JP5728527B2 - Utterance candidate generation device, utterance candidate generation method, and utterance candidate generation program - Google Patents

Description

  The present invention relates to an utterance candidate generation device, an utterance candidate generation method, and an utterance candidate generation program, and more particularly to an utterance candidate generation device, an utterance candidate generation method, and an utterance candidate for system utterance in an interactive system that responds to user utterances. It relates to the generation program.

  In general, in a dialogue system that interacts with a user, in order to realize a human-like response, a large number of responses regarding user utterances are manually constructed as utterance pairs and used as responses of the dialogue system 10 in many cases. Examples of the utterance pair include user utterance: “thank you”, response: “you are welcome”, and the like. There is a system that generates a response by such a method (see, for example, Non-Patent Document 1).

  In addition, regarding the current topic in dialogue with the user, the relevant sentence is extracted using a web search engine using the Internet, and one or more sentences are extracted as utterances from among the extracted sentences. There exists a system (for example, refer nonpatent literature 2).

R. S. Wallace, “The Anatomy of A.L.I.C.E.”, A.L.I.C.E.Artificial_Intelligence Foundation, Inc., 2004. Shibata, M., Nishiguchi, T., and Tomiura, Y. (2009). “Dialog system for open-ended conversation using web documents.” Informatica, 33 (3), pp. 277-284.

  Both the technique described in Non-Patent Document 1 and the technique described in Non-Patent Document 2 generate a response with a single algorithm, but user utterances vary, and a dialogue system is appropriate only with a single algorithm. It is difficult to respond to. For example, in the technique of Non-Patent Document 1, an accurate answer can be given to a user utterance assumed by the developer, but a response cannot be made to a user utterance not assumed by the developer. In the technology of Non-Patent Document 2, since an Internet search engine is used, responses can be returned for various topics. However, search results from the Internet including various contents are low in quality to be used as utterances. Also, it is difficult to respond to a standard response such as returning a greeting as a greeting. In addition, even if the response methods according to Non-Patent Document 1 and Non-Patent Document 2 are combined, system utterances that can respond appropriately are limited, and it is not always possible to respond appropriately to user utterances. Absent.

  The present invention has been made in consideration of the above-described problems, and provides an utterance candidate generation device, an utterance candidate generation method, and an utterance candidate generation program capable of appropriately responding to a user utterance in a dialog system. For the purpose.

In order to achieve the above object, an utterance candidate generation device of the present invention is an utterance candidate generation device that generates an utterance candidate used for a response in an interactive system that interacts with a user by responding to a user utterance. Selecting an utterance pair determined to match the input user utterance from a plurality of predetermined utterance pairs consisting of a user utterance and a response to the user utterance, and using the response of the selected utterance pair The first module for generating utterance candidates, the user extracted from the input user utterance from utterances using related words related to the word predetermined for each word representing the topic of user utterance A second module for selecting an utterance using the related word related to a word representing an utterance topic, and generating an utterance candidate using the selected utterance; A third module, which searches for a search query using a word extracted from the input user utterance from an utterance database storing a number of utterances and generates utterance candidates using the searched utterance; In addition, a fourth module that generates an utterance candidate using at least one utterance out of a plurality of utterances that are determined to be utterable regardless of the topic of the user utterance, and changes a predetermined topic of the user utterance comprising a fifth module for generating an utterance candidate with at least one utterance of a plurality of utterances for, when the utterance candidate by the first module is not generated, the process proceeds to the second module, the second When the utterance candidate is not generated by two modules, the process proceeds to the third module, and the utterance candidate is generated by the third module. If not, the utterance candidate generated by the first module, the utterance candidate generated by the second module, the utterance candidate generated by the third module, by proceeding to the fourth module, The utterance candidate is output in preference to the utterance candidate generated by the fourth module and the utterance candidate generated by the fifth module.

  In addition, at least one of the first module to the fifth module of the utterance candidate generation device of the present invention is determined based on dialogue action estimation information obtained by estimating a dialogue action representing the user's utterance intention. Preferably, the utterance candidate is generated based on an interactive action of the interactive system.

An utterance candidate generation method according to the present invention is an utterance candidate generation method for generating an utterance candidate used for a response in an interactive system that interacts with a user by responding to a user utterance. The utterance candidate generation method includes: And a response to the user utterance, a utterance pair that matches the input user utterance is selected from a plurality of predetermined utterance pairs, and an utterance candidate is generated using the response of the selected utterance pair The topic of the user utterance extracted from the input user utterance from the utterance using a related word related to the word, which is predetermined for each word representing the topic of the user utterance by the second module. Selecting an utterance using the related word related to the word representing, and generating an utterance candidate using the selected utterance; Searching by a search query using a word extracted from the input user utterance from an utterance database storing a plurality of utterances, and generating utterance candidates using the searched utterances; A step of generating an utterance candidate using at least one utterance of a plurality of utterances that can be uttered regardless of a topic of the user utterance determined by the module, and a user predetermined by the fifth module A step of generating an utterance candidate using at least one utterance of a plurality of utterances for changing a topic of utterance, and when the utterance candidate is not generated by the first module, the process proceeds to the second module If the utterance candidate is not generated by the second module, the process proceeds to the third module, and the third module If the utterance candidate is not generated by the process, the utterance candidate generated by the first module, the utterance candidate generated by the second module, and the third module are advanced to the fourth module. The utterance candidate is output in the order of the generated utterance candidate, the utterance candidate generated by the fourth module, and the utterance candidate generated by the fifth module .

  The utterance candidate generation program of the present invention is for causing a computer to function as each module of the utterance candidate generation device of the present invention.

  According to the utterance candidate generation device, the utterance candidate generation method, and the utterance candidate generation program of the present invention, it is possible to obtain an effect that an appropriate response can be made according to the user utterance in the dialog system.

It is a block diagram which shows the outline of an example of the dialogue system of this Embodiment. It is explanatory drawing which showed the table showing the relationship between the set of dialogue action used in the dialogue system of this Embodiment, and the example of a corresponding utterance. It is explanatory drawing which shows a specific example of the utterance pair of this Embodiment. It is explanatory drawing which shows a specific example of the word-utterance pair of this Embodiment. It is explanatory drawing which shows a specific example of the utterance list | wrist obtained from an utterance database by using "ramen" as a search query in the utterance search part of this Embodiment. It is explanatory drawing which shows a specific example of the utterance list | wrist obtained from an utterance database by making "5: self-hill maintenance evaluation +" which is a dialogue action of a system and "5 ramen" into the search query in the utterance search part of this Embodiment. . It is explanatory drawing which shows a specific example of the speech list | wrist of this Embodiment. It is explanatory drawing which shows a specific example of the speech list | wrist of this Embodiment. It is a flowchart of an example of the speech candidate production | generation process performed with the speech candidate production | generation apparatus of this Embodiment. It is a flowchart of an example of the speech candidate production | generation process performed with the speech candidate production | generation apparatus of this Embodiment. It is explanatory drawing for demonstrating a specific example of the dialogue with the dialogue system of this embodiment, and a user.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. Note that this embodiment does not limit the present invention.

  The dialogue system 10 according to the present embodiment is a dialogue system in which the system utters in response to a user's utterance and performs a dialogue such as chatting with the user, based on the user's utterance and the dialogue action estimated from the user's utterance. The utterance candidate 46 is generated on the basis of the dialog action of the dialog system 10 determined in this manner, and a response is made to the user by the utterance selected from the generated utterance candidate 46.

  First, the configuration of the interactive system 10 according to the present embodiment will be described. FIG. 1 is a configuration diagram illustrating an example of a schematic configuration of a dialogue system 10 according to the present embodiment. The dialogue system 10 according to the present embodiment includes an input unit 12, a dialogue control unit 14, an output unit 18, and an utterance candidate generation device 20.

  The input unit 12 has a function of outputting a user utterance input by the user to the utterance candidate generation device 20. In addition, the input unit 12 of the present embodiment has a function of outputting the input user utterance also to the dialogue control unit 14. Furthermore, the input unit 12 according to the present embodiment has a function of outputting a conversation history with the user to the utterance candidate generation device 20.

  The dialogue control unit 14 includes a dialogue act estimator 16. The dialogue act estimator 16 estimates the dialogue act of the input user utterance, and determines the utterance intention of the dialogue system according to the estimated dialogue act. It has a function. Information representing the utterance intention on the dialog system 10 side is called a dialog action of the system. In the present embodiment, “dialogue action” refers to a symbol representing the intention of utterance.

  The dialogue act of this embodiment is composed of a set of dialogue act ID and dialogue act name. FIG. 2 shows a table representing a relationship between a set of dialogue actions and a corresponding utterance example used in the present embodiment as a specific example. The dialogue action set is not limited to the one shown in FIG. 2, and a DAMSL tag set or the like generally used in discourse research may be used.

  The dialogue act estimator 16 is an estimator that estimates a dialogue act based on a word being uttered, and is constructed by a machine learning technique. For example, the dialogue action estimator 16 can be constructed using a support vector machine or the like, which is a method generally used in document classification. In the present embodiment, a multi-class classifier that learns about tens of thousands of utterances by hand and uses this data as learning data to estimate the dialogue action from the utterances by using a support vector machine is learned.

  In addition, for the determination of the dialogue action of the system, for example, the literature (Toyomi Meguro, Yasuhiro Minami, Ryuichiro Higashinaka, Koji Dosaka, Evaluation of the hearing dialogue control unit using POMDP by the Wizard of Oz experiment, The 26th Annual Meeting of the Japanese Society for Artificial Intelligence Manual rules such as those described in the National Conference Organized Session “OS-18 Intelligent Dialogue System”, 2012.) may be used. In addition, the interactive action of the system may be determined using a statistical dialog control technique called MDP or POMDP which has become common in recent years.

  The dialogue action of the system determined by the dialogue control unit 14 is output to the utterance candidate generation device 20. In addition, in the generation of utterance candidates in the utterance candidate generation device 20, the system interaction is not essential. When the utterance candidate generation device 20 generates an utterance candidate without using a dialogue action, the dialogue control unit 14 may not include the dialogue action estimator 16 and outputs the estimated dialogue action to the utterance candidate generation device 20. It is not necessary to have the function to do.

  The output unit 18 has a function of selecting one or more utterances from the utterance candidates 46 output by the utterance candidate generation device 20 and responding (presenting) to the user as system utterances.

  The utterance candidate generation device 20 of the present embodiment has a function of generating an utterance candidate 46 for responding to a user utterance input from the input unit 12 using a plurality of modules and outputting the utterance candidate 46 to the output unit 18. The utterance candidate generation device 20 of the present embodiment includes five modules. In the utterance candidate generation device 20, the utterance pair selection unit 22 corresponds to the first module, the related utterance selection unit 26 corresponds to the second module, the utterance search unit 32 corresponds to the third module, and the versatile utterance selection unit 38. Corresponds to the fourth module, and the new topic utterance selection unit 42 corresponds to the fifth module. Hereinafter, the utterance pair selection unit 22, the related utterance selection unit 26, the utterance search unit 32, the versatile utterance selection unit 38, and the new topic utterance selection unit 42 are collectively referred to as “module”.

  In addition, the utterance candidate generation device 20 of the present embodiment includes an utterance pair 24, a word-utterance pair 30, an utterance database 36, an utterance list 40, and an utterance list 44 as databases used in each module.

  Each of these processing units and the above-described processing units (input unit 12, dialogue control unit 14, and output unit 18) in the utterance candidate generation device 20 are a CPU (Central Processing Unit), a RAM (Random Access Memory), and The processing is realized by a computer having a ROM (Read Only Memory) or the like, and the CPU executes a program stored in the ROM, whereby processing in each processing unit described in detail later is executed.

  The utterance pair selection unit 22 has a function of generating utterance candidates by selecting a response to the user utterance itself from the utterance pair 24 given in advance.

  The utterance pair 24 stores one or more pairs (utterance pairs) of a user utterance and a response to the user utterance. The data of the utterance pair 24 (utterance pair) is created in advance by, for example, manual work. FIG. 3 shows a specific example of the utterance pair 24 of the present embodiment. In the present embodiment, the dialogue action of the response of the utterance pair is assigned to the utterance pair and stored in the utterance pair 24. In addition, what is necessary is just to provide the dialog act estimated using the dialog act estimator similar to the above-mentioned dialog act estimator 16 as the response dialog act. In the case of “0: greeting” shown in FIG. 3, “0” is the dialogue action ID, and “greeting” indicates the dialogue action name. In this embodiment, as will be described later, the utterance candidate 46 is generated also using the system interaction action, and thus the conversation action of each utterance pair is also stored in the utterance pair selection unit 22 in this way. However, when the utterance candidate 46 is generated without using the dialogue action, the dialogue action may not be stored in the utterance pair selection unit 22.

  The utterance pair selection unit 22 extracts responses of utterance pairs including user utterances that match the input user utterance from the utterance pair selection unit 22, and generates an utterance list including the extracted responses as utterance candidates 46. The number of responses extracted by the utterance pair selection unit 22 is not particularly limited. Further, in the utterance pair selection unit 22 according to the present embodiment, when a dialogue action of the system is inputted in addition to the user utterance, the inputted dialogue utterance and the dialogue action are matched with the inputted user utterance. Responses in the matched utterance pairs are extracted from the utterance pairs 24, and an utterance list including the extracted responses is generated as the utterance candidates 46. By using the dialogue action, the utterance candidate 46 in accordance with the utterance intention of the dialogue system 10 can be generated.

  Since the utterance candidate 46 generated by the utterance pair selection unit 22 is an utterance corresponding to the user utterance itself, an appropriate response to the user utterance can be expected.

  The related utterance selection unit 26 extracts a word indicating the current topic in the dialogue with the user from the input user utterance, and selects an utterance corresponding to the extracted word from the word-utterance pair 30, thereby utterance candidates. 46 is generated. In the present embodiment, the “input user utterance” is not limited to the utterance input immediately before, but a plurality (for example, a predetermined number) of users input so far Speech may be targeted.

  The word-utterance pair 30 stores at least one word-utterance data including a combination of a word indicating a topic and an utterance using a related word related to the word. The word-utterance data is created in advance by hand, for example. FIG. 4 shows a specific example of the word-utterance pair 30 of the present embodiment. In the present embodiment, the dialogue action of the utterance pair is assigned to each data and stored in the word-utterance pair 30. In addition, what is necessary is just to give the dialog act estimated using the dialog act estimator similar to the above-mentioned dialog act estimator 16 to the dialog act of speech.

  The related utterance selection unit 26 performs morphological analysis on the input user utterance by the morphological analyzer 28 and extracts words indicating topics in the dialogue with the user based on the morphological analysis result. For example, the word indicating the topic may be a noun included in the user utterance. As a specific example, when a user utterance “I like ramen” is input, morphological analysis is performed on the user utterance by the morphological analyzer 34, so that “ramen” can be determined as a noun. In this case, the related utterance selection unit 26 extracts “ramen” as a word indicating a topic. Further, the related utterance selection unit 26 extracts utterances corresponding to the words that match the extracted word indicating the topic from the word-utterance pair 30, and generates an utterance list including the extracted utterances as the utterance candidates 46. In the present embodiment, as a specific example, an utterance list including the utterances shown in FIG. The number of utterances extracted by the related utterance selection unit 26 is not particularly limited. Further, in the related utterance selection unit 26 according to the present embodiment, when a dialogue action of the system is input in addition to the user utterance, it matches the word indicating the extracted topic, and the inputted dialogue action and dialogue action The utterances in the data that coincide with each other are extracted from the word-utterance pair 30, and the utterance list including the extracted utterances is generated as the utterance candidate 46. By using the dialogue action, the utterance candidate 46 in accordance with the utterance intention of the dialogue system 10 can be generated.

  Since the utterance candidate 46 generated by the related utterance selection unit 26 corresponds to the topic in the dialogue with the user, not the user utterance itself, the utterance candidate selection unit 22 (first module) directly deals with the user utterance. Although it is not compatible, it is possible to realize a response that matches the topic of dialogue.

  The utterance search unit 32 has a function of generating utterance candidates 46 by extracting words from the input user utterance and searching the utterance database 36 for utterances including the extracted words.

  As a specific example, the utterance database 36 of the present embodiment crawls Twitter (registered trademark) site data independently, divides these into sentences according to punctuation marks and symbols, and sets a single sentence (utterance). Although what was registered as a record is used, it is not specifically limited. Note that the utterance database 36 holds a transposed index so that a full-text search is possible. The method for constructing the utterance database 36 is not particularly limited. For example, a procedure for constructing a general full-text search engine database may be taken, and construction may be performed using free software Lucene, Namazu, or the like. In the present embodiment, as a specific example, an utterance database 36 constructed using Lucene is used. In the present embodiment, an utterance dialogue act is assigned to the utterance and stored in the utterance database 36. In addition, what is necessary is just to give the dialog act estimated using the dialog act estimator similar to the above-mentioned dialog act estimator 16 to the dialog act of speech.

  The utterance search unit 32 performs morphological analysis on the input user utterance by the morphological analyzer 34, extracts words from the user utterance based on the morphological analysis results, and constructs a search query using the extracted words. Then, the utterance database 36 is searched. As a specific example, when a user utterance “I like ramen” is input, a morphological analysis is performed on the user utterance by the morphological analyzer 34 to extract words included in the utterance sentence. . In the user utterance sentence, “ramen” and “like” can be extracted as content words (nouns, verbs, and adjectives). A search query consisting of these two words may be created, or a search query consisting of either “ramen” or “like” may be created. The utterance database 36 is searched by the search query obtained in this way, and the utterances 46 as the utterance list are obtained as utterance lists by obtaining the utterances as the top N search results. The number of utterances acquired by the utterance search unit 32 is not particularly limited. Further, in the utterance search unit 32 according to the present embodiment, when a system interaction action is input in addition to a user utterance, only the utterances that match the system interaction action are included by including the system interaction action in the search query. Can be obtained as a search result. As a specific example, FIG. 5 shows an utterance list obtained from the utterance database 36 using “ramen” as a search query. FIG. 6 shows an utterance list obtained from the utterance database 36 by using “5: Self-disclosure evaluation +” as a specific example as a search query.

  The utterance candidate 46 generated by the utterance search unit 32 can respond using an utterance directly corresponding to the topic in the dialogue with the user, compared to the utterance candidate 46 generated by the related utterance selection unit 26 (second module). Although it is not limited, a response suitable for the topic can be realized by obtaining an appropriate search result.

  The versatile utterance selection unit 38 has a function of generating an utterance candidate 46 by selecting an utterance from the utterance list 40. In general, “versatile” means various and includes various meanings. For this reason, an utterance that can be interpreted in any sense and is less likely to cause a large problem in context when uttered is referred to as “versatile utterance” in the present embodiment.

  The utterance list 40 is an utterance list composed of a plurality of utterances that do not greatly affect the contents of the dialogue regardless of the current topic in the dialogue with the user. By storing the Versatile utterances in the utterance list 40, the dialogue system 10 can connect the places using these utterances when he / she is deceived. The utterance list 40 stores, for example, utterances such as “Yes” and “Yes”, and utterances that express sympathy such as “It ’s amazing” or “Good” without specific contents. ing. The utterance list 40 is created in advance, for example, manually. FIG. 7 shows a specific example of the utterance list 40 of the present embodiment. In the present embodiment, an utterance dialogue act is assigned to each utterance and stored in the utterance list 40. As the dialog action of the utterance, the dialog action estimated using the dialog action estimator similar to the dialog action estimator 16 described above may be given.

  The versatile utterance selection unit 38 generates an utterance candidate 46 by selecting an utterance from the utterance list 40. The versatile utterance selection unit 38 may set the entire utterance list 40 as the utterance candidate 46, or may appropriately select one or more from the utterance list 40 as the utterance candidate 46. In the present embodiment, an utterance dialogue act is assigned to the utterance and stored in the utterance list 40. In addition, when a system dialogue action is input, the versatile utterance selection unit 38 of the present embodiment selects a utterance to which a dialogue action that matches the system dialogue action is given from the utterance list 40, and is a speech candidate. 46 is generated.

  The utterance candidates 46 generated by the versatile utterance selection unit 38 are user utterances such as the utterance pair selection unit 22 (first module), the related utterance selection unit 26 (second module), and the utterance search unit 32 (third module). In addition, it is possible to realize a response capable of continuing the current topic, although an utterance corresponding to the current topic in the dialog with the user cannot be responded.

  The new topic utterance selection unit 42 has a function of generating an utterance candidate 46 by selecting an utterance from the utterance list 44.

  The utterance list 44 is an utterance list including utterances for guiding the user to a new topic in order to change the current topic in the dialog with the user. The utterance list 44 is created in advance, for example, manually. FIG. 8 shows a specific example of the utterance list 44 of the present embodiment. In the present embodiment, an utterance dialogue act is assigned to each utterance and stored in the utterance list 44. As the dialog action of the utterance, the dialog action estimated using the dialog action estimator similar to the dialog action estimator 16 described above may be given.

  The new topic utterance selection unit 42 generates an utterance candidate 46 by selecting an utterance from the utterance list 44. Note that the new topic utterance selection unit 42 may set the entire utterance list 44 as the utterance candidate 46, or may appropriately select one or more from the utterance list 44 as the utterance candidate 46. In addition, in the new topic utterance selection unit 42 according to the present embodiment, when a system dialogue action is input, the utterance list 44 selects an utterance provided with a dialog action that matches the inputted system dialogue action. The utterance candidate 46 is generated.

  The utterance candidates 46 generated by the new topic utterance selection unit 42 are the utterance pair selection unit 22 (first module), the related utterance selection unit 26 (second module), the utterance search unit 32 (third module), and versatile utterance selection. Unlike the unit 38 (fourth module), by performing a response that guides the user to a new topic, it is possible to overcome the situation where the user cannot speak or respond to the current topic in the dialog with the user. A response can be realized.

  Next, the flow of operation of the utterance candidate generation device 20 of the present embodiment will be described. In the utterance candidate generation device 20 of the present embodiment, by appropriately using a plurality of (first to fifth) modules, an appropriate and high-quality response is always performed according to the user utterance. In addition, the utterance candidate generation device 20 according to the present embodiment sequentially generates utterance candidates 46 in the order of the first module (utterance pair selection unit 22) to the fifth module (new topic utterance selection unit 42). By generating the utterance candidates 46 in such an order, the utterance candidates 46 are sequentially generated in the order of modules capable of providing a more appropriate response to the user utterances. A response can be realized. Note that the order of the modules for generating the utterance candidates 46 is not limited to the order of the present embodiment, but for the reasons described above, the order of proceeding from the module with the smallest number to the module with the larger number as in this embodiment. Is preferred.

  FIG. 9 and FIG. 10 show a flowchart of an example of the flow of the utterance candidate generation process in the utterance candidate generation device 20. The utterance candidate generation process is executed when a user utterance and a dialogue history are input from the input unit 12 to the utterance candidate generation device 20. The input of the dialogue history is not essential, but when each module cannot generate the utterance candidate 46, such as when a word indicating a topic cannot be extracted from one input user utterance, the input user utterance is input. Instead, since the utterance candidate 46 can be generated using the user utterance uttered in the past from the user utterance, it is preferable to input the dialogue history as in the present embodiment.

  First, the utterance pair selection unit 22 generates an utterance candidate 46. Therefore, in step S <b> 100, a user utterance is input to the utterance pair selection unit 22. In the next step S102, the utterance pair selection unit 22 determines whether or not an utterance pair including a user utterance that matches the input user utterance is stored in the utterance pair 24. In the present embodiment, the utterance pair selection unit 22 determines whether or not an utterance pair including a user utterance that is regarded as a match is stored in the utterance pair 24 as well as a complete match. Specifically, for example, using the editing distance or the similarity of the word set, whether or not they match is determined based on whether or not the threshold value is exceeded based on a preset threshold value. As described above, the utterance pair selection unit 22 stores not only the utterance pair including the user utterance that completely matches the input user utterance but also whether the utterance pair including the user utterance considered to match is stored in the utterance pair 24. It is preferable to judge whether or not.

  If an utterance pair including a user utterance that matches the input user utterance is not stored in the utterance pair 24, an utterance candidate 46 cannot be generated, and therefore the following module (related utterance selection unit 26 in the present embodiment) is used. In order to generate the utterance candidate 46, the process proceeds to step S112. On the other hand, if an utterance pair including a user utterance that matches the input user utterance is stored in the utterance pair 24, the process proceeds to step S104.

  In step S104, it is determined whether or not there is a dialogue action input from the dialogue control unit 14. If there is an interactive action, the process proceeds to step S106, and after generating an utterance candidate 46 from an utterance pair that matches the input user utterance and the interactive action, the process proceeds to step S110. Specifically, an utterance pair including the input user utterance and provided with the input dialogue action is extracted from the utterance pair 24, and a response of the extracted utterance pair is generated as the utterance candidate 46. On the other hand, if there is no dialogue act, the process proceeds from step S104 to step S108. In step S108, the utterance candidate 46 is generated from the utterance pair that matches the input user utterance, and the process proceeds to step S110. Specifically, an utterance pair including a user utterance that matches the input user utterance is extracted from the utterance pair 24, and a response of the extracted utterance pair is generated as an utterance candidate 46.

  In step S110, it is determined whether or not the utterance candidate 46 is generated in the next module. In this embodiment, as described above, the utterance candidates 46 are sequentially generated in the order of the first module (the utterance pair selection unit 22) to the fifth module (the new topic utterance selection unit 42). The utterance candidate 46 may not be generated. For example, one or more modules may be skipped, and when the utterance candidate 46 including a predetermined number of responses or utterances is generated, the utterance candidate 46 is not generated by the subsequent module. May be. In these cases, the generation of the utterance candidate 46 is set in advance in the utterance candidate generation device 20, and it is sufficient to determine whether or not to generate the utterance candidate 46 in each module according to the setting.

  In the present embodiment, when the utterance pair selection unit 22 generates only the response to which “conformity” is given as the dialogue action as the utterance candidate 46, the utterance candidate 46 is generated in the next module. Deciding.

  In step S110, if the utterance candidate 46 is not generated in the next module according to the setting or the like, the process proceeds to step S162. On the other hand, when generating the utterance candidate 46 in the next module, the process proceeds to step S112. Note that in the utterance candidate generation device 20 of the present embodiment, the utterance candidate 46 is sequentially generated by the module of the next number in order to generate the utterance candidate 46 by the subsequent module, but as described above, the module Is set to be skipped, the utterance candidate 46 is generated by the setting destination module.

  Next, the related utterance selection unit 26 generates an utterance candidate 46. Therefore, in step S112, the user utterance is input to the related utterance selection unit 26. In the next step S114, morphological analysis of the user utterance is performed by the related utterance selection unit 26, and a word indicating a topic is extracted from the input user utterance.

  In the next step S116, it is determined whether or not the word-utterance pair 30 has an utterance corresponding to a word that matches the extracted word. If there is no utterance corresponding to the extracted word in the word-utterance pair 30, the process proceeds to step S126. On the other hand, when the utterance corresponding to the word that matches the extracted word is in the word-utterance pair 30, the process proceeds to step S118.

  In step S118, it is determined whether or not there is a dialogue action input from the dialogue control unit 14. If there is an interactive action, the process proceeds to step S120, and after generating the utterance candidate 46 from the utterance that matches the extracted word and the input interactive action, the process proceeds to step S124. Specifically, from the word-utterance pair 30, an utterance corresponding to a word that matches the extracted word and provided with an interactive action that matches the input interactive action is extracted, and the extracted utterance is the utterance candidate. 46 is generated. On the other hand, if there is no interactive action, the process proceeds from step S118 to step S122. In step S122, the utterance candidate 46 is generated from the utterance for the word that matches the extracted word, and then the process proceeds to step S124. Specifically, an utterance corresponding to a word that matches the extracted word is extracted from the word-utterance pair 30 and generated as an utterance candidate 46 by the extracted utterance.

  In the next step S124, as in step S110 described above, it is determined whether or not the utterance candidate 46 is generated in the next module. When the utterance candidate 46 is not generated in the next module, the process proceeds to step S162. On the other hand, when the utterance candidate 46 is generated in the next module, the process proceeds to step S126.

  Next, the utterance search unit 32 generates an utterance candidate 46. Therefore, in step S126, the user utterance is input to the utterance search unit 32. In the next step S128, the utterance search unit 32 performs morphological analysis of the user utterance and extracts words from the user utterance.

  In the next step S <b> 130, it is determined whether or not there is a dialogue action input from the dialogue control unit 14. If there is an interactive action, the process proceeds to step S132, and in step S132, a search query is constituted by the extracted word and the input interactive action. On the other hand, when there is no dialogue act input from the dialogue control unit 14, the process proceeds to step S134, and in step S134, a search query is constituted by the extracted words. In the next step S136, the search database 36 is searched using the search query, and the search result is acquired.

  In the next step S138, it is determined whether or not an utterance is included in the search result. If the search result does not include an utterance, that is, if the search result is zero, the process proceeds to step S144. On the other hand, if the search result includes an utterance, that is, if the search result is not 0, the process proceeds to step S140. In step S140, the utterance candidate 46 is generated from the utterance list including utterances included in the search result, and then the process proceeds to step S142.

  In step S142, as in step S110 described above, it is determined whether or not to generate the utterance candidate 46 in the next module. When the utterance candidate 46 is not generated in the next module, the process proceeds to step S162. On the other hand, when generating the utterance candidate 46 in the next module, the process proceeds to step S144.

  Next, the versatile utterance selection unit 38 generates an utterance candidate 46. Therefore, in step S144, the user utterance is input to the versatile utterance selection unit 38. In the next step S146, it is determined whether or not there is a dialogue action input from the dialogue control unit 14. If there is an interactive action, the process proceeds to step S148, where an utterance candidate 46 is generated from the utterance list 40 and the input interactive action. Specifically, an utterance candidate 46 is generated by selecting an utterance provided with an interactive action that matches the input interactive action from the utterance list 40. On the other hand, if there is no dialogue action input from the dialogue control unit 14, the process proceeds to step S 150, and the utterance candidate 46 is generated from the utterance list 40.

  In the next step S152, as in step S110 described above, it is determined whether or not to generate the utterance candidate 46 in the next module. When the utterance candidate 46 is not generated in the next module, the process proceeds to step S162. On the other hand, when the utterance candidate 46 is generated in the next module, the process proceeds to step S154.

  Next, the new topic utterance selection unit 42 generates an utterance candidate 46. Therefore, in step S154, the user utterance is input to the new topic utterance selection unit 42. In the next step S156, it is determined whether or not there is a dialogue act input from the dialogue control unit 14. If there is an interactive action, the process proceeds to step S158, and in step S158, an utterance candidate 46 is generated from the utterance list 44 and the input interactive action. Specifically, an utterance candidate 46 is generated by selecting an utterance to which an interactive action that matches the input interactive action is selected from the utterance candidates 46. On the other hand, if there is no dialogue action input from the dialogue control unit 14, the process proceeds to step S 160, and the utterance candidate 46 is generated from the utterance list 44.

  In the next step S162, the utterance candidate 46 generated by each module is output to the output unit 18, and then this process is terminated. In the present embodiment, the utterance candidates 46 generated in each module are collectively output to the output unit 18 as described above. However, the present invention is not limited thereto, and the utterance candidates 46 generated in each module are sequentially output. You may make it output to the part 18. FIG.

  The output unit 18 selects an utterance from the utterance candidates 46 generated and output by the utterance candidate generation device 20, and responds to the user.

  Further, in the dialogue system 10 of the present embodiment, the utterance to which the dialogue act such as “conformity”, which is regarded as insufficient in the content as the dialogue, is given so that only the utterance (response) having no content is answered. The response using only (response) is not made. Specifically, when the output unit 18 of the dialog system 10 according to the present embodiment selects an utterance (response) to which “conversation” is given as a dialog action from the utterance candidates 46 and responds to the user Then, another utterance is selected from the utterance candidates 46 and responded to the user. In the dialogue system 10, a method for not responding only to an utterance to which a dialogue act such as “conformity” is given is not particularly limited. For example, a flag is given to such an utterance (response) Then, the utterance candidate 46 is generated, and when the output unit 18 selects the utterance to which the flag is assigned from the utterance candidate 46 and responds to the user, the output unit 18 continues another utterance from the utterance candidate 46. Select and respond to the user.

  The dialogue system 10 performs a dialogue such as the dialogue example shown in FIG.

  FIG. 11 shows a specific example of a dialog that can be realized between the user and the dialog system 10. Conversations example shown in FIG. 11, first, the interactive system 10 (system) is the system prompts (initial utterances) presented to the user to "Hello".

  In response to the presentation, when the user (you) makes a user utterance “Tomo”, the user is input from the user utterance input unit 12 to the utterance candidate generation device 20. In response to the user utterance, the utterance pair selection unit 22 responds to the user utterance, and the utterance selected by the output unit 18 from the utterance candidate 46 generated by the response selected from the utterance pair 24 is uttered to the user. To do.

  Further, when the user utters “I want to go to Italy” in response to the response, the user utterance is input from the input unit 12 to the utterance candidate generation device 20. When the user utters in this way, the user utterance is sequentially input from the input unit 12 to the utterance candidate generation device 20, and the utterance is referred to by each module with reference to the input user utterance and conversation history as described above. Candidate 46 is generated. The output unit 18 selects an utterance from the generated utterance candidates 46 and responds to the user.

  In the dialogue example shown in FIG. 11, in response to the user utterance “I want to go to Italy”, the output unit 18 selects “Italy go to Italy” selected from the utterance candidates 46 generated by the related utterance selection unit 26 (second module). "?" In response to the user utterance “I want to go”, the output unit 18 responds “I want to go” selected from the utterance candidates 46 generated by the utterance pair selection unit 22 (first module). In response to the user utterance “Dane”, the output unit 18 responds “Yes” selected from the utterance candidates 46 generated by the utterance pair selection unit 22 (first module). In this case, since “Yes” is assigned as “conversation” to “Yes”, the output unit 18 continues to the related utterance selection unit 26 (second module) and the utterance search unit 32 (third module). The utterance search unit 32 selected from among the utterance candidates 46 generated by the user responds “Italian wine is good” by the utterance candidate 46 generated. In response to the user utterance “delicious”, the output unit 18 responds “Yes” selected from the utterance candidates 46 generated by the utterance pair selection unit 22 (first module). In this case, since “Yes” is assigned as “conversation” to “Yes”, the output unit 18 continues to the related utterance selection unit 26 (second module) and the utterance search unit 32 (third module). The utterance search unit 32 selected from the utterance candidates 46 generated by the user responds “Italian cuisine delicious” by the utterance candidates 46 generated. For the user utterance “pasta also”, the output unit 18 selects the utterance selected from the utterance candidates 46 generated by the related utterance selection unit 26 (second module) and the utterance search unit 32 (third module). The search unit 32 responds with the utterance candidate 46 generated, “Is Italian item still a pasta?”. Further, in response to the user utterance “I think”, the output unit 18 responds “I see” selected from the utterance candidates 46 generated by the versatile utterance selection unit 38 (fourth module). Furthermore, in response to the user utterance “What is it?”, The output unit 18 selects from the utterance candidates 46 generated by the new topic utterance selection unit 42 (fifth module) “What did you eat today? ".

  In FIG. 11, each time a user utterance is performed, a user utterance is input from the input unit 12 to the utterance candidate generation device 20, an utterance candidate 46 is generated by each module, and an utterance selected from the generated utterance candidate 46 is used. The output unit 18 responds to the user. Note that the output unit 18 repeatedly selects an utterance from the utterance candidates 46 generated by the utterance candidate generation device 20 and responds to the user until a predetermined timing in the dialogue with the user (for example, a timing at which the topic can be considered changed). You may make it do.

  As described above, in the dialogue system 10 of the present embodiment, a user utterance is input from the input unit 12 to the utterance candidate generation device 20 in a dialogue with the user. In addition, the dialogue control unit 14 estimates the dialogue act of the user utterance by the dialogue act estimator 16, and outputs the dialogue act of the dialogue system 10 determined according to the estimated result to the utterance candidate generation device 20. In the utterance candidate generation device 20, the utterance pair selection unit 22, the related utterance selection unit 26, the utterance search unit 32, the versatile utterance selection unit 38, and the new topic utterance selection unit 42 are sequentially input to the input user utterance and dialogue action. A corresponding utterance candidate 46 is generated. The utterance pair selection unit 22 generates an utterance candidate 46 based on a response selected from the utterance pair 24 according to the input user utterance. The related utterance selection unit 26 performs morphological analysis on the input user utterance by the morphological analyzer 28, extracts a word indicating a topic, and corresponds to the extracted word and outputs the utterance to which the input interactive action is given. An utterance candidate 46 is generated from the word-utterance pair 30. The utterance search unit 32 performs morphological analysis on the input user utterance by the morphological analyzer 34, extracts words, searches the utterance database 36 using the extracted words and the input interactive action as a search query, and searches the search results. The speech candidate 46 is generated based on the above. The versatile utterance selection unit 38 selects an utterance provided with an interactive action that matches the input interactive action from the utterance list 40 and generates an utterance candidate 46. The new topic utterance selection unit 42 selects an utterance provided with an interactive action that matches the input interactive action from the utterance list 44, and generates an utterance candidate 46. The output unit 18 selects an utterance from the utterance candidates 46 generated and output by the utterance candidate generation device 20, and responds to the user.

  As described above, the utterance candidate generation device 20 of the interactive system 10 according to the present embodiment generates the utterance candidates 46 in order from the module corresponding to the user utterance and generating a more appropriate utterance candidate 46 for the user utterance. be able to. In addition, the utterance candidate generation device 20 can generate utterance candidates 46 having different utterance qualities by including a plurality of modules. Therefore, an appropriate response can be made in the dialog system 10 according to the user utterance.

  Thereby, the interactive system 10 that performs a high-quality response can be realized. Dialogue such as chat is used as a lubricant for conversation between humans. By enabling a high-quality dialogue, the interaction between the human and the dialogue system becomes smoother, and the efficiency of collaboration between the computer and the user can be improved.

  In the present embodiment, five modules (the utterance pair selection unit 22, the related utterance selection unit 26, the utterance search unit 32, the versatile utterance selection unit 38, and the new topic utterance selection unit 42) are provided. Not limited to these, it is only necessary to include at least two of these.

  In addition, the utterance pair 24, the word-utterance pair 30, the utterance database 36, the utterance list 40, and the utterance list 44 may be provided outside the utterance candidate generation device 20 without being limited to the present embodiment.

  Although the present embodiment has been described in detail with reference to the drawings, the present embodiment is an example, and the specific configuration is not limited to the present embodiment, and departs from the gist of the present invention. Needless to say, the range of designs that are not included is included and can be changed according to the situation.

DESCRIPTION OF SYMBOLS 10 Dialogue system 22 Utterance pair selection part 24 Utterance pair 26 Related utterance selection part 30 Word-utterance pair 32 Utterance search part 36 Utterance database 38 Versatile utterance selection part 40 Utterance list 42 New topic utterance selection part 44 Utterance list

Claims (4)

An utterance candidate generation device that generates an utterance candidate for use in a response in a dialog system that interacts with a user by responding to a user utterance,
From a plurality of predetermined utterance pairs consisting of a user utterance and a response to the user utterance, select an utterance pair determined to match the input user utterance, and using the response of the selected utterance pair, A first module for generating utterance candidates;
The word related to the word representing the topic of the user utterance extracted from the input user utterance from the utterance using the related word related to the word predetermined for each word representing the topic of the user utterance. A second module for selecting an utterance using a related word and generating an utterance candidate using the selected utterance;
A third module that searches an utterance database storing a plurality of utterances using a search query using a word extracted from the input user utterance and generates an utterance candidate using the searched utterance;
A fourth module that generates an utterance candidate using at least one utterance of a plurality of utterances determined to be utterable regardless of a predetermined user utterance topic, and a predetermined user utterance topic comprising a fifth module for generating an utterance candidate with at least one utterance of a plurality of utterances for changing,
If the utterance candidate is not generated by the first module, the process proceeds to the second module. If the utterance candidate is not generated by the second module, the process proceeds to the third module, and the utterance candidate is generated by the third module. Is not generated, the utterance candidate generated by the first module, the utterance candidate generated by the second module, and the utterance candidate generated by the third module by proceeding to the fourth module An utterance candidate generation device that outputs the utterance candidates in preference to the utterance candidates generated by the fourth module and the utterance candidates generated by the fifth module .   At least one of the first module to the fifth module is based on a dialogue act of the dialogue system, which is determined based on dialogue act estimation information obtained by estimating a dialogue act representing an intention of the user's utterance. The utterance candidate generation device according to claim 1, wherein the utterance candidate is generated. An utterance candidate generation method for generating an utterance candidate for use in a response in a dialog system that interacts with a user by responding to a user utterance,
The first module selects an utterance pair that matches the input user utterance from a plurality of predetermined utterance pairs including a user utterance and a response to the user utterance, and uses the response of the selected utterance pair Generating utterance candidates,
The second module represents the topic of the user utterance extracted from the input user utterance from the utterance using a related word related to the word, which is predetermined for each word representing the topic of the user utterance. Selecting an utterance using the related word related to a word, and generating an utterance candidate using the selected utterance;
Searching with a search query using a word extracted from the input user utterance from an utterance database storing a plurality of utterances according to the third module, and generating utterance candidates using the searched utterances;
A step of generating an utterance candidate using at least one utterance of a plurality of utterances that can be uttered regardless of a topic of user utterance determined in advance by the fourth module; and predetermined by a fifth module Generating an utterance candidate using at least one utterance of a plurality of utterances for changing the topic of the user utterance ,
If the utterance candidate is not generated by the first module, the process proceeds to the second module. If the utterance candidate is not generated by the second module, the process proceeds to the third module, and the utterance candidate is generated by the third module. Is not generated, the utterance candidate generated by the first module, the utterance candidate generated by the second module, and the utterance candidate generated by the third module by proceeding to the fourth module The utterance candidate generation method of outputting the utterance candidate in preference to the utterance candidate generated by the fourth module and the utterance candidate generated by the fifth module .   The utterance candidate production | generation program for functioning a computer as each module of the utterance candidate production | generation apparatus of Claim 1.

Images