in "Term structure analysis", Proceedings of the 14th Annual Conference of the Association for Natural Language Processing, 2008, p.556-559).

  Since utterances written by microblog users are stored in the utterance database, many natural utterances are stored. However, there are not always many utterances corresponding to dialogue actions that the system wants to speak.

  On the other hand, the predicate term structure database stores sentence components (predicate term structure), and can generate sentences corresponding to various interactive actions. However, automatically generated sentences can be simple, unnatural, or non-sentenced. Since both databases have merits and demerits, by using these two in a complementary manner, it becomes possible to generate natural utterances in accordance with the dialogue action of the system.

  When generating an utterance about a word indicating a conversation topic (hereinafter referred to as a focus) and a dialogue action of a system given by a higher-level module in the dialogue system, first, a predicate term structure is searched from a predicate term structure database, A sentence is generated according to a predetermined template, but if a sentence cannot be generated or the generated sentence does not satisfy a predetermined condition, the utterance database is searched, and if there is a corresponding utterance, the utterance is uttered. Alternatively, when the utterance database is searched and the corresponding utterance is not found, the predicate term structure is retrieved from the predicate term structure database, and if the predicate term structure is found, a sentence is generated according to a predetermined template and uttered.

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

<System configuration of utterance database construction device>
FIG. 1 is a block diagram showing an utterance database construction device 100 according to an embodiment of the present invention. This utterance database construction device 100 is constituted by a computer including a CPU, a RAM, and a ROM storing a program for executing an utterance database construction processing routine to be described later, and is functionally configured as follows. Has been.

  As shown in FIG. 1, utterance database construction apparatus 100 according to the present embodiment includes utterance data input unit 1, utterance database construction unit 2, and utterance database 3.

  The utterance data input unit 1 receives an input of an utterance set including a plurality of utterances acquired from a microblog site. In the present embodiment, about several million utterances are crawled from a microblog site to form an utterance set.

  The utterance database construction unit 2 gives a dialogue action to each of a plurality of utterances included in the utterance set received by the utterance data input unit 1 by the dialogue action estimation unit 22 described later. Then, the utterance database construction unit 2 registers a pair of the utterance and the dialogue action as one record in the utterance database 3 described later. The utterance database 3 may be constructed by following a general procedure for constructing a full-text search engine database, and may be constructed using free software Lucene, Namazu, or the like. In this embodiment, Lucene is used. The utterance database construction unit 2 includes an utterance candidate database 20 and a dialogue action estimation unit 22.

  The utterance candidate database 20 stores the utterance set accepted by the utterance data input unit 1.

  The dialogue action estimation unit 22 estimates the dialogue action of the utterance for each of a plurality of utterances included in the utterance set stored in the utterance candidate database 20, and assigns the estimated dialogue action to the utterance, Store in the utterance database 3. Specifically, the dialogue action estimation unit 22 pairs the utterance and the dialogue action for each utterance, and then stores them in the utterance database 3. There are 33 types of dialogue actions handled in this embodiment. A list of dialogue actions is shown in FIG.

  The dialogue action estimation unit 22 may extract a feature amount based on the words in the utterance, and use an estimator that estimates the dialogue action from the feature amount. The estimator is constructed in advance by a machine learning technique. Just keep it. For example, it can be constructed using a support vector machine, which is a technique generally used in document classification. As the feature amount, for example, a frequency vector of words in the utterance may be used. In this embodiment, about tens of thousands of utterances prepared separately, a dialogue action is manually added, and this data is used as learning data to learn a multi-class classifier that estimates the dialogue action from the utterance by a support vector machine. did.

  Dialog act estimation unit 22, for example, to the utterance "Hello", estimated the dialogue act "greeting", with respect to the utterance "I like ramen", to estimate the dialogue act "self-disclosure evaluation +" .

  The utterance database 3 stores a plurality of pairs of utterances and dialogue actions estimated for the utterances by the dialogue action estimation unit 22. Further, the utterance database 3 holds a transposed index for each pair so that a full-text search is possible.

<System configuration of predicate term structure database construction device>
FIG. 3 is a block diagram showing predicate term structure database construction apparatus 200 according to the embodiment of the present invention. This predicate term structure database construction device 200 is constituted by a computer including a CPU, a RAM, and a ROM storing a program for executing a predicate term structure database construction processing routine to be described later. It is configured as shown.

  As shown in FIG. 3, the predicate term structure database construction device 200 according to the present embodiment includes a text data input unit 4, a predicate term structure database construction unit 5, and a predicate term structure database 6.

  The text data input unit 4 accepts input of a large-scale text set including a plurality of texts. In the present embodiment, text collected from a blog site is used as a large-scale text set. In this embodiment, the text is collected from the blog site. However, the present invention is not limited to this, and any text may be used.

  The predicate term structure database construction unit 5 extracts a predicate term structure from each of a plurality of texts included in the large-scale text set received by the text data input unit 4 and stores it in the predicate term structure database 6. The predicate term structure database construction unit 5 includes a large-scale text database 50, a morpheme analysis unit 52, a dependency analysis unit 54, and a predicate term structure extraction unit 56.

  The large-scale text database 50 stores a large-scale text set received by the text data input unit 4.

  The morpheme analysis unit 52 performs morpheme analysis on each sentence of a plurality of texts included in the large-scale text set stored in the large-scale text database 50.

  The dependency analysis unit 54 performs dependency analysis on each sentence of a plurality of texts included in the large-scale text set based on the morphological analysis result analyzed by the morpheme analysis unit 52, identifies a phrase, and Determine the dependency structure.

  In addition, what is used for free may be used for the morphological analysis by the morphological analysis part 52 and the dependency analysis by the dependency analysis part 54. For example, ChaSen or CaboCha. In this embodiment, JTAG and JDEP developed by the applicant are used for morphological analysis and dependency analysis, respectively.

  The predicate term structure extraction unit 56, for each sentence of the plurality of texts included in the large-scale text set, based on the dependency structure determined by the dependency analysis unit 54, the predicate of the sentence and the case corresponding to the predicate. The case element that is the element is identified and extracted as a predicate term structure. Then, the predicate term structure extraction unit 56 stores the predicate term structure extracted for each sentence of the plurality of texts in the predicate term structure database 6. In the present embodiment, the predicate term structure extraction unit 56, for each predicate in the sentence, as the case element of the predicate, the ga case, wo case, ni case, de case, g case, kara case, and made case Are extracted, and combinations of predicates and one or more case elements are extracted as a group. Then, the predicate term structure extracting unit 56 collects the same extracted predicate term structures together and registers them in the predicate term structure database 6 together with the frequency.

  The predicate term structure database construction unit 5 will be described with a specific example. For example, if “Taro meets Hanako”, “Meet” is a predicate, and from the clauses related to the clause of the predicate, the element of ga rating is “Taro” and the element of Ni rating is “Hanako” I understand. Then, the predicate term structure “predicate: meet ga case: Taro ni case: hanako” is extracted from here. Such a predicate term structure is extracted from all sentences of the text included in the large-scale text set.

  The predicate term structure database 6 stores a plurality of predicate term structures extracted by the predicate term structure extraction unit 56. Since the predicate term structure data stored in the predicate term structure database 6 is table data, the predicate term structure database 6 may be a general RDB (Relational database). Further, as the predicate term structure database 6, a database capable of full text search may be used. In the present embodiment, RDB is used, and PostgreSQL is used as the implementation.

  FIG. 4 shows an example of predicate term structure table data stored in the predicate term structure database 6. “ID” in FIG. 4 is a serial number of the predicate term structure, and the last column is the frequency. For example, the first data is a predicate term structure corresponding to the sentence “Wagtail looks for food”.

<System configuration of utterance generation device>
FIG. 5 is a block diagram showing the utterance generation device 300 according to the embodiment of the present invention. The utterance generation device 300 is composed of a computer including a CPU, a RAM, and a ROM that stores a program for executing an utterance generation processing routine to be described later, and is functionally configured as follows. Yes.

  As shown in FIG. 5, the utterance generation device 300 according to the present embodiment includes an input unit 7, a predicate term structure database 8, a conversion rule database 9, a language model database 10, an utterance database 11, and a calculation unit. 12 and an output unit 13.

  The input unit 7 receives a focus, a user's utterance, and a dialogue act.

  The predicate term structure database 8 stores table data having the same predicate term structure as the predicate term structure table data stored in the predicate term structure database 6 of the upper descriptive term structure database construction apparatus 200.

  The conversion rule database 9 stores in advance conversion rules for converting a plain text to a sentence expressing the interactive action for each interactive action.

  The language model database 10 stores an N-gram language model obtained in advance.

  The utterance database 11 stores a plurality of pairs of utterances and dialogue actions that are the same as a plurality of pairs of utterances and dialogue actions stored in the utterance database 3 of the utterance database construction apparatus 100.

  The computing unit 12 first selects the predicate term structure database 8 and attempts to acquire utterance candidates (one or more utterances). And when an utterance candidate cannot be acquired, it refers to the utterance database 11 and tries to acquire an utterance candidate. If the utterance still cannot be acquired, information indicating that the utterance is impossible is returned to the output unit 13. If the utterance candidate has been acquired, the utterance candidate is returned to the output unit 13. The arithmetic unit 12 includes a control unit 110, a predicate term structure search unit 112, a sentence generation unit 114, a sentence determination unit 116, and an utterance search unit 118.

  In the control unit 110, information for searching for a speech candidate to be used for speech preferentially from the predicate term structure database 8 is recorded as a set value. The control unit 110 selects the predicate term structure database 8 according to the set value. In the present embodiment, a description will be given of a case where an utterance candidate is searched with priority from the predicate term structure database 8.

Based on the focus received by the input unit 7, the predicate term structure search unit 112 searches the predicate term structure database 8 for a predicate term structure including a case element that matches the focus.
At this time, the searched predicate term structures may be sorted in descending order of frequency, and the top J items may be output. Alternatively, the user's utterance may be used as an input to search the predicate term structure database 8 that can be searched in full text. In that case, a predicate term structure including as many user-uttered words as possible may be searched from the predicate term structure database 8. Further, ranking may be performed using a vector space model similar to the utterance search unit 118 described later, and the top J items may be output.

  FIG. 6 shows a predicate term structure in which “mystil” is input as a focus and searched based on the focus. Ranking in descending order of frequency. In FIG. 6, numbers 0 and 1 are given to predicates and case elements. The number is a simple notation of the predicate term structure used in the present embodiment, and the predicates are number 0, ga case, wo case, ni case, de case, to case, kara case, and made case element. This is expressed by numbers 7 to 7. In other words, in this data, “0_sing” indicates that the predicate is “sing”, and “2_theme song” indicates that the case is “theme song”. In addition, the predicate term structure search unit 112 outputs the top J search results for the predicate term structure to the sentence generation unit 114.

  The sentence generation unit 114, for each of the predicate term structures searched by the predicate term structure search unit 112, in accordance with a pre-determined order for the predicate, case element, and case element, A plain text in which elements and case elements are arranged is generated. Further, the sentence generation unit 114 is generated based on the predicate structure for each predicate term structure based on the conversion rule stored in the conversion rule database 9 corresponding to the dialogue act accepted by the input unit 7. The plain text is converted into a sentence expressing the dialogue action, and an utterance sentence is generated.

  For example, first, the sentence generation unit 114 generates a plain text in the order of words “arg case pred”, where the predicate is pred, the case is case, and the case element is arg. If the predicate term structure is “0_singing 2_music”, a plain text “singing music” is generated. In addition, from the characteristics of Japanese, case elements are arranged side by side so as to be close to predicates in the order of wo case, ni case, ga case, de case, to case, made case, and empty case. If there are multiple case elements, the “case pred” portion is repeated. If the predicate term structure is “0_sing 2_song 4_television”, a plain text “sing a song on television” is generated.

  Next, the sentence generation unit 114 uses the rewrite rule based on the word sequence and the part-of-speech sequence at the end of the sentence as the conversion rule corresponding to the dialogue action stored in the conversion rule database 9 for the generated plain text. Convert to a sentence that represents an action.

  For example, if the adjective verb is the last word, such as a plain text “I like mystil”, and the dialogue act is “question evaluation”, by adding “?” Like “do you like mystil?” , It is converted into a sentence representing the dialogue act “question evaluation”. If the last word is a verb and the dialogue act is “Question_Information Request”, as in the plain text “Mystil sings”, add “Is it?” And “Is mystil sing?” Convert to Or, referring to a given dictionary, the last verb is made into a continuous form (“singing”), and “is it?” Is converted to a sentence “Would you like to sing mystil?”. When there are a plurality of conversion candidates for a sentence representing a dialogue action, one of them is selected at random and set as a conversion result. By preparing a conversion rule for each dialogue action, a plain text that matches the conversion rule can be converted into a sentence representing a given dialogue action.

  If a conversion rule matching the generated plain text is not found, the plain text is rejected. If conversion is possible according to the conversion rule, the converted sentence is sent to the sentence determination unit 116 as an uttered sentence. If the input is J utterance, M (M <= J) converted utterances are sent to the sentence determination unit 116.

  Here, since the utterance sentence produced | generated by the sentence production | generation part 114 is a sentence produced | generated automatically, there exists a possibility that it is unnatural than what was uttered by the human. For example, the word order determined by the sentence generation unit 114 may become unnatural depending on the predicate term structure. For example, the sentence “I am interested in ramen” may be more natural than “I am interested in ramen”. In addition, the sentence as a result of conversion by the conversion rule so as to follow the dialogue act may become unnatural. For example, if you convert a plain text “Cherry Blossoms” into a sentence that represents a dialogue act “Proposal”, the sentence ends with “Let ’s do it”. It is a sentence.

  In order to exclude such an unnatural sentence, in this Embodiment, filtering by perplexity using an N-gram language model is performed.

  That is, the sentence determination unit 116 perplexity indicating the difficulty of generating a sentence based on the N-gram language model stored in the language model database 10 for each utterance sentence generated by the sentence generation unit 114. An utterance that is not removed by removing a non-natural sentence from each of the utterance sentences based on a rule that calculates a value and determines that the perplexity value is equal to or less than a threshold value. Each sentence is output to the output unit 13 as an utterance candidate.

  Specifically, an N-gram language model is constructed from large-scale text data. For example, N is 5. The 5-gram language model includes the probability of how much a given sequence of five words appears in text data. Perplexity is a value that indicates how hard a sentence appears when compared to the language model. Perplexity is a general index in language processing that represents the complexity (difficulty of being generated) of a language. When there are a plurality of utterance sentences whose perplexity value is less than the threshold, the plurality of utterance sentences are sent to the output unit 13 as utterance candidates.

  FIG. 7 shows the result of calculating perplexity for the utterance sentence generated by the system. The first sentence has a very high perplexity of over 500. If you actually read the sentence, it will be considered unnatural as Japanese. The last sentence has a very low perplexity and is understood to be a natural sentence even in Japanese.

  The control unit 110 selects the utterance database 11 when the utterance candidate is not output by the sentence determination unit 116.

  When the utterance database 11 is selected by the control unit 110, the utterance search unit 118 matches the dialogue act from the utterance database 11 based on the user's utterance and dialogue act accepted by the input unit 7, and Searches the utterance corresponding to the user's utterance, outputs the searched utterance as the utterance candidate when the corresponding utterance is searched, and indicates that the utterance is impossible when the corresponding utterance is not searched Output.

  For example, the utterance search unit 118 searches the utterance database 11 for an utterance in which the dialogue act received by the input unit 7 matches the dialogue act and includes as many words as the user is uttering. Note that the utterance search unit 118 searches the utterance database 11 based on the dialogue action and the focus received by the input unit 7, and searches for utterances that have the same dialogue action and include the focused word. Good.

  In the ranking of the search result in the utterance search unit 118, a vector space model weighted by a TF / IDF of a word that is often used in a full-text search engine may be used, or OKAPI / BM25 that is also frequently used may be used. Good. In this embodiment, Lucene's default ranking based on a vector space model weighted by TF / IDF is used. When the corresponding utterance is searched, the top K cases are output to the output unit 13 as utterance candidates. If the corresponding utterance is not searched, information indicating that the search result is empty is output to the output unit 13.

  For example, FIG. 8 shows search results when “self-disclosure evaluation +” is input as the dialogue action and “ramen” is input as the focus. The score is a search result score based on the vector space model.

  The output unit 13 outputs the utterance candidate output from the sentence determination unit 116 or the utterance search unit 118. The output utterance candidate is used for utterance by an upper module such as a dialogue system. In the upper module, for example, if there are a plurality of utterance candidates, one of them is uttered at random.

<Operation of utterance database construction device>
Next, the operation of the utterance database construction device 100 according to the present embodiment will be described. First, when an utterance set including a plurality of utterances acquired from a microblog site is input to the utterance database construction device 100, the utterance database construction device 100 executes an utterance database construction processing routine shown in FIG.

  First, in step S100, the input utterance set is received by the utterance data input unit 1 and stored in the utterance candidate database 20.

  In step S102, the dialogue action estimation unit 22 sets one utterance among a plurality of utterances included in the utterance set stored in the utterance candidate database 20 in step S100.

  In step S104, the dialogue action estimation unit 22 estimates the dialogue action of the utterance for the utterance set in step S102, and assigns the estimated dialogue action to the utterance.

  In step S106, the dialogue action estimation unit 22 stores a pair of the utterance and the dialogue action given in step S104 in the utterance database 3 for the utterance set in step S102.

  In step S108, it is determined whether or not the processes in steps S102 to S106 have been executed for all utterances stored in the utterance candidate database 20. If there is an utterance for which the processes of steps S102 to S106 are not executed, the process returns to step S102. On the other hand, when the processes in steps S102 to S106 are executed for all utterances stored in the utterance candidate database 20, the utterance database construction process routine is terminated.

<Operation of predicate term structure database construction device>
Next, the operation of the predicate term structure database construction device 200 according to the present embodiment will be described. First, when a large-scale text set including a plurality of texts is input to the predicate term structure database construction device 200, the predicate term structure database construction device 200 executes a predicate term structure database construction processing routine shown in FIG.

  First, in step S <b> 200, the text data input unit 4 receives a large text set and stores it in the large text database 50.

  In step S202, the predicate term structure database construction unit 5 sets one text among a plurality of texts included in the large-scale text set stored in the large-scale text database 50 in step S200.

  In step S204, one sentence is set out of the sentences included in the text set in step S202.

  In step S206, the morpheme analysis unit 52 performs morpheme analysis on the sentence set in step S204.

  In step S208, the dependency analysis unit 54 performs dependency analysis on the sentence set in step S204 based on the morphological analysis result analyzed in step S206, identifies the phrase, and determines the relation between phrases. Determine the receiving structure.

  In step S210, for the sentence set in step S204 by the predicate term structure extraction unit 56, the predicate of the sentence set in step S204 and the predicate are set based on the dependency structure determined in step S208. Case elements that are corresponding case elements are identified and extracted as predicate term structures. In step S210, the extracted predicate term structure is temporarily stored in a memory (not shown).

  In step S212, it is determined whether or not the processing in steps S204 to S210 has been executed for all sentences included in the text set in step S202. If there is a sentence that does not execute the processes in steps S204 to S210, the process returns to step S204. On the other hand, if the processes in steps S204 to S210 have been executed for all sentences included in the text set in step S202, the process proceeds to step S214.

  In step S214, it is determined whether or not the processes in steps S202 to S212 have been executed for all the texts stored in the large-scale text database 50. If there is a text that has not been subjected to the processes in steps S202 to S212, the process returns to step S202. On the other hand, if the processes in steps S202 to S212 are executed for all texts stored in the large-scale text database 50, the process proceeds to step S216.

  In step S216, the predicate term structure stored in the memory (not shown) in step S210 and the frequency of the predicate term structure are stored in the predicate term structure database 6.

<Operation of utterance generator>
Next, the operation of the utterance generation device 300 according to the present embodiment will be described. First, when a plurality of pairs of utterances and dialogue actions stored in the utterance database 3 of the utterance database construction device 100 are input to the utterance generation device 300, they are stored in the utterance database 11. Next, when the predicate term structure table data stored in the predicate term structure database 6 of the predicate term structure database construction device 200 is input to the utterance generation device 300, it is stored in the predicate term structure database 8. When a user's utterance, focus, and dialogue action are input to the utterance generation device 300 from a higher-level module such as a dialog system, the utterance generation processing routine shown in FIG. 11 is executed by the utterance generation device 300. .

  First, in step S300, the input unit 7 accepts a focus, a user's speech, and a dialogue act.

  In step S <b> 301, the control unit 110 selects the predicate term structure database 8.

  In step S302, the predicate term structure search unit 112 searches the predicate term structure database 8 for a predicate term structure including a case element that matches the focus, based on the focus received in step S300.

  In step S304, for each of the predicate term structures retrieved in step S302 by the statement generation unit 114, the predicate of the predicate term structure according to a pre-determined order for the predicate, the case element, and the case element case, A plain text in which case elements and case elements are arranged is generated.

  In step S306, for each of the predicate term structures by the sentence generation unit 114, the predicate structure in step S304 based on the conversion rule stored in the conversion rule database 9 corresponding to the dialogue act accepted in step S300. Is converted into a sentence expressing the dialogue action to generate an utterance sentence.

  In step S308, the sentence determination unit 116 calculates a perplexity value based on the N-gram language model stored in the language model database 10 for each of the uttered sentences generated in step S306. Based on the rule for determining that the sentence is a natural sentence if the plexity value is less than or equal to the threshold value, the unnatural sentence is removed from each of the utterance sentences generated in step S306, and the remaining natural utterance sentence is determined. It is determined whether or not there is. If at least one natural utterance remains from each of the utterances generated in step S306, the process proceeds to step S314. On the other hand, if all of the utterance sentences generated in step S306 are removed as unnatural sentences, the process proceeds to step S309.

  In step S <b> 309, the utterance database 11 is selected by the control unit 110.

  In step S310, the utterance search unit 118 searches the utterance database 11 for an utterance corresponding to the user's utterance that matches the utterance based on the user's utterance and the dialog act accepted in step S300. .

  In step S312, it is determined whether or not the dialog act matches in step S310 and an utterance corresponding to the user's utterance has been searched. If the corresponding utterance is searched, the process proceeds to step S314. On the other hand, if the corresponding utterance has not been searched, the process proceeds to step S316.

  In step S314, the output unit 13 outputs each utterance sentence determined to be a natural sentence in step S308 or each corresponding utterance searched in step S312 as an utterance candidate, The generation processing routine is terminated.

  In step S316, the output unit 13 outputs information indicating that the utterance is impossible, and the utterance generation processing routine is terminated.

  Note that the utterance candidate output in step S314 is used for utterance by a higher-level module such as a dialogue system. In the upper module, for example, if there are a plurality of utterance candidates, one of them is uttered at random.

  As described above, according to the utterance generation device according to the present embodiment, a predicate term structure including a case element matching the focus is retrieved from a predicate term structure database storing a plurality of predicate term structures. A natural sentence of each of the utterance sentences generated based on each of the utterances is output as an utterance candidate for the utterance of the user, and when the utterance candidate is not output, a plurality of utterances are generated based on the utterance of the user Utterance database that stores a pair of dialogue actions, and searches for utterances that match the dialogue actions and include the word being spoken by the user, and outputs the utterance as a utterance candidate when the utterance is searched Thus, it is possible to stably generate natural utterance candidates for user utterances.

  In addition, the dialogue system, which is a higher-level module, can perform natural utterance in accordance with the dialogue action (utterance intention) of the dialogue system itself, and the accuracy of information transmission of the system is improved.

  In addition, communication with the user is facilitated, and the interaction between the system and the user becomes smooth.

  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 utterance generation device 300 according to the present embodiment, the predicate term structure database is preferentially searched, and the utterance database is searched when no utterance candidate is generated based on the search result of the predicate term structure database. However, the present invention is not limited to this, and the utterance database may be preferentially searched, and if no utterance candidate is searched in the utterance database, the predicate term structure database may be searched. .

  Moreover, although the above-mentioned speech database construction apparatus 100 demonstrated the case where the speech candidate database 20 and the speech database 3 were provided, the speech candidate database 20 and the speech database 3 are provided in the external apparatus of the speech database construction apparatus 100, for example. The utterance database construction device 100 may refer to the utterance candidate database 20 and the utterance database 3 by communicating with an external device using communication means.

  Further, the above-described predicate term structure database construction device 200 also communicates with the large-scale text database 50 and the predicate term structure database 6 provided in the external device using the communication means, so that the large-scale text database 50 and the predicate The term structure database 6 may be referred to.

  In addition, the utterance generation apparatus 300 described above also communicates with the predicate term structure database 8, the conversion rule database 9, the language model database 10, and the utterance database 11, which are provided in the external device, by using communication means. You may make it refer to the term structure database 8, the conversion rule database 9, the language model database 10, and the speech database 11. FIG.

  In the above embodiment, the case where the utterance database construction device 100, the predicate term structure database construction device 200, and the utterance generation device 300 are configured as separate devices has been described as an example. At least two of the structure database construction device 200 and the utterance generation device 300 may be configured as one device.

  The utterance database construction device 100, the predicate term structure database construction device 200, and the utterance generation device 300 described above have a computer system therein, but the “computer system” may be a case where a WWW system is used. For example, the homepage provision environment (or display environment) is also included.

  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.

DESCRIPTION OF SYMBOLS 1 Speech data input part 2 Speech database construction part 3 Speech database 4 Text data input part 5 Predicate term structure database construction part 6 Predicate term structure database 7 Input part 8 Predicate term structure database 9 Conversion rule database 10 Language model database 11 Speech database 12 Calculation unit 13 Output unit 20 Utterance candidate database 22 Dialogue action estimation unit 50 Large scale text database 52 Morphological analysis unit 54 Dependency analysis unit 56 Predicate term structure extraction unit 100 Utterance database construction device 110 Control unit 112 Predicate term structure search unit 114 sentence Generation unit 116 Sentence determination unit 118 Utterance search unit 200 Predicate term structure database construction device 300 Utterance generation device

Claims (4)

An input unit for receiving a word indicating a conversation topic, a user's utterance, and a dialogue act indicating an intention of the utterance ;
Based on a word indicating the topic of the dialogue received by the input unit, from a predicate term structure database that stores a plurality of predicate term structures that are combinations of a predicate and a case element that is a case element corresponding to the previous descriptive word, A predicate term structure search unit for searching a predescription term structure including the case element corresponding to a word indicating the topic of the dialogue;
Based on each of the previous description term structure searched by the previous description term structure search unit, the previous description term structure search is performed according to a predetermined order for the previous description word, the case element, and the case of the case element. A predicate of the predescription term structure searched by the section, the case element, and a plain text in which the cases of the case element are arranged, and predetermined for the dialogue act received by the input section, A sentence generation unit that converts the generated plain text into a sentence that expresses the dialogue action based on a conversion rule for converting the sentence into a sentence that expresses the dialogue action, and generates each spoken sentence;
From each of the utterance sentences generated by the sentence generation unit, utterance sentences that are not natural sentences are removed based on a predetermined rule for determining whether or not the sentence is a natural sentence, and are not removed. A sentence determination unit that outputs each of the utterance sentences as an utterance candidate for the utterance of the user received by the input unit;
When the utterance candidate is not output by the sentence determination unit, based on the user's utterance, from the utterance database storing a plurality of utterances for each of the dialogue actions , the user's utterance and the dialogue action The corresponding utterance is searched, and when the utterance is searched, the searched utterance is output as the utterance candidate, and when the utterance is not searched, information indicating that the utterance cannot be output is output. A search section;
An utterance generating device including The sentence determination unit calculates a perplexity value indicating difficulty in generating a sentence based on an N-gram language model obtained in advance for each of the uttered sentences generated by the sentence generation unit, Based on the rule that determines that the perplexity value is equal to or less than a threshold value, the utterance sentence that is not a natural sentence is removed from each of the utterance sentences and the utterance sentence that has not been removed is determined. The utterance generation device according to claim 1, wherein each is output as the utterance candidate. An utterance generation method in an utterance generation apparatus including an input unit, a predicate term structure search unit, a sentence generation unit, a sentence determination unit, and an utterance search unit,
Accepting, by the input unit, a word indicating a conversation topic, a user's utterance, and a dialogue action representing the intention of the utterance ;
A predicate term structure that is a combination of a predicate and a case element that is a case element corresponding to the previous descriptive word based on a word indicating the topic of the dialogue received by the input unit by a predescription term structure search unit. Retrieving the case element corresponding to the word indicating the topic of the conversation from a plurality of stored predicate term structure databases, and extracting the pre-description term structure corresponding to the searched case element;
Based on each of the previous description term structure extracted by the previous description term structure search unit by the sentence generation unit, according to a predetermined order for the previous description word, the case element, and the case element, Generates a plain text in which the predicate of the previous description term structure searched by the previous description term structure search unit, the case element, and the case element are arranged, and the dialogue action received by the input unit Converting the generated plain text into a sentence expressing the dialogue action based on a predetermined conversion rule for converting into the sentence expressing the dialogue action, and generating each spoken sentence; ,
The sentence determination unit removes an utterance sentence that is not a natural sentence from each of the utterance sentences generated by the sentence generation unit, based on a predetermined rule for determining whether or not the sentence is a natural sentence. And outputting each of the utterance sentences that have not been removed as utterance candidates for the utterance of the user accepted by the input unit;
When the utterance search unit does not output the utterance candidate by the utterance search unit, based on the utterance of the user, from the utterance database storing a plurality of utterances for each of the dialogue actions , the user's utterance Searching the utterance corresponding to the utterance and the dialogue action , outputting the searched utterance as the utterance candidate when the utterance is searched, and being unable to speak when the utterance is not searched Outputting information indicating:
Utterance generation method including The program for functioning a computer as each part of the speech production | generation apparatus of Claim 1 or Claim 2 .