JP6280342B2 - Function execution instruction system and function execution instruction method - Google Patents

Description

  The present invention relates to a function execution instruction system and a function execution instruction method that can instruct execution of a function.

  In recent years, especially in mobile terminals, there are an increasing number of voice agent applications that can perform various operations and searches in accordance with voice utterances of voice-recognized users. For example, there is a system capable of searching for contents such as music and moving images and operating functions such as e-mail and telephone by user's voice (for example, see Non-Patent Document 1). In order to realize such a system, the agent is connected to various search servers and various terminal functions. And the function (task) which a user wants to perform from a user's utterance is determined, and a task is performed in a server or a portable terminal according to the result.

  For example, in the system shown in Non-Patent Document 1, a task to be executed is determined as follows. When the user's utterance is “I want to know the news”, it is determined that a news search task is executed. When the user's utterance is “Call Mr. XX”, it is determined that the task of the telephone function is executed. If the utterance of the user is that "Hello", it is determined that chat tasks are performed.

  As described above, the system shown in Non-Patent Document 1 supports chatting in addition to normal search tasks and operation tasks. Chatting means that when a user's utterance does not execute a normal search task or operation task, the user responds with a voice or text output corresponding to the user's utterance (chat utterance), and the user This is a function for conversation (chat).

Ken Yoshimura, “Talking Concierge and Language Processing”, Information Processing Society of Japan Research Report. SLP, Spoken Language Information Processing 2012-SLP-93 (4), 1-6, 2012-10-19

  An example of a dialog system that realizes such a mechanism is shown in FIG. The dialogue system includes a dialogue agent, a chat response system, and a content search system. The dialogue agent is a system that receives an utterance from a user and returns a response. The dialogue agent determines which task is executed based on the user's utterance, acquires a response corresponding to the determined task from another system, and outputs the response to the user. The task to be executed is any one of chat, search and operation tasks.

  When the determined task is a chat task, the dialogue agent outputs the user's utterance to the chat response system. The chat response system inputs an utterance, generates a response sentence for the utterance (chat), and outputs the response sentence to the dialog agent as a response to the utterance. When the determined task is a search task, the dialogue agent requests the content search system to search according to the utterance. The content search system performs the requested search, and outputs the search result to the dialog agent as a response to the utterance. When the determined task is an operation task, the conversation agent activates and operates a function corresponding to the utterance that operates on the terminal (user terminal).

As an extension of such a system, it is also conceivable to realize searches and operations (tasks) by multiple voice conversations. An example of dialogue is shown below.
User “I am hungry”
Agent "What do you eat?"
User "Ramen"
Agent “There are 3 ramen shops nearby.
User "Yes"

  These are all dialogues within one task (for example, a gourmet search task), but the state changes as the dialogue progresses. For example, the first user's utterance is the utterance in the state where the task is activated (utterance for activating the task), and the second user's utterance is the utterance in the state where the cooking genre is searched (the cooking genre is specified). And the third user's utterance is an utterance in the confirmation input state.

An advanced chat response system not only simply responds to a user's utterance (chat response), but also generates a context corresponding to the interaction with the user so far, and responds according to the generated context. An example of a chat conversation that takes context into consideration is shown below.
User “I like ramen”
Agent “I like it too”
User “what you like best”
Agent "It's salt after all"

  In the above example, the agent sets “ramen” as the context by the first user's utterance. In the second user's utterance, the object of “like” is not shown, but the type of “ramen” can be answered by considering the set context.

Here, a dialogue system as shown in FIG. 9 is considered. In the dialog system described above, the tasks executed are independent of each other. Therefore, the dialogue agent passes the utterance to the chat response system only when the task to be executed is a chat task. When the chat response system is given an utterance, it creates a context based on the utterance and returns a chat response sentence. Here, consider the following dialogue.
User "I want to eat ramen" (task: gourmet search)
Agent “There are 3 ramen shops nearby.
User "Yes"
User “what you like best” (task: chat)
Agent "This is useful for my husband"

  In the above example, the user may want to answer what they like best about ramen, but the agent returns an answer that has nothing to do with ramen. This is because the user's first utterance of “I want to eat ramen” is for the gourmet search task to be executed, the utterance was not sent to the chat response system, and the context of “ramen” was not generated Is the cause. As described above, only when the task to be executed is a chat task, when the utterance is input to the chat response system, the conversation between the user and the agent becomes unnatural.

In order to solve this problem, a method of inputting all utterances to the chat response system and updating the context can be considered. However, in such a case, there is a possibility that the context will be generated even for the following utterances that are not meaningful to the user.
User "Search for sightseeing spots around Yokohama" (task: sightseeing spot search, context: Yokohama)
Agent “Search results displayed on the map”
User “Enlarge the map a little” (Task: Enlarge Map, Context: Map)
Agent "Enlarge Map"
User “Recommended” (task: chat)
Agent "I recommend the map function"

  In the above example, since the user operates the map, the agent makes a response related to the “map” in response to the user's utterance “recommended”. However, it is natural that a response related to “Yokohama” of the user's previous utterance is returned to the user's “recommended” utterance. Thus, even if the context is updated for each user utterance, the conversation between the user and the agent becomes unnatural.

  The present invention has been made in view of the above problems, and even when a function that considers a context is executed based on a series of sentences, a function execution instruction that can appropriately execute the function It is an object to provide a system and a function execution instruction method.

  In order to achieve the above object, a function execution instruction system according to the present invention includes a plurality of functions including a sentence input unit for inputting a sentence, and a function executed based on the sentence and the context input by the sentence input unit. A function execution instructing unit for instructing execution of the function, an execution function determining unit for determining a function to be instructed to be executed by the function execution instructing unit from a plurality of functions based on the text input by the text input unit, Based on the function determined by the means or the execution state of the function, the context update determination means for determining whether or not to update the context, and input by the sentence input means according to the determination by the context update determination means Context updating means for updating the context based on the received text.

  In the function execution instruction system according to the present invention, execution of a plurality of functions including functions executed based on text and context is instructed. It is determined whether or not the context is updated based on the function determined to be executed based on the text or the execution state of the function. Thus, in the function execution instruction system according to the present invention, the context can be updated when a sentence for executing the function whose context is to be updated is input, and the context is updated in other cases. It can be maintained without. Therefore, according to the function execution instruction system according to the present invention, the function can be appropriately executed even when the function considering the context is executed based on the series of sentences.

  The function execution instructing unit may be configured to instruct execution of a function that generates and outputs a response sentence based on the context with respect to the sentence as a function that is executed by inputting the sentence and context input by the sentence input unit. Good. According to this configuration, for example, the chat response function can be appropriately executed in consideration of the context based on a series of sentences.

  The function execution instruction system may further include response sentence generation means for inputting a sentence and context from the function execution instruction means, and generating and outputting a response sentence based on the context for the sentence. According to this configuration, for example, the chat response function according to the present invention can be reliably implemented.

  The text input unit may input text for each user, and the context update unit may update the context for each user. According to this configuration, the function can be appropriately executed for each user.

  The context update unit may reset the context according to the text input by the text input unit. According to this configuration, the context can be reset appropriately and reliably, and the function can be executed more appropriately.

  The function execution instruction system may further include voice recognition means for inputting voice, performing voice recognition on the input voice, and inputting the result of the voice recognition to the sentence input means. According to this configuration, the function can be executed by the user's voice.

  By the way, the present invention can be described as the invention of the function execution instruction system as described above, as well as the invention of the function execution instruction method as follows. This is substantially the same invention only in different categories, and has the same operations and effects.

  That is, the function execution instructing method according to the present invention is a function execution instructing execution of a plurality of functions including a sentence input step for inputting a sentence, and a function executed based on the sentence and context input in the sentence input step. An instruction function; an execution function determination step for determining a function to be executed in the function execution instruction step from a plurality of functions based on the sentence input in the sentence input step; and a function determined in the execution function determination step; Alternatively, based on the execution state of the function, a context update determination step for determining whether or not to update the context, and depending on the determination at the context update determination step, the context is determined based on the sentence input at the sentence input step. A context update step for updating.

  In the present invention, when a sentence for executing a function whose context is to be updated is input, the context can be updated. In other cases, the context can be maintained without being updated. Therefore, according to the present invention, even when a function considering a context is executed based on a series of sentences, the function can be appropriately executed.

It is a figure which shows the structure of the task determination apparatus which is a function execution instruction system which concerns on embodiment of this invention, and the structure of the system containing the said task determination apparatus. It is a table | surface which shows the context management table hold | maintained in a chat response apparatus. It is a figure which shows the flow of the response sentence production | generation in a chat response apparatus. It is a table | surface which shows the information used for response sentence preparation. It is a table | surface which shows the context update policy used for determination of the update of a context. It is a table | surface which shows another example of the context update policy used for determination of the update of a context. It is a figure which shows the hardware constitutions of the task determination apparatus which is a function execution instruction system which concerns on embodiment of this invention. It is a flowchart which shows the process (function execution instruction method) performed with the function execution instruction system which concerns on embodiment of this invention. It is a figure which shows an example of the dialogue system which performs the process according to a user's utterance.

  DESCRIPTION OF EMBODIMENTS Hereinafter, embodiments of a function execution instruction system and a function execution instruction method according to the present invention will be described in detail with reference to the drawings. In the description of the drawings, the same elements are denoted by the same reference numerals, and redundant description is omitted.

  FIG. 1 shows a task determination apparatus 10 that is a function execution instruction system according to the present embodiment. The task determination device 10 is a device that instructs execution of a plurality of preset functions. In the present embodiment, the function to be executed is called a task. The task execution instruction is issued to any one of the chat response device 30, the search task processing device 40, and the communication terminal 50. Further, as will be described later, a task to be executed is determined based on the speech of the user of the communication terminal 50. In addition, the task execution instruction is performed using the user's utterance of the communication terminal 50 as a trigger. That is, the function is executed when the user speaks. Since the task determination device 10 issues a task execution instruction, the communication terminal 50 is configured to be able to transmit and receive information to and from the voice recognition device 20.

  In addition, the tasks instructed to be executed by the task determination device 10 include, for example, types such as a search task, an operation task, and a chat task. The search task searches for information desired by the user. Specifically, the search task is a search for navigation information such as facility search such as gourmet search (restaurant search) and transfer guidance. The search is executed by the search task processing device 40, and search result information is output to the communication terminal 50.

  The operation task is for operating a function of the communication terminal 50. The functions of the communication terminal 50 to be operated include, for example, functions such as mail creation, music player, scheduler, incoming and outgoing calls, and camera. Specifically, an application program for executing these tasks is started in the communication terminal 50. Further, depending on the content of the task, an operation after the application program is started is also performed. For example, in the mail creation task, information is entered into the mail text and destination.

  The chat task outputs a response sentence constituting an appropriate conversation in response to the user's utterance. The response sentence is generated by the chat response device 30 and output to the communication terminal 50.

  In addition, the task instructed to execute may have a function of inputting words (arguments). This is called a task slot. The slot is predetermined for each task. For example, in the case of a transfer guidance task, the task is executed with the departure station or arrival station as an input to the slot. Alternatively, in the case of a gourmet search task, a store relating to the input dish is searched using the dish name as an input to the slot. In the case of a telephone directory search task, information such as a telephone number associated with the input person name is searched using the person name as an input to the slot. Note that zero or more slots may be defined for a task. That is, there may be a task without a slot. The above is the outline of each device 10 to 50 according to the present embodiment. Subsequently, each of the devices 10 to 50 will be described in more detail.

  The communication terminal 50 is a device that can communicate with the task determination device 10 and the voice recognition device 20 via a communication network (for example, a mobile communication network), such as a mobile phone (including a smartphone) or a PC. (Personal Computer) The communication terminal 50 may be a terminal equipped with a car navigation function used in a vehicle such as an automobile. Further, the communication terminal 50 may be capable of directly communicating with the chat response device 30 and the search task processing device 40.

  Since the above task is executed using the user's voice as a trigger, the communication terminal 50 has a function (voice recognition interface) for inputting the user's voice. In addition, the communication terminal 50 receives a command from the task determination device 10 and executes a task, and receives and outputs a task execution result from the task determination device 10, for example, an information reception function, information processing And a display function (for example, a browser for displaying various contents).

  That is, the communication terminal 50 has a function execution unit that executes one or more (or a plurality of) tasks. Specifically, the communication terminal 50 stores an application program corresponding to the task in advance, and executes the task by executing (starting) the stored application program. The task executed by the communication terminal 50 is instructed by the task determination device 10 as will be described later. In addition to executing the application, the task may be executed by acquiring information corresponding to the task from the network.

  The communication terminal 50 may execute the above function by executing an application (dialog application) for executing the above function on its own terminal.

  An instruction to execute a task by the task determination device 10 is performed as follows. First, the user speaks according to a task to be executed on the communication terminal 50. When the utterance is made, the communication terminal 50 inputs the user's voice (utterance). The input voice is transmitted from the communication terminal 50 to the voice recognition device 20. The voice recognition device 20 receives voice from the communication terminal 50 and performs voice recognition on the received voice. The voice recognition device 20 transmits a sentence (text) as a result of the voice recognition to the communication terminal 50. The communication terminal 50 receives the result of speech recognition and further transmits it to the task determination device 10. In this transmission, the communication terminal 50 transmits the task determination apparatus 10 together with a context ID that is information for specifying the context. The context will be described later. Note that the context ID is preset for each user and stored in the communication terminal 50. The task determination device 10 receives the result of the voice recognition, determines a task to be executed based on the result of the voice recognition, and instructs the execution of the task.

  If the task to be executed is a search task, the task determination device 10 makes a search request to the search task processing device 40. For example, the task determination device 10 transmits information indicating a search to be executed (for example, a search keyword or a search condition) to the search task processing device 40. The search task processing device 40 receives the search request and executes a search according to the request. The search task processing device 40 transmits information indicating the search result to the task determination device 10. The task determination device 10 receives information indicating the search result and transmits it to the communication terminal 50. The communication terminal 50 receives information indicating the search result from the task determination device 10 and performs output such as display output.

  If the task to be executed is an operation task, the task determination device 10 instructs the communication terminal 50 to perform an operation. For example, the task determination device 10 transmits information indicating the function to be operated to the communication terminal 50. The communication terminal 50 receives an instruction from the task determination device 10 and executes an operation according to the instruction.

  If the task to be executed is a chat task, the task determination device 10 inputs a sentence that is a result of speech recognition to the chat response device 30 and requests to generate a response sentence that is a response sentence to the sentence. I do. In addition, the task determination device 10 inputs a context ID, which is information for specifying a context, to the chat response device 30 when a response sentence generation request is made. The context is information indicating a topic in a series of user utterances (sentences based on the user utterances). The chat response device 30 receives the text and context ID from the task determination device 10, specifies the context from the context ID, and generates a response text for the text based on the text and context. The chat response device 30 transmits the generated response sentence to the task determination device 10. The task determination device 10 receives the response sentence and transmits it to the communication terminal 50. The communication terminal 50 receives the response text from the task determination device 10 and displays the response text or outputs a voice or the like as a response to the user's utterance.

  As described above, the communication terminal 50 interactively responds to the user's utterance. That is, the task determination device 10 constitutes an interactive agent for the communication terminal 50.

  Note that the text that is the result of voice recognition may be directly transmitted from the voice recognition device 20 to the task determination device 10 without being sent to the communication terminal 50 once. Furthermore, the task determination device 10 and the voice recognition device 20 may be integrated.

  The voice recognition device 20 is a device that is a voice recognition unit that inputs voice (voice data), performs voice recognition on the input voice, and outputs a result of the voice recognition. Specifically, as described above, the voice recognition device 20 receives voice data from the communication terminal 50. The voice recognition device 20 has a voice recognition engine, and performs voice recognition using the voice recognition engine. For speech recognition itself, any conventional speech recognition method can be used. The voice recognition device 20 acquires the voice recognition result as a sentence and transmits it to the communication terminal 50. Note that the sentence here may include arbitrary character information that can be obtained by voice recognition such as a set of words (a group of words including one or more words) and used for task determination.

  As described above, the chat response device 30 is a device that generates a response sentence for a sentence produced by a user's utterance. The chat response device 30 is connected to the task determination device 10 via a communication network or the like, and can communicate with the task determination device 10. As shown in FIG. 1, the chat response device 30 includes a context generation unit 31 and a chat response generation unit 32.

  The context generation unit 31 is a unit that receives a context generation request from the task determination device 10 and generates a context based on the request. Specifically, the context is a word indicating a topic such as “ramen”, “Yokohama”, or “weather” as shown in FIG. The chat response device 30 stores a context management table that is information that associates a context and a context ID that is information for specifying the context as shown in FIG. 2, so that the context can be specified from the context ID. ing. The context ID is, for example, information for specifying a context for each user of the communication terminal 50, and is assigned in advance for each user. In the example illustrated in FIG. 2, there is no context having the context ID “0001” (user), and the context having the context ID “0002” (user) is “ramen”.

  The context generation request includes a context ID and a sentence by the user's utterance. The context generation unit 31 extracts (generates) a context from the text included in the request. Specifically, a conventional method can be used to generate the context. For example, it is described in Barbara J. Grosz and Candace L. Sidner, “Attention, intentions, and the structure of discourse Computational Linguistics”, Volume 12 Issue 3, 1986, Pages 175-204, MIT Press. It can be generated from the sentence by the user's utterance according to the center theory. In particular, Yoshida Atsuko, who applied this theory to Japanese, Humanities Review: Bulletin of the Department of Humanities, Faculty of Humanities, Mie University 20, pp. 193-202, 2003, Mie University. (Non-Patent Document 3), Mitsuko Takei, Miho Fujiwara, Teruaki Aizawa, Centering Theory and Zero Pronoun: From the results of Japanese corpus analysis and native speaker survey. According to the method described in the 12th Annual Conference of the Language Processing Society, pp. 292-295, 2006. (Non-Patent Document 4) Nouns can be extracted as context. Note that the method for generating the context is not limited to the above, and any method can be used as long as the context is generated from the text.

  When generating the context, the context generation unit 31 updates the context corresponding to the context ID included in the context generation request in the context management table with the generated context. For example, when the sentence included in the request is “I like ramen”, the context generation unit 31 generates a context with “ramen” and corresponds to the context ID included in the request in the context management table. Update the context to "Ramen".

  Note that if the text included in the context generation request is not appropriate for context generation (extraction), context generation is not performed. The sentence that is not appropriate for generating the context is, for example, a sentence that does not include a word that can become a topic. In that case, the context generation unit 31 does not update the context management table. That is, the context updated before that is directly associated with the context ID.

  When the context generation unit 31 receives a context reset request from the task determination device 10, the context generation unit 31 deletes the context corresponding to the context ID included in the request from the context management table. That is, there is no context corresponding to the context ID. The context generation unit 31 may store the timeout time in advance, and delete the context from the context management table when the timeout time elapses from the generation or update of the context. In other words, the context may be deleted not only by reset but also by timeout. Further, the context generation unit 31 may delete the context ID at the same time when deleting the context from the context management table. In this case, when there is a context generation request, the context generation unit 31 newly generates a context ID included in the request in the context management table. By deleting the context ID as described above, the size of the context management table can be reduced. When the processing according to the request from the task determination device 10 is completed, the context generation unit 31 sends a response to that effect to the task determination device 10.

  The chat response generation unit 32 is a response sentence generation unit that inputs a sentence and a context, and generates and outputs a response sentence (chat response) that is a response sentence based on the context with respect to the sentence. The chat response generation unit 32 receives a request for generating a response sentence from the task determination device 10 and generates a response sentence based on the request. The request for generating a response sentence includes a context ID and a sentence by the user's utterance. The chat response generation unit 32 refers to the context management table and specifies (inputs) a context corresponding to the context ID included in the request. The chat response generation unit 32 generates a response sentence based on the specified context for the sentence included in the request.

  Specifically, a conventional method can be used to generate a response sentence considering the context. For example, H. Sugiyama, T. Meguro, R. Higashinaka, and Y. Minami, “Open-domain Utterance Generation for Conversational Dialogue Systemsusing Web-scale Dependency Structures”, Proceedings of the SIGDIAL 2013 Conference, pages 334-338, Metz, France , 22-24 August 2013. (Non-patent document 5).

  FIG. 3 shows a flow of response sentence generation processing by the method described in the document. First, the chat response generation unit 32 abstracts a sentence included in the request into a concept of an action (corresponding to a Dialogue act in the document) based on information such as the ending of the sentence. For example, the chat response generation unit 32 specifies an action using information (correspondence table between the ending and the action) indicating the correspondence between the ending and the action shown in FIG. 4A stored in advance. For example, if the ending of a sentence included in the request is "~", the action is "question". If the ending is “I want to do”, the act is “aspiration”. If the ending is "~", the action is "past fact".

  Subsequently, the chat response generation unit 32 determines the next action connected to the abstracted action, and sets this as the action of the response sentence (chat response). For example, the chat response generation unit 32 stores information indicating the correspondence between the previous action and the next action shown in FIG. 4B stored in advance (the next action correspondence table which is a rule for determining the next action). Use to determine the next action. For example, if the abstracted action is a “question”, the next action is “information provision”. If the abstracted action is “aspiration”, the next action is “agreement”. If the abstracted action is “past facts”, the next action is a “question”.

  On the other hand, the chat response generation unit 32 extracts a related word that is a word related to the context from the specified context (corresponding to Topic in the document). For example, the chat response generation unit 32 uses information (a related word dictionary that is a dictionary in which words related to a certain noun are described) indicating correspondence between the nouns and related words shown in FIG. 4C stored in advance. To extract related terms.

  Subsequently, the chat response generation unit 32 relates to the determined next action from among a large number of response sentences prepared and stored in advance, and the response sentence including the most extracted related words is included in the request. Select as a response to For example, the chat response generation unit 32 selects (generates) a response sentence using information (utterance index) indicating correspondence between the text (candidate response sentence) and the action shown in FIG. Note that the method of generating the response sentence is not limited to the above, and any method can be used as long as the response sentence is generated from the sentence and the context. The chat response generation unit 32 transmits the generated response sentence to the task determination device 10 as a reply to the request.

As described above, the following chat conversation is possible by considering the context in the generation of the response sentence.
User “I like ramen”
Agent “I like it too”
User “what you like best”
Agent "It's salt after all"

  In the chat conversation, the context becomes “ramen” by the user's first utterance. And by holding this information, a chat response that considers “ramen” is also returned for the user's next utterance. In the above example, a new context is not generated when the user utters “what you like best”.

  The search task processing device 40 is a means that receives (inputs) a search request from the task determination device 10 and executes a search according to the request. The search task processing device 40 is connected to the task determination device 10 via a communication network or the like, and can communicate with the task determination device 10. The search request includes information indicating a search to be executed (for example, a search keyword, a search condition, etc.). The search task processing device 40 performs a search based on information indicating a search to be executed. The search itself is performed in the same manner as the conventional method. A plurality of search task processing devices 40 may be provided according to the type of search. Search types include facility searches such as gourmet searches for searching for restaurant locations and navigation information searches such as transfer guidance. The search task processing device 40 transmits (outputs) the executed search result to the task determination device 10.

  As illustrated in FIG. 1, the task determination device 10 includes a sentence input unit 11, a command generation unit 12, a task determination unit 13, a context update determination unit 14, and a context update unit 15.

  The text input unit 11 is text input means for inputting text (input text, text data). Specifically, the sentence input unit 11 inputs a sentence from the communication terminal 50 by receiving a sentence that is a result of voice recognition by the voice recognition device 20. In addition, the text input unit 11 receives a context ID together with the document from the communication terminal 50. The text input to the text input unit 11 is a user's utterance unit for the communication terminal 50. That is, a sentence included in one utterance is treated as one unit. The text input unit 11 inputs text for each user. The following processing is performed for each user. The user is specified by, for example, the context ID described above. For example, when the task determination apparatus 10 stores information in which a user ID, which is information for specifying a user, is associated with a context ID in advance, when the communication terminal 50 transmits a sentence, the user ID is included in the sentence. , And the context ID may be specified from the user ID. The text input unit 11 outputs the input text and context ID to the task determination unit 13.

  The command generation unit 12 is a function execution instruction unit that instructs execution of a plurality of tasks. The task instructed to be executed by the command generation unit 12 includes a task executed based on the text and context input by the text input unit 11. Such a task is, for example, a chat task that is a task for generating and outputting an appropriate response sentence (chat response) based on the context of a sentence. In addition, the task that is instructed to be executed by the command generation unit 12 includes a task that is executed without being based on the context. Such a task is, for example, the above-described search task or operation task. The task instructed to be executed by the command generation unit 12 is determined by the task determination unit 13 as described later.

  The command generator 12 generates a command that is information to be transmitted to the communication terminal 50 and transmits the command to the communication terminal 50. For example, when the task instructing execution is a search task (content search), the command generation unit 12 sends a search request to the search task processing device 40 based on information input from the task determination unit 13. Send. Subsequently, the command generation unit 12 receives a search result transmitted from the search task processing device 40 in response to the request. The command generator 12 includes information indicating the executed (executed) task and search result information (search content) in the command. For example, if the user's utterance is “display nearby restaurants”, the command includes information indicating restaurant information display as information indicating a task, and peripheral restaurant information as search content. The communication terminal 50 that has received the command displays nearby restaurant information included in the command.

  In addition, when the task instructing execution is an operation task, the command generation unit 12 uses the information input from the task determination unit 13 on the communication terminal 50 such as telephone, mail, or camera activation. Include content (information indicating the task to be executed) instructing the operation of the function in the command. For example, when the user's utterance is “camera activation”, the command includes information on camera activation as information indicating a task. The communication terminal 50 that has received the command activates the camera function.

  When the task instructing execution is a chat task (chat response), the command generation unit 12 responds based on the information (text ID and text by the user's utterance) input from the task determination unit 13. A request for generation is transmitted to the chat response device 30. Subsequently, the command generation unit 12 receives a response sentence transmitted from the chat response device 30 in response to the request. The command generator 12 includes information indicating the executed (executed) task and a response sentence (chat response) in the command. For example, when the user's utterance is “Thank you always”, the command includes information about chat as information indicating a task, and “I am glad to help” as a response sentence. The communication terminal 50 that has received the command outputs a response sentence included in the command by voice or display.

  The task determination unit 13 is an execution function determination unit that determines a task for which execution is instructed by the command generation unit 12 from a plurality of functions based on the text input from the text input unit 11. For example, the task determination unit 13 may determine (determine) a task using a learning model (determination rule) obtained by machine learning. The task decision results in a document classification problem of which task a sentence is classified into. Therefore, for example, utterance cases associated with tasks are collected in advance. Machine learning is performed using the utterance example as correct answer data (sample data), and a task is determined using a learning model obtained by machine learning. The task determination unit 13 inputs the text input from the text input unit 11 to a task classifier based on the learning model, and determines a task by the task classifier. The task determination unit 13 only needs to be able to use a task classifier based on a learning model obtained by machine learning, and the task determination device 10 does not necessarily need to perform machine learning. In that case, the task determination apparatus 10 acquires in advance information indicating a learning model from the apparatus that performed the machine learning described above.

  In addition to the above-described method by machine learning, the task determination unit 13 sets a score for each word or category in advance for each task, and from the word included in the sentence or the category associated with the word, A score may be identified and a task may be determined based on the score. For example, the task with the highest total score may be determined as a task for which execution is instructed. The score of the word or category in this case is determined according to the degree of association with the task.

  Moreover, it is good also as a task which performs a chat task when an appropriate task is not determined by said task determination method. Further, the task determination unit 13 may determine a task to be instructed for execution as long as it is based on the input text, and any method other than the above can be used. For example, Non-Patent Document 1 and methods described in the specifications of PCT / JP2013 / 064766 (basic application: Japanese Patent Application No. 2012-137842) and Japanese Patent Application No. 2012-137844, which are international applications based on the Patent Cooperation Treaty by the present applicant Etc. can be used.

  When the task determination unit 13 determines a task to be executed, the task determination unit 13 notifies the context update determination unit 14 of the task and the context ID input from the text input unit 11. Further, in accordance with the task notification, the task determination unit 13 outputs the text used for the task determination to the context update determination unit 14. When a response to the task notification is input from the context update determination unit 14, the task determination unit 13 requests the command generation unit 12 to instruct execution of the determined task. The request may include information specifying a task to be executed and information necessary for executing the task (information input to the slot).

  Specifically, for example, if the task is a chat task, the content of the execution of the task is a sentence serving as a basis for creating a response sentence. The sentence is input from the sentence input unit 11. In the case of a chat task, the context ID input from the text input unit 11 is included. In the case of a search task, it is information indicating a search to be executed (for example, a search keyword or a search condition). The task determination unit 13 extracts information indicating the search to be executed from the text input from the text input unit 11. A conventionally used method can be used for the extraction. In the case of an operation task, it is information indicating an operation to be executed. The task determination unit 13 extracts information indicating an operation to be executed from the text input from the text input unit 11 as well as information indicating a search to be executed.

  Moreover, the task determination part 13 is good also as holding | maintaining the information which shows the dialog state which is an execution state of a task, when the task by a user's voice is performed, and managing the said dialog state. Processing related to the task according to the dialog state is performed. For example, when a mail creation task is activated, the communication terminal 50 enters an “email creation” dialog state. In the communication terminal 50, the mail text and the destination are input by the user's voice in the dialog state “mail creation”. There may be a plurality of dialog states in one task. For example, when a navigation task is executed, there may be a plurality of dialog states corresponding to information to be input by the user's voice such as “task activation”, “destination input”, “confirmation input”, etc. . The task determination unit 13 determines these dialog state transitions based on the text input from the text input unit 11. This determination is performed in the same manner as the determination of the task instructed to execute.

  For example, in the case of a navigation task, the following dialog state is obtained. First, when the user utters “navigation activation”, the “task activation” dialog state of the navigation task is entered. In this case, the navigation task is activated, and the agent outputs “Where are you going?” To prompt the destination to be input. On the other hand, when the user speaks “I want to go to Yokohama”, the dialogue state of “destination input” is entered. In this case, information for navigation is acquired based on the input utterance, and the agent performs an output prompting confirmation of the output of the information “Yokohama was found. Do you want to start guidance?”. On the other hand, when the user utters “Guide”, the dialogue state of “Confirmation Input” is entered. In this case, navigation to Yokohama by the navigation task (for example, output of information indicating a route to Yokohama) is executed.

  Further, for example, in the case of a gourmet search task, the following dialogue state is set. First, when the user speaks “I am hungry”, a gourmet search task “task activation” is entered. In this case, the gourmet search task is activated, and the agent outputs “Who wants to eat?” And prompts the user to enter a food genre subject to gourmet search. On the other hand, when the user utters “ramen”, the dialog state “select food category” is entered. In this case, a gourmet search is executed based on the input utterance, and the agent performs an output prompting confirmation of the output of the information “There are three cases. On the other hand, when the user speaks “Yes”, the dialogue state of “confirmation input” is entered. In this case, presentation of information by a gourmet search task (for example, output of information indicating a route to the nearest ramen shop) is executed.

  Also, the dialog state transitions when, for example, another task is activated or when the task being executed is terminated. When another task is activated, the conversation state relating to the other task is entered. Further, when the task being executed is completed, the initial state in which no task is executed is obtained. The task is terminated when the task to be executed is completed normally (for example, in the case of an e-mail creation task, e-mail creation is completed and e-mail transmission is performed), or the execution of the task is terminated by the user's voice, etc. Is performed (for example, when the user speaks “cancel”).

  When managing the dialog state, the task determination unit 13 notifies the context update determination unit 14 of the dialog state in addition to or instead of the task instructed to be executed.

  The task determination unit 13 may determine whether to reset the context based on the text input from the text input unit 11. For example, the task determination unit 13 determines to instruct execution of a task that resets the task activation state when a text “cancel” is input (when the user's utterance is “cancel”) or the like. To do. In this case, the task determination unit 13 determines to reset the context. When the context is reset, the task determination unit 13 notifies the context update unit 15 of the fact together with the context ID input from the text input unit 11.

  The context update determination unit 14 is a context update determination unit that determines whether to update (generate) a context based on the task determined by the task determination unit 13. Specifically, the context update determination unit 14 stores the context update policy illustrated in FIG. 5 in advance in order to perform context update determination. The context update policy is information that defines whether or not the context should be updated for each task. Specifically, the context update policy is information in which a task and an update flag are associated with each other. When receiving the notification of the task determined from the task determination unit 13 and inputting a sentence, the context update determination unit 14 refers to the update flag associated with the task notified from the task determination unit 13 in the context update policy. . The context update determination unit 14 determines to update the context if the referenced update flag is “1”, and determines not to update the context if it is “0”.

  The context update policy is generated in advance by, for example, an administrator of the task determination device and stored in the context update determination unit 14. The genre (category) of words included in the utterance is determined to some extent for each task, and a context update policy can be generated depending on whether these genres (categories) of words are suitable as contexts. For example, taking FIG. 5 as an example, the navigation task is a task for guiding to a destination, and the utterance related to the task always includes the name of the destination and the facility. These are suitable as chatting contexts. Alternatively, the gourmet search task includes a gourmet genre, which is also suitable as a chat context. On the other hand, such a word is not included in a task for manipulating the volume and the size of the map, so it is not suitable as a context. In addition, in the case of outgoing calls and mails, destinations and the like are included, but these person names are not suitable as a chatting context.

  In addition, when the dialog state is notified in addition to the task from the task determination unit 13, the context update determination unit 14 may determine whether to update the context based on the dialog state. . In this case, specifically, in order to perform context update determination, the context update policy shown in FIG. 6 is stored in advance. The context update policy is information that defines whether the context should be updated for each combination of task and dialog state. Specifically, the context update policy is information in which a combination of a task and dialog state is associated with an update flag. .

  The context update determination unit 14 refers to the update flag associated with the combination of the task and the dialog state notified from the task determination unit 13 in the context update policy. The context update determination unit 14 determines to update the context if the referenced update flag is “1”, and determines not to update the context if it is “0”.

  In the context update policy in this case, for example, an update flag in a dialog state for filling a specific slot such as a place name or a restaurant genre is set as “1”.

  When the context update determination unit 14 determines to update the context, the context update determination unit 14 notifies the context update unit 15 to that effect. The context update determination unit 14 outputs the text input from the task determination unit 13 to the context update unit 15 for context update together with the context ID input from the task determination unit 13. In this case, when receiving a response indicating that the context update is completed from the context update unit 15, the context update determination unit 14 responds to the task determination unit 13 with respect to notification of a task or a task and a dialog state. When the context update determination unit 14 determines not to update the context, the context update determination unit 14 does not notify the context update unit 15 but responds to the task determination unit 13 with the notification of the task or the task and the conversation state. I do.

  The context update unit 15 is a context update unit that updates the context based on the text input from the context update determination unit 14 according to the determination by the context update determination unit 14. The context update unit 15 updates the context by transmitting a context generation request to the chat response device 30 when receiving a notification from the context update determination unit 14 that it is determined to update the context. . The context update unit 15 includes the sentence input from the context update determination unit 14 in the request as a sentence for generating a context. Further, the context update unit 15 includes the context ID input from the context update determination unit 14 that is a context update target in the request. The context ID corresponds to the user related to the utterance input from the communication terminal 50. That is, the context update unit 15 updates the context for each user.

  In addition, when the context update unit 15 receives a notification to reset the context from the task determination unit 13, the context update unit 15 transmits a context reset (erase) request to the chat response device 30. That is, the context update unit 15 resets the context in accordance with the text input by the text input unit 11. The above is the functional configuration of each device 10 to 50 according to the present embodiment.

  FIG. 7 shows a hardware configuration of a server device constituting the task determination device 10, the speech recognition device 20, the chat response device 30, and the search task processing device 40 according to the present embodiment. As shown in FIG. 7, the server device includes a central processing unit (CPU) 101, a random access memory (RAM) 102 and a read only memory (ROM) 103 that are main storage devices, a communication module 104 for communication, and a hard disk. The computer is configured to include a computer including hardware such as the auxiliary storage device 105. When these components operate according to a program or the like, the functions of the task determination device 10, the speech recognition device 20, the chat response device 30, and the search task processing device 40 described above are exhibited. The above is the configuration of each device according to the present embodiment.

  Subsequently, a function execution instruction method that is a process executed by the task determination apparatus 10 that is the function execution instruction system according to the present embodiment will be described with reference to the flowchart of FIG.

  Since the process by the task determination apparatus 10 according to the present embodiment is performed, before the process is performed, first, a user operation on the communication terminal 50 is performed, and the user who causes the communication terminal 50 to execute a function Voice (utterance) is input. Subsequently, the voice is transmitted from the communication terminal 50 to the voice recognition device 20. The voice recognition device 20 receives and inputs the voice (not shown, voice recognition step). Subsequently, the voice recognition device 20 performs voice recognition on the input voice (not shown, voice recognition step). Information indicating a sentence as a voice recognition result is transmitted from the voice recognition device 20 to the communication terminal 50. The communication terminal 50 receives the information and transmits it to the task determination device 10. At this time, the context ID is also transmitted from the communication terminal 50 to the task determination device 10.

  In the task determination device 10, the sentence input unit 11 receives and inputs the sentence and the context ID that are the voice recognition results (S01, sentence input step). The input information is output from the text input unit 11 to the task determination unit 13. Subsequently, the task determination unit 13 determines a task to be instructed to be executed by the command generation unit 12 based on the text input from the text input unit 11 (S02, execution function determination step). At this time, the dialog state, which is the task execution state, may be determined based on the text.

  The determined task is notified from the task determination unit 13 to the context update determination unit 14 together with the text and the context ID. Based on the notified task, the context update determination unit 14 refers to a context update policy stored in advance and determines whether or not to update (generate) a context (S03, context update determination step). Note that the above dialog state may be taken into consideration in this determination. When it is determined that the context is to be updated (YES in S03), the context update determination unit 14 notifies the context update unit 15 of the fact and the text and the context ID are input.

  Subsequently, a request for generating a context including a sentence and a context ID is transmitted from the context update unit 15 to the chat response device 30 (S04, context update step). In the chat response device 30, the request is received by the context generation unit 31, and the context is generated from the text included in the request (S04, context generation step). When the context is generated, the context corresponding to the context ID included in the request in the context management table is updated with the generated context. If the sentence is not appropriate for context generation, the context is not updated and remains in the previously updated state. When the processing corresponding to the request from the task determination device 10 is completed by the context generation unit 31, a response to that effect is sent from the context generation unit 31 to the task determination device 10. If it is determined not to update the context (NO in S03), the process in S04 is not performed.

  Subsequent to the NO and S04 processing in S03, the task determination unit 13 requests the command generation unit 12 to instruct execution of the determined task. Subsequently, when the determined task is a chat task (YES in S05), a request for generating a response sentence is transmitted from the command generation unit 12 to the chat response device 30 (S06, function execution instruction step). ). In the chat response device 30, the request is received by the chat response generation unit 32, and a response sentence is generated from the text included in the request in consideration of the context (S06, response sentence generation step). The generated response sentence is transmitted from the chat response generation unit 32 to the task determination device 10 and received by the command generation unit 12. If the determined task is not a chat task (NO in S05), the process in S06 is not performed.

  Following the processing of NO in S05 and the processing of S06, a command is generated by the command generation unit 12 based on a request from the task determination unit 13 to instruct execution of the determined task (S07, function execution instruction step). If the task instructing execution is a search task, the search content is acquired from the search task processing device 40 when the command is generated. In this case, an instruction to execute the task is included in the command generation process. Subsequently, the generated command is transmitted from the command generation unit 12 to the communication terminal 50 (S08, function execution instruction step).

  The communication terminal 50 receives the command and performs an operation according to the received command. For example, in the case of a chat task, voice output or display output of a response sentence (chat response) is performed. In the case of a search task, display output of search content and the like are performed. If it is an operation task, the operation of the communication terminal 50 corresponding to the operation task is executed. As described above, particularly in the case of an operation task, an instruction to execute the task may be included in the command output process.

  Note that the above process corresponds to one utterance of the user, and if the user's utterance is repeated, the above process is repeated according to the repetition of the user's utterance. The above is the processing executed by the task determination device 10 that is the function execution instruction system according to the present embodiment.

By repeating the user's utterance, according to the above processing, the following interaction between the user and the agent becomes possible.
User "Search for sightseeing spots around Yokohama" (task: sightseeing spot search, context: Yokohama)
Agent “Search results displayed on the map”
User “make the map a little larger” (task: enlarge map, no context update)
Agent "Enlarge Map"
User “Recommend” (task: chat, no context update)
Agent "XX Tower is recommended"

  In the above dialogue, the chat task is executed only by the user's utterance for the utterance “Recommended”. However, a context is generated from the user's utterance “search for sightseeing spots around Yokohama” to execute the previous task, and the chat response takes the context into consideration. In addition, regarding the user's utterance “slightly enlarge the map” after the above utterance, considering only the utterance content, it is also possible to generate a context such as “map”. However, it is determined not to update the context from the task based on the utterance. Since the user's “Recommendation” utterance is thought to be a recommendation for “sightseeing spots around Yokohama” based on the flow of the story, the above response is natural. .

  As described above, according to the present embodiment, the context used in the execution of the chat task is a task determined as a task instructed to be executed based on the input sentence, or an execution state of the task. It is determined whether or not to update based on the dialog state. Thereby, in this embodiment, when an utterance for executing a function whose context is to be updated is made, the context can be updated, and in other cases, the context can be maintained without being updated. it can. Therefore, according to this embodiment, even if it is a case where the task which considered the context based on a series of sentences is performed, execution of a task can be performed appropriately. Specifically, in the case of a chat task, a response sentence to a series of utterances of the user can be made natural without causing the user to feel uncomfortable. At this time, the user does not need to be aware of whether the task to be executed is a chat or a search, and the task can be realized by interaction with the agent.

  In the present embodiment, it is determined whether or not the context should be updated using a task determined from the utterance, not the utterance (input sentence) itself. This has the following advantages. First, since the utterance is freely input by the user and there are infinite patterns, it is difficult to determine whether or not to update each utterance. On the other hand, since a task is a finite one generated from the utterance, it is easy to prepare a policy for each task.

  Furthermore, the genre (category) of words included in the utterance is determined to some extent for each task, and it is possible to generate a context update policy depending on whether the genre (category) of these words is suitable as a context. For example, utterances related to navigation tasks and gourmet search tasks usually include words (genres) suitable for context. On the other hand, utterances related to a task for manipulating the volume and the size of a map or a task for making a call or sending a mail usually do not include a word (genre) suitable for context. The context update policy can be easily created in consideration of the above. Note that whether or not the task is suitable as a task for updating the context is appropriately determined according to the task executed in consideration of the context, and is not necessarily limited to the above example.

  Moreover, it is good also considering the task performed considering a context like this embodiment as a chat task. According to this configuration, for example, the chat task can be appropriately performed in consideration of the context based on a series of sentences. In the present embodiment, the function execution instruction system does not include the chat response device 30 that creates a response sentence. However, the chat response device 30 may be included. According to this configuration, the chat task can be reliably performed.

  However, the task executed in consideration of the context is not necessarily a chat task, and the present invention can be applied to any task as long as the task is executed in consideration of the context.

  Moreover, a sentence may be input for every user like this embodiment, and a context may be updated for every user. According to this configuration, task execution can be instructed for a plurality of users, and task execution can be appropriately performed for each user.

  For a task that resets the state such as “cancel” utterance as in the present embodiment, the previous context may be reset (erased). According to this configuration, the context can be reset appropriately and reliably, and the task can be executed more appropriately.

  Further, as in the present embodiment, voice recognition may be performed by inputting voice, and the task may be executed as a sentence in which the voice recognition result is input. According to this configuration, the task can be executed by the user's voice. In the present embodiment, the function execution instruction system is configured not to include the voice recognition device 20, but may be configured to include the voice recognition device 20. That is, the function execution instruction system may include the task determination device 10 and the voice recognition device 20. When the function execution instruction system is configured not to include the voice recognition device 20, a sentence not based on voice recognition may be input to the function execution instruction system.

  Moreover, in this embodiment, although the task determination apparatus 10 and the speech recognition apparatus 20 were comprised separately, they may be comprised integrally. In this case, it is not necessary to transmit / receive the voice recognition result via the communication terminal 50. Further, the chat response device 30 may be configured integrally with the task determination device 10. A terminal used by a user such as the communication terminal 50 may be provided with at least one of the functions of the task determination device 10, the speech recognition device 20, and the chat response device 30. In that case, the terminal constitutes a function execution instruction system according to the present invention. Alternatively, the task determination device 10 may have a function of executing a task, and the task execution result may be provided to the communication terminal 50.

  In the present embodiment, in the chat response device 30, even when the chat task is executed, the context update and the response sentence generation are performed independently. That is, the chat response device 30 transmits information indicating the generated context to the task determination device 10, and then receives a request for generating a response message and generates a response message in response to the request. However, the context update and response sentence generation in the execution of the chat task may be performed as a series of processes. That is, after a context is generated, information indicating the context may not be transmitted to the task determination device 10 and a response sentence may be generated using the generated context.

  DESCRIPTION OF SYMBOLS 10 ... Task determination apparatus, 11 ... Text input part, 12 ... Command generation part, 13 ... Task determination part, 14 ... Context update determination part, 15 ... Context update part, 20 ... Speech recognition apparatus, 30 ... Chat response apparatus, 31 DESCRIPTION OF SYMBOLS ... Context production | generation part, 32 ... Chat response production | generation part, 40 ... Search task processing apparatus, 101 ... CPU, 102 ... RAM, 103 ... ROM, 104 ... Communication module, 105 ... Auxiliary storage device, 50 ... Communication terminal.

Claims (7)

A sentence input means for inputting sentences;
A function execution instruction means for instructing execution of a plurality of functions including a function executed based on a sentence and a context input by the sentence input means;
An execution function determination means for determining a function to be instructed to be executed by the function execution instruction means from the plurality of functions based on the text input by the sentence input means;
A context update determination unit that determines whether to update the context based on the function determined by the execution function determination unit or the execution state of the function;
Context updating means for updating the context based on the text input by the text input means in response to the determination by the context update determining means;
A function execution instruction system comprising:   The function execution instructing unit instructs execution of a function that generates and outputs a response sentence based on the context with respect to the sentence as a function executed with the sentence and context input by the sentence input unit as inputs. Item 4. The function execution instruction system according to Item 1.   The function execution instruction system according to claim 2, further comprising response sentence generation means for inputting a sentence and a context from the function execution instruction means, and generating and outputting a response sentence based on the context for the sentence. The sentence input means inputs a sentence for each user,
The function execution instruction system according to claim 1, wherein the context update unit updates the context for each user.   The function execution instruction system according to any one of claims 1 to 4, wherein the context updating unit resets the context according to a sentence input by the sentence input unit.   6. The speech recognition unit according to claim 1, further comprising speech recognition means for inputting speech, performing speech recognition on the input speech, and inputting the result of speech recognition to the sentence input means. Function execution instruction system. A sentence input step for inputting a sentence;
A function execution instruction step for instructing execution of a plurality of functions including a function executed based on the sentence and context input in the sentence input step;
An execution function determination step for determining a function to be executed in the function execution instruction step from the plurality of functions based on the sentence input in the sentence input step;
A context update determination step for determining whether to update the context based on the function determined in the execution function determination step or the execution state of the function;
In response to the determination in the context update determination step, a context update step for updating the context based on the text input in the text input step;
Function execution instruction method including

Images