JP2018010610A - Agent device, dialog system, dialog method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an efficient dialog system that can perform a large scale task based on small scale dialog control.SOLUTION: An agent 2001 that performs transfer of a dialog act which is information constituted by each of a front-end 2002 performing user interface processing and a back-end 2005 performing information processing, a dialog act type and incidental information in a dialog system 10, comprises a determination unit 200 that determines a dialog act for providing, according to the dialog act acquired from one of the front-end 2002 and the back-end 2005, and a provision unit 205 that provides the dialog act for providing which is determined by the determination unit 200 to one of the front-end 2002 and the back-end 2005, and the determination unit 200 determines the dialog act type in the dialog act for providing, not based on an incidental information sequence but based on a dialog act type sequence, in the dialog act sequentially acquired from a collection of the front-end 2002 and the back-end 2005.SELECTED DRAWING: Figure 2

Description

  The present invention relates to an interactive system for performing information processing by interacting with a user, and an agent device, an interactive method, and a program used in the interactive system.

  2. Description of the Related Art Conventionally, a system that identifies necessary information by interacting with a user using an apparatus such as a computer is known. In this system, it is necessary to perform appropriate control according to the state of dialogue with the user. Patent Document 1 discloses a technique for determining an action using a POMDP (Partially Observable Markov Decision Process) model according to a conversation state. Non-Patent Document 1 describes a dialogue action composed of a dialogue action type and an action item.

JP 2010-129033 A

Steve Young and 6 others, “The Hidden Information State model: A practical framework for POMDP-based spoken dialogue management”, Computer Speech & Language 24 (2010) 150-174

  By the way, in order to perform a large-scale task that draws thousands or more conclusions in a dialog system, it is necessary to efficiently perform control corresponding to the dialog state.

  Therefore, the present invention provides a dialog system that can perform control according to the dialog state more efficiently than the conventional technique. The present invention also provides an agent device, a dialogue method, and a program used in the dialogue system.

  In order to solve the above-described problem, an agent device according to an aspect of the present invention includes a predetermined configuration including a front end that performs user interface processing, a back end that performs information processing, a dialogue action type, and incidental information. An agent device that exchanges a dialogue act that is information in a dialogue act format, the decision unit deciding a providing dialogue act according to a dialogue act acquired from one of the front end and the back end, and the decision A providing unit that provides the providing interactive action determined by a unit to one of the front end and the back end, and the determining unit determines the interactive action type in the providing interactive action as the front action. In the dialogue action sequentially obtained from the end and the set of back ends, the dialogue action tie is not based on the series of incidental information. Based on the series, it is determined.

  In order to solve the above problems, an interactive system according to an aspect of the present invention includes the agent device described above, the front end for exchanging information expressed in a natural language with a user, and a knowledge base search. And the back end for providing information based on the above.

  In addition, in order to solve the above problem, a dialogue method according to an aspect of the present invention provides a front end for exchanging information expressed in a natural language with a user, and provides information based on a knowledge base search. Used in a dialog system including a back end, and each of the front end and the back end, and an agent device that exchanges a dialog action that is information of a predetermined dialog action format including a dialog action type and incidental information. A dialog method, wherein the agent device determines a providing dialog action according to a dialog action acquired from one of the front end and the back end, and the agent device is determined in the determination step. A providing step for providing the provided interactive action to one of the front end and the back end; In the determining step, the dialogue action type in the providing dialogue action is not based on the series of incidental information in the dialogue action sequentially obtained from the set of the front end and the back end. Based on the decision.

  In order to solve the above problem, a program according to an aspect of the present invention includes a front end that performs user interface processing, a back end that performs information processing, a dialogue action type, and incidental information. A program for causing a computer to execute a dialog control process for exchanging a dialog act that is information in a predetermined dialog act format, wherein the dialog control process is acquired from one of the front end and the back end. A determination step for determining a providing interactive action according to the determination, and a providing step for providing the providing interactive action determined in the determining step to one of the front end and the back end. In the step, the dialogue action type in the providing dialogue action is changed to the front end and the back environment. In sequentially acquired dialogue act from the set of de, based on the dialogue act type sequence not based on a series of supplementary information, determined.

  According to the present invention, control according to the conversation state can be performed efficiently.

It is a conceptual diagram which shows an example of the aspect of the service which the dialogue system which concerns on embodiment provides. It is a conceptual diagram which shows an example of the data center operating company which concerns on embodiment. It is a schematic block diagram of the dialogue system which concerns on embodiment. It is a flowchart which shows an example of the dialogue process in the dialogue system which concerns on embodiment. It is a figure which shows an example of the knowledge base which concerns on embodiment. It is a conceptual diagram which shows an example of the knowledge process which concerns on embodiment. It is a sequence diagram which shows an example of the sequence of the dialog in the dialog system which concerns on embodiment. It is a state transition diagram (the 1) about a dialog state. It is the state transition diagram (the 2) about a dialog state. It is the state transition diagram (the 3) about a dialog state. It is a state transition diagram (the 4) about a dialog state. It is a figure which shows the result of the experiment using the dialogue system which concerns on embodiment. It is a conceptual diagram which shows the service (type 1) which a dialogue system provides. It is a conceptual diagram which shows the service (type 2) which a dialogue system provides. It is a conceptual diagram which shows the service (type 3) which a dialogue system provides. It is a conceptual diagram which shows the service (type 4) which a dialogue system provides.

(Knowledge that became the basis of the present invention)
There are many scenes in which a task of specifying information as a conclusion that both parties are satisfied by exchanging information through dialogue between a specialist with rich expertise and a general person with little expertise. An example of such a task is an inquiry performed between a doctor and a patient. In an interview, there is a relationship between a doctor who has expertise but no information about the patient's subjective symptoms, and a patient who has limited expertise but can provide information about his or her symptoms by answering questions or by giving himself / herself. Therefore, the amount of information that becomes a clue by the dialogue is increased and reflected in the conclusion (diagnosis result). An interactive system that uses a machine such as a computer to perform a task of identifying information as a conclusion through dialogue with a user who is a general person, such as this inquiry, has been studied. For example, in the dialogue control by the POMDP model, uncertain input including errors such as speech recognition and language understanding can be handled, and it is expected that the action selection for performing the confirmation action for the user without excess or deficiency can be expected.

  However, in the case of an interview, the number of disease candidates that can be the conclusion (diagnosis result) of the interview is several thousand or more. In other words, in order to represent a process of identifying a disease candidate as a state transition based on information obtained from a patient, several thousand or more states including a logical conditional branch and a conclusion are required. When the number of states (dialogue states) necessary for dialog control increases to several thousand or more, the computation load for optimizing action selection such as confirmation actions for the user in dialog control increases. Large computer resources may be required. Until now, no technology has been studied for performing large-scale tasks with small-scale dialog control.

  In view of this, the present inventors have conceived a method of reducing the number of dialog states related to dialog control so that a large-scale task can be performed in a small dialog system. Then, by using this method, by reducing the number of dialogue states, a dialogue system is provided that can control the dialogue state more efficiently than the conventional technology. Also provided are an agent device, a dialogue method, and a program which are a part of the dialogue system.

  An agent device according to an aspect of the present invention is information of a predetermined dialogue action format including a front end that performs user interface processing, a back end that performs information processing, a dialogue action type, and incidental information. An agent device that exchanges a dialogue act, wherein the decision unit decides a dialogue act for provision according to a dialogue act acquired from one of the front end and the back end, and the provision decided by the decision unit A providing unit that provides an interactive action for one of the front end and the back end, and the determining unit determines the interactive action type in the providing interactive action as a set of the front end and the back end. Determined based on interactive action type sequence, not based on incidental information sequence, That. Here, the incidental information corresponds to the action item described above. According to this configuration, an action is determined based on a series of dialogue action types obtained by removing incidental information from the dialogue action. For this reason, since the action is determined from the series of dialogue action types, except for incidental information that causes the variation of the dialogue action, the number of dialogue states can be reduced. As a result, the optimization load in the case of optimizing action selection by reinforcement learning can be suppressed, and optimization can be facilitated. As a result, for example, control according to the conversation state can be efficiently performed so that a large-scale task can be performed in a small-scale conversation system.

  For example, the determination unit may determine a dialogue action type in the providing dialogue action based on a series of dialogue action types determined in the past. As a result, the agent device can appropriately determine the action from the series of dialogue action types as observation and action.

  In addition, the determination unit, based on a series of dialogue action types as observations, an action determination unit that determines a dialogue action type as an action, and sequentially acquires a dialogue action from the set of the front end and the back end, A separation unit that sequentially inputs a series of dialogue action types as observations to the action determination unit by separating incidental information from the acquired dialogue action, and a dialogue action type determined by the action determination unit and the separation unit. A synthesizing unit that generates a providing dialogue action by synthesizing the supplementary information, and the behavior determination unit determines the dialogue action type as the behavior as a behavior decided in the past. It may be performed based on the type series. Thereby, the agent device can use the supplementary information in the acquired dialogue act for the provided dialogue act, and the behavior determining unit can determine the dialogue act type as the behavior without depending on the supplementary information.

  The behavior determination unit may determine a behavior based on a POMDP (Partially Observable Markov Decision Process) model. Thereby, optimization of action selection can be performed by reinforcement learning.

  Further, the dialogue action type as the observation represents one of a plurality of predetermined dialogue action type values, and the dialogue action type as the action is a first kind of dialogue action type value directed to the front end. And a plurality of predetermined interactive action type values that are distinguished from the second kind of interactive action type value for the back end, and the providing unit is for providing determined by the determining unit Based on whether the dialogue action type in the dialogue action of the first type or the second kind represents the dialogue action type value, the destination of the dialogue action is selected and selected from the front end and the back end. It is good also as providing the said dialog act to a provision place. As a result, the agent device can provide not only the dialogue action to the front end for the user but also the dialogue action to the back end by distinguishing according to the dialogue action type.

  Further, the incidental information may represent a keyword. As a result, although various conversations using keywords (accompanying information) that can represent various information are possible, the action is determined based on the dialogue action type that excludes the incidental information from the dialogue action. The number of states can be reduced.

  In addition, a dialogue system according to an aspect of the present invention provides information based on the above-described agent device, the front end for exchanging information expressed in natural language with a user, and a knowledge base search. The back end. Thereby, for example, a task that requires specialized knowledge such as an inquiry task for identifying a disease by a question about a symptom can be realized. And in the dialogue for such tasks (that is, the exchange of dialogue acts), the behavior is determined from the series of dialogue act types except for incidental information that causes the variation of the dialogue act. The number can be reduced.

  Also, for example, the incidental information represents a keyword that can be used for the search of the knowledge base, and the back end uses the keyword represented by the incidental information in the interactive action acquired from the agent device to search the knowledge base. It is good also as providing the said agent apparatus with the interactive action produced | generated based on the result of the search process. This makes it possible to reduce the number of dialogue states even if there are a variety of keywords that can be used for knowledge-based searches and acquired from the user. Can be easily performed.

  Further, the back end determines whether or not the result of the knowledge base search process has been sufficiently narrowed down, and when it is determined that the result has been sufficiently narrowed down, provides the agent device with an interactive action representing the result. When it is determined that the search has not been sufficiently narrowed down, an interactive action representing a valid keyword may be provided to the agent device in order to narrow down the result of the search process. Thereby, narrowing down for specifying desired information can be performed interactively.

  The interactive method according to one aspect of the present invention includes a front end for exchanging information expressed in a natural language with a user, a back end for providing information based on a knowledge base search, and the front end. And a dialogue method used in a dialogue system including each of the back ends, and an agent device that exchanges a dialogue act which is information of a predetermined dialogue act type composed of a dialogue act type and incidental information, A determining step in which an agent device determines a providing dialogue action in accordance with a dialogue act acquired from one of the front end and the back end; and the providing device determines the providing dialogue determined in the determining step. Providing an action to one of the front end and the back end; and In the dialog, the interactive action type in the providing interactive action is determined based on the interactive action type sequence not based on the accompanying information sequence in the interactive action sequentially obtained from the set of the front end and the back end. . As a result, actions are determined from a series of dialogue action types, except for incidental information that causes a large variation in dialogue actions, so the number of dialogue states can be reduced, and the number of dialogue states can be reduced efficiently. Control can be performed.

  Further, for example, the determining step sequentially acquires dialogue actions from the set of the front end and the back end, and separates sub-steps for separating incidental information and dialogue action types in the obtained dialogue actions; and the separation sub-steps The action decision that determines the dialogue action type as an action by using the dialogue action type separated in step 1 as an observation and referring to the belief state information that is the standard for deciding the action corresponding to the observation based on the observation. A sub-step, a synthesizing sub-step for generating a dialogue act for provision by synthesizing the dialogue act type determined in the action determining sub-step and the incidental information separated in the separating sub-step; and the separating sub-step And a dialogue action type system determined in the action determination substep. It may include an update step of updating the belief state information in response to the actions and is. As a result, the belief state can be efficiently updated by the interactive action type sequence without using the incidental information.

  Further, the program according to an aspect of the present invention is information on a predetermined dialogue action format including a front end that performs user interface processing and a back end that performs information processing, a dialogue action type, and incidental information. A program for causing a computer to execute a dialog control process for sending / receiving a certain dialog action, wherein the dialog control process is a dialog for providing according to a dialog action acquired from one of the front end and the back end. A determination step of determining an action; and a providing step of providing the providing interactive action determined in the determining step to one of the front end and the back end, wherein the providing step includes the providing interaction The dialogue action type in the action is sequentially acquired from the set of the front end and the back end. In dialog act, on the basis of the dialogue act type sequence not based on a series of supplementary information, determined. If this program is installed in a computer, the computer functions as a device as a dialog agent, and the action is determined from the series of dialog action types except for incidental information that causes variations in dialog action. Can be reduced. Thereby, the optimization load in the case of optimizing action selection by reinforcement learning can be suppressed, and optimization can be facilitated.

  Note that these comprehensive or specific aspects may be realized by a system, method, integrated circuit, computer program, or recording medium such as a computer-readable CD-ROM, and the system, method, integrated circuit, computer. You may implement | achieve with arbitrary combinations of a program or a recording medium.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. Note that each of the embodiments described below shows a comprehensive or specific example of the present invention. The numerical values, shapes, components, arrangement of components, steps, order of steps, and the like shown in the following embodiments are merely examples, and do not limit the present invention. In addition, among the constituent elements in the following embodiments, constituent elements that are not described in the independent claims can be arbitrarily added. Each figure is a mimetic diagram and is not necessarily illustrated strictly.

(Overview of services provided)
First, an aspect of a service provided by the interactive system in the present embodiment will be described.

  FIG. 1A is a diagram illustrating an example of a service provided by the interactive system 10 according to the present embodiment. The dialogue system 10 includes a group 1100, a data center operating company 1110, and a service provider 1120.

  The group 1100 is, for example, a facility such as a company, an organization, or a home (for example, a building, a house, etc.), and may be of any size. The group 1100 includes a device 1101 and a home gateway 1102. The device 1101 has a user interface for exchanging information with the user in some way. The device 1101 is, for example, a device that can be connected to the Internet (for example, a smartphone, a personal computer (PC), a television, or the like). The device 1101 itself may be a device that cannot be connected to the Internet (for example, a washing machine, a refrigerator, or the like), or may be a device that can be connected to the Internet via the home gateway 1102. Note that the home gateway 1102 may not be necessary. A user 1010 uses the device 1101. The device 1101 may be a portable device that can be wirelessly connected to the Internet (for example, a smartphone, a notebook PC, etc.). In this case, the user 1010 and the device 1101 are not necessarily located in the group 1100. Also good. Note that the device 1101 is not limited to a single device, and the device 1101 may include a plurality of devices.

  The data center operating company 1110 includes a cloud server 1111. The cloud server 1111 is a computer that cooperates with various devices via the Internet, and is, for example, a virtualization server. The cloud server 1111 manages, for example, huge data (big data) that is difficult to handle with a normal database management tool or the like. The data center operating company 1110 performs data management, cloud server 1111 management, and the like. Note that the data center operating company 1110 is not limited to a management company that only manages data or manages the cloud server 1111, and may be a company that also performs other businesses. It may be an equipment manufacturer that is developing or manufacturing. Further, the data center operating company 1110 is not limited to a single company. For example, as shown in FIG. 1B, when a device manufacturer and a management company jointly or share data management or management of the cloud server 1111 Both the device manufacturer and the management company can correspond to the data center operating company 1110.

  The service provider 1120 includes a server 1121. The server 1121 is realized by, for example, one or a plurality of computers, and may be provided with a storage such as a large-capacity hard disk as a storage medium regardless of the scale, for example, only provided with a memory in a PC. May be. Note that the server 1121 does not necessarily include a storage medium.

  Next, the flow of information in the dialog system 10 of the above-described aspect will be described.

  First, the device 1101 of the group 1100 transmits the input information acquired via the user interface to the cloud server 1111 of the data center operating company 1110 directly or via the home gateway 1102. The user interface used for inputting input information by the user 1010 in the device 1101 is, for example, a keyboard, a touch panel, or a voice input device using existing voice recognition technology. The cloud server 1111 acquires input information from the device 1101.

  Next, the cloud server 1111 of the data center operating company 1110 transmits information based on the input information from the device 1101 to the server 1121 of the service provider 1120. The information based on the input information is information as a result of performing processing such as calculation on the input information without including the same information even if the information includes the same information as at least a part of the input information. There may be. The unit of information transmission based on this input information may be any unit. The server 1121 acquires information from the cloud server 1111.

  Then, the server 1121 of the service provider 1120 specifies information for provision to the user in response to the information from the cloud server 1111, and sends the provision information to the user 1010 to the cloud server 1111. Send. The cloud server 1111 transmits information to be provided from the server 1121 to the device 1101 or transmits information to the device 1101 as a result of processing such as computation on the information to be provided. As a result, information is provided to the user 1010 from the user interface of the device 1101. A user interface used for providing information to the user 1010 in the device 1101 is, for example, a display or a voice output device using an existing voice synthesis technique.

  The server 1121 may acquire input information from the user 1010 and provide information for provision by communicating with the device 1101 without using the cloud server 1111. Further, the dialog system 10 may be in a mode different from the above-described example. For example, in the interactive system 10, the data center operating company 1110 and the service provider 1120 may be omitted. Based on the input information acquired by the device 1101 from the user, the providing information is specified, and the providing information is displayed. It may be provided to the user.

(Embodiment 1)
(Constitution)
Hereinafter, as an example, a configuration for performing a dialogue method for exchanging information with a user in the dialogue system 10 using the POMDP model will be described.

  In general, the POMDP model is a model of a state set S, an action set A, an observation set Z, a state transition matrix T, an observation matrix O, a reward matrix R, and a belief state initial state B0. Defined by element. In a conventional dialogue system using the POMDP model in which an agent observes the current state in a certain environment and decides an action to be taken, the user is associated with the environment, the system (computer system) is associated with the agent, and the user's dialogue is performed. As an observation, the dialogue action of the system is regarded as an action. On the other hand, the dialog system 10 performs modeling with a structure of one agent and two environments.

  FIG. 2 is a schematic configuration diagram of the dialogue system 10.

  As shown in FIG. 2, the dialogue system 10 includes an agent 2001, a front end 2002 as a first environment, a back end 2005 as a second environment, and a knowledge base 206.

  The agent 2001 is a device (agent device) as a dialogue agent. The agent 2001 can be implemented in a computer including a processor (microprocessor), a memory, and the like, for example. Specifically, for example, the function of the agent 2001 can be realized by a processor executing a program stored in a memory. The agent 2001 mediates the exchange of dialogue actions between the front end 2002 and the back end 2005. That is, the agent 2001 exchanges a dialogue action with each of the front end 2002 and the back end 2005. The dialogue action is information in a predetermined format (predetermined dialogue action format) composed of a dialogue action type and incidental information.

  The front end 2002 is a device that performs user interface processing, and exchanges information expressed in a natural language with a user. The front end 2002 may be implemented on the same computer as the agent 2001 or on a different computer, for example. For example, the functions of the front end 2002 can be realized by a computer user interface and a processor that executes a program stored in a memory of the computer.

  The back end 2005 is a device that performs information processing such as search processing, and performs knowledge processing including provision of information based on an inquiry (search processing) to the knowledge base 206. The back end 2005 can be implemented on, for example, the same computer as the agent 2001 or a different computer. For example, the functions of the back end 2005 can be realized by a communication interface of a computer and a processor that executes a program stored in a memory of the computer.

  When the agent 2001 and each of the front end 2002 and the back end 2005 are realized by a plurality of devices separated from each other, the agent 2001 and each of the front end 2002 and the back end 2005 are directly or other devices. Communicate through.

  The knowledge base 206 is a device that holds information. For example, the knowledge base 206 is a database management device that manages data representing information in a structured manner. The knowledge base 206 can be implemented, for example, on the same computer as the back end 2005 or on a different computer. For example, the knowledge base 206 can be realized by a storage medium such as a memory or a hard disk in a computer. In the interactive system 10 for performing the inquiry task, the knowledge base 206 holds information useful for executing the inquiry task (for example, information related to symptoms and illness). The knowledge base 206 may be composed of a plurality of devices.

  FIG. 2 illustrates an interactive action in the interactive system 10 for performing an inquiry task. Observation 2003 is an example of an interactive action acquired by the agent 2001 from the front end 2002. An observation 2006 is an example of an interactive action acquired by the agent 2001 from the back end 2005. The action 2004 is an example of an interactive action that the agent 2001 provides to the front end 2002. An action 2007 is an example of an interactive action provided by the agent 2001 to the back end 2005. Observation 2003, observation 2006, action 2004, and action 2007 are all examples of information (dialogue action) in a predetermined dialogue action format. In the dialogue action of the character string information in this example, the part before the parenthesis (eg “Provide”, “Confirm”, etc.) is the dialogue action type, and the part in the parenthesis (eg “vertigo”, “headache”, etc.) Is incidental information. The dialogue action type represents one of a plurality of values (dialogue action type values) such as “Provide”, “Confirm”, “Check”, “push”, etc., which are set in advance. Represented by a column. In this example, the dialogue action type value is expressed in English words, but this is only an example, and may be expressed based on any language or rule. The incidental information is, for example, character string information indicating a keyword.

  In the dialogue system 10, tasks are performed in accordance with a series of dialogue actions exchanged between the agent 2001 and a set of the front end 2002 and the back end 2005. In the example of the inquiry task, the back end 2005 identifies a disease based on information collected by the agent 2001 through the front end 2002 interactively from the user.

  Hereinafter, more detailed configurations of the agent 2001, the front end 2002, and the back end 2005 will be described.

  As shown in FIG. 2, the front end 2002 includes an input processing unit 201 and an output processing unit 208 related to user interface processing. The agent 2001 includes a determining unit 200 and a providing unit 205. The back end 2005 includes a knowledge processing unit 207.

  The determination unit 200 has a function of determining an interactive action for provision to one of the front end 2002 and the back end 2005 according to the interactive action acquired from one of the front end 2002 and the back end 2005. In order to realize this function, the determination unit 200 includes a separation unit 202, an action determination unit 203, and a synthesis unit 204. Note that the determination unit 200 may be configured to determine the interactive action type in the providing interactive action based on a series of interactive action types determined in the past.

  The input processing unit 201 uses, for example, a natural language expression (character string) that is input from a user through a user interface such as a keyboard, a touch panel, or a voice input device, using an existing natural language processing technology (for example, a language understanding technology). Convert to a dialogue act and output the dialogue act with confidence. Thereby, the dialogue action and the reliability are acquired by the separation unit 202. The input processing unit 201 may perform preprocessing for converting the input voice into text using a voice recognition technique when the user input is voice.

  The separation unit 202 separates the dialogue action acquired from the input processing unit 201 of the front end 2002 or the knowledge processing unit 207 of the back end 2005 into a dialogue action type and incidental information. For example, the dialogue action “Provide (headache)” is separated into a dialogue action type “Provide” and accompanying information “headache”. The separation unit 202 inputs the separated dialogue action type to the action determination unit 203. That is, the separation unit 202 sequentially acquires the dialogue action from the set of the front end 2002 and the back end 2005, and separates the incidental information from the acquired dialogue action, so that the series of dialogue action types is sequentially observed as the observation by the action determination unit 203. input. Further, the separation unit 202 inputs the separated incidental information to the synthesis unit 204.

  The action determination unit 203 determines and outputs a dialog action type as an action based on a series of dialog action types sequentially input as observations by an existing action determination technique using a POMDP model. For example, the action determining unit 203 determines and outputs the dialogue action type “push”. Existing techniques can be used to learn the policies referenced in the POMDP model. In the case where the probability distribution (also referred to as belief state) of the state handled by POMDP takes only a value of 0 or 1, the model can be regarded as an MDP (Markov Decision Process) model. That is, the action determination by the action determination unit 203 includes an MDP model. In these models (POMDP model, MDP model), it is possible to optimize a policy referred to for selective action determination by reinforcement learning (Reinforcement Learning). Note that the action determination unit 203 determines an action with reference to the belief state and the policy based on the observation. The action determination unit 203 holds belief state information as information indicating the belief state, and can update the belief state information according to the observation and the determined action. Moreover, the action determination unit 203 may hold a history of dialog action types as the determined actions. And the action determination part 203 is good also as determining the dialog action type as an action based also on the series of the dialog action type as an action determined in the past.

  The combining unit 204 combines the interactive action type determined and output by the action determining unit 203 based on the interactive action type separated by the separating unit 202 and the incidental information separated by the separating unit 202, Generate and output an interactive action for provision. For example, the synthesizing unit 204 outputs a dialogue action “push (headache)”.

  The providing unit 205 has a function of providing the providing interactive action determined by the determining unit 200 to one of the front end 2002 and the back end 2005. The providing unit 205 can switch the environment that is the output destination of the interactive action output from the combining unit 204 to either the front end 2002 or the back end 2005. Specifically, the providing unit 205 determines in advance whether the output destination of the interactive action for provision should be the output processing unit 208 or the knowledge processing unit 207 in order to distinguish the output destination of the interactive action. Judgment based on the criteria. The criterion that has been determined in advance is, for example, a criterion that distinguishes the output destination depending on whether the first character of the dialogue action is uppercase or lowercase. For example, as shown by action 2004, providing section 205 outputs the interactive action to output processing section 208 if the first character of the providing interactive action (that is, the first character of the interactive action type) is capitalized. Further, for example, as shown by action 2007, providing section 205 outputs the interactive action to knowledge processing section 207 if the first character of the interactive action for provision (that is, the first letter of the interactive action type) is lowercase.

  The output processing unit 208 converts the semantic expression indicating the interactive action output from the providing unit 205 into a response sentence using a response sentence generation technique (for example, an existing sentence generation technique), and provides it to the user as a character string. The provision of the character string to the user is performed by a user interface such as a display or a voice output device, for example. The output processing unit 208 may perform post-processing for converting a character string to be provided to the user into a voice by a voice synthesis technique.

  The knowledge processing unit 207 refers to the knowledge base 206, performs information processing according to the dialogue action output from the providing unit 205, and provides the dialogue act indicating the result to the separation unit 202. Information processing in the knowledge processing unit 207 is, for example, search processing for searching for information from the knowledge base 206 and extracting it. For example, keyword processing as a character string represented by supplementary information in an interactive action acquired from the agent 2001 so far is performed. The list is set as a set of search keywords (search keywords), and a search process is performed to AND search the contents of the knowledge base. The knowledge processing unit 207 generates a dialogue action based on the search processing result, and provides the generated dialogue action to the separation unit 202 of the agent 2001. For example, the knowledge processing unit 207 determines whether or not the result of the search process has been sufficiently narrowed down, and a predetermined condition (for example, the number of pieces of retrieved information) is set to indicate that the search has been sufficiently narrowed down. If it is determined that the condition is sufficiently narrowed down, an interactive action indicating the search result is generated and output (provided to the separation unit 202). On the other hand, if it is determined that the search has not been sufficiently narrowed down, the knowledge processing unit 207 estimates an effective keyword (search keyword) to narrow down the search processing result, and generates a dialogue act indicating the keyword. And output (provide to the separation unit 202). An example of an interactive action indicating a keyword effective for narrowing down the search processing result is, for example, “suggest (fever)” including “fever” as incidental information.

  By providing such a configuration, the determining unit 200 of the agent 2001 extracts the interactive action types in the sequentially acquired interactive actions and provides the interactive actions to be output as actions based on the series of past interactive action types. Can determine the type of dialogue action. That is, the agent 2001 can determine a dialogue action type to be output based on a series of past dialogue action types without making a determination based on incidental information in the dialogue action.

  Thus, by determining the dialogue action type based on the type of dialogue action (each of the dialogue action type values) rather than the type of dialogue action (each of the values of the whole dialogue action), It can be kept low. Minimizing the number of dialogue states is advantageous for optimizing action selection by reinforcement learning. As a result of this optimization, the agent 2001 appropriately selects an action 2004 related to a question type or the like for acquiring (collecting) information as the observation 2003 from the front end 2002 which is one of environments. As a result of the information processing, it is possible to appropriately select the action 2007 for obtaining the observation 2006 from the back end that is another one of the environments. Through these selections (for example, selection of dialogue action types with different question types), the agent 2001 can perform dialogue control processing such as control of the subject of dialogue and control of confirmation of uncertain information.

  Note that the agent 2001 only needs to retain the latest acquired one for the incidental information separated from the dialog action type in the acquired dialog action.

  Further, in the back end 2005, if the acquired series of incidental information is accumulated and managed on a stack which is one area such as a memory after the start of the dialogue, it is possible to perform an advanced task using a plurality of incidental information. By the information processing in the back end 2005, for example, incidental information representing symptoms such as “headache”, “vertigo”, and the like is collected. Based on these symptoms, “sickness A”, “sickness B”, “sickness C”. Or the like can be identified as the desired information. In the back end 2005, the interactive action type of the acquired interactive action may be handled as, for example, a command type, and processing may be performed only according to the current interactive action type, and it is not necessary to manage the history of the interactive action type.

(Operation)
Hereinafter, the operation of the interactive system 10 having the above-described configuration will be described.

  FIG. 3 is a flowchart showing an example of dialogue processing in the dialogue system 10. Hereinafter, the operation of the agent 2001 related to the dialogue will be described with reference to FIG.

  In starting the dialogue, first, the agent 2001 sets a belief state (step S31). The belief state is a set of probability values (probability distribution) such that the sum is 1 with respect to the dialog state defined by combining the progress of the dialog and the user's intention. For example, assuming that the user is in one dialog state at the start of the dialog, the probability value of that state is 1, and the probability values of other states are 0. The belief state is set by the determination unit 200 of the agent 2001 holding belief state information indicating a belief state as an initial state in a storage medium such as a memory.

  The determination unit 200 of the agent 2001 determines an action based on the belief state indicated by the belief state information (step S32). Specifically, the behavior determination unit 203 of the determination unit 200 refers to a policy that is a set of criteria for what kind of behavior should be taken in what kind of belief state, and as an action based on the belief state Determine the interaction action type. The determination unit 200 informs the providing unit 205 of the providing interactive action including the interactive action type determined by the action determining unit 203.

  Next, the agent 2001 determines whether the environment to which the action determined in step S32 is directed is the front end 2002 or the back end 2005 (step S33). Specifically, the providing unit 205 determines an output destination according to a providing interactive action based on a predetermined criterion. The providing unit 205 provides the providing interactive action to either the output processing unit 208 of the front end 2002 or the knowledge processing unit 207 of the back end 2005 according to the determination result.

  As a result of the determination in step S33, when an interactive action is provided from the agent 2001 to the front end 2002 as one of the environments, the front end 2002 executes an interaction with the user (step S34). That is, the output processing unit 208 of the front end 2002 converts an action (dialogue action) in a predetermined dialogue action format acquired from the providing unit 205 into a natural language and presents it to the user. Further, the input processing unit 201 of the front end 2002 acquires a response (answer) from the user in a natural language format, converts it into a dialogue action format, and provides it to the agent 2001.

  On the other hand, as a result of the determination in step S33, when an interactive action is provided from the agent 2001 to the back end 2005 as another one of the environments, the back end 2005 executes knowledge processing (step S35). . That is, the knowledge processing unit 207 of the back end 2005 performs a search process of the knowledge base 206 based on the action (dialogue action) in the predetermined dialogue action format acquired from the providing unit 205, and the information based on the result of the search process is displayed as the dialogue action. And is provided to the agent 2001. The knowledge processing unit 207 executes a command that is an interactive action if the interactive action can be used as it is as a command for searching the knowledge base 206 (for example, a command in machine language). Convert to command and execute. In addition, the knowledge processing unit 207 provides information based on a return value from the knowledge base 206 as a result of the search processing to the agent 2001 in the form of a dialogue action.

  Subsequent to step S34 or step S35, the agent 2001 receives (acquires) observation (dialogue action) of a predetermined dialogue action form from the front end 2002 or the back end 2005 (step S36).

  Next, the agent 2001 determines whether or not the dialogue process should be ended based on the acquired dialogue action (step S37). For example, when the agent 2001 obtains a dialogue act “Acknowledge ()” from the front end 2002 as an observation, the agent 2001 determines that the dialogue processing should be terminated and terminates the dialogue processing.

  If it is not determined to end in step S37, the agent 2001 determines the current belief state, the determined action (action in the form of dialogue action type), and the observation obtained from the environment (observation in the form of dialogue action type). Based on the above, the belief state is updated (step S38). That is, the determination unit 200 updates the belief state information indicating the belief state according to the series of dialog action types in the acquired dialog action and the determined series of dialog action types.

  Subsequent to step S38, the agent 2001 returns to step S32 and determines an action corresponding to the dialogue action acquired in step S36.

  As described above, the dialogue processing includes the dialogue control processing including the following determination step and provision step by the agent 2001. In the interactive process, every time the interactive action is acquired from either the front end 2002 or the back end 2005 in the repeated step S36, the agent 2001 determines a providing interactive action in accordance with the acquired interactive action. (Step S32) is performed. Then, the agent 2001 performs a providing step (steps S33 to S35) for providing the providing interactive action determined in the determining step to either the front end 2002 or the back end 2005. In the determination step, the agent 2001 determines the dialogue action type in the providing dialogue action based on the dialogue action type sequence in the dialogue action sequentially obtained from the set of the front end 2002 and the back end 2005, based on the series of incidental information. And decide. The agent 2001 may update the belief state based on the series of interaction action types, and may determine the interaction action type as an action based on the belief state.

  As described above, the belief state in the agent 2001 is sequentially updated according to the dialogue action type of action and the dialogue action type of observation, and is a belief state reflecting all of the series of dialogue action types after the start of the dialogue processing. For this reason, if the agent 2001 determines the action based on the belief state, the agent 2001 can take an action that matches the progress of the dialogue and also an action that matches the user's intention.

  Hereinafter, an example of knowledge processing, which is a type of information processing in the back end 2005, will be described with reference to FIGS. 4A and 4B.

  FIG. 4A is a diagram illustrating an example of the knowledge base 206 referred to in the back end 2005.

  The knowledge base 206 is a database to be searched in knowledge processing by the knowledge processing unit 207 of the back end 2005. For example, knowledge for searching for diseases using symptoms as search keywords is structured and managed. The table in FIG. 4A shows an example of an image of knowledge content in the knowledge base 206. In the table of FIG. 4A, a cell that is the intersection of a row of a certain symptom and a column of a certain illness indicates “+” if the symptom is likely to appear in the case of the illness, and “−” if the symptom is difficult to appear. Is described. As a search keyword, for example, a search may be performed by distinguishing between a specified keyword that specifies that symptoms appear and an excluded keyword that specifies that symptoms do not appear. It is possible to narrow down corresponding diseases from a set of diseases by performing an AND search with a combination of one or more of a specified keyword for specifying symptoms and an excluded keyword. In the example of FIG. 4A, by performing an AND search using the symptom A as the designated keyword (+) and the symptom D as the exclusion keyword (−), the corresponding disease is only disease 1. FIG. 4A is merely an example for convenience of explanation. In the knowledge base 206, the search keyword (for example, symptom) and the search target name (for example, illness) can each be a large number (for example, tens or thousands or more).

  FIG. 4B is a conceptual diagram illustrating an example of knowledge processing by the knowledge processing unit 207 of the back end 2005 using a question tree.

  The example of FIG. 4B is an example of a query tree associated with performing an interrogation task of identifying a disease by a symptom question using the knowledge base 206 illustrated in the table of FIG. 4A, where: For convenience of explanation, a relatively simple question tree is shown. The root node of the question tree in FIG. 4B represents the first question, and the question as to whether or not symptom A has occurred is expressed as “symptom A”. The leaf node indicates the disease identified by the question. Each intermediate node located between the root and the leaf represents the second and subsequent questions. The arrow (edge) between the nodes indicates whether the answer to the question is “+” (an answer indicating that the symptom of the question has developed) or “−” (an answer indicating that the symptom of the question has not occurred) The entire question tree is a binary tree. In the root node, all disease sets are associated with each other, and the corresponding disease set is sequentially divided by the question to narrow down the diseases. When there are no more questions that divide the disease set, the disease set (one or more diseases) remaining at that time is the conclusion (result of the search process). In this question tree, the efficiency of narrowing down candidates varies depending on the order of questions. In order to reduce the worst-case question count as much as possible, it is preferable to select a question whose ratio of “+” and “−” in the remaining set of diseases is as close as possible to 1: 1.

  The operation of the back end 2005 as the knowledge processing unit 207 acquires a dialogue act indicating a command from the agent 2001, performs knowledge processing (search processing, etc.) based on the command, and performs a dialogue act indicating a return value for the command. Return to (provide). More specifically, the knowledge processing unit 207 accumulates incidental information (keyword indicating a symptom) incidental to the dialogue action type in the acquired dialogue action as a search keyword, and searches the knowledge base 206 according to the search keyword list. Perform the search process to be performed. As a result of the search processing, the knowledge processing unit 207 determines whether or not the search target disease is sufficiently narrowed down. When the search target disease is sufficiently narrowed down, the knowledge processing unit 207 concludes that the search target information (disease) is narrowed down as a conclusion. Return to 2001. When it is determined that the narrowing down is insufficient, the knowledge processing unit 207 proposes (provides) to the agent 2001 search keywords that are effective for narrowing down the search target. Note that the process of specifying a disease by the back end 2005 shown here is merely an example of information processing by the back end 2005. The information processing performed in the back end 2005 is performed, for example, by performing an Internet search (searching for information on the Internet) using a keyword, and until the search result is equal to or less than a predetermined number (for example, one or several) The next effective search keyword may be proposed (provided), and the process may be such that information as a search result is provided when the number is less than a predetermined number. In this case, the back end 2005 can be configured to include a wrapper to an existing Internet search program.

  Hereinafter, a dialogue sequence by the dialogue system 10 will be described. FIG. 5 is a sequence diagram showing an example of a dialogue sequence by the dialogue system 10.

  In FIG. 5, the processing section (active section) in each of the front end 2002, the agent 2001, and the back end 2005 (each object) is represented by a vertically long rectangle, and the vertical broken line from the object to the bottom represents the passage of time. Among the arrows (solid arrows), the solid line arrow represents a call message (that is, the provision of a dialogue action from the agent 2001), and the broken line arrow represents a procedure result notification message (that is, the acquisition of the dialogue action by the agent 2001). ing. In FIG. 5, the contents (a list of search keywords) of the stack 50 managed in the back end 2005 are appended. A positive (+) sign added to a keyword in the stack 50 means that the keyword is treated as a designated keyword when searching the knowledge base 206, and a negative (−) sign added to the keyword is knowledge This means that the keyword is treated as an excluded keyword when searching the base 206.

  Hereinafter, referring to FIG. 5, each step in the sequence according to the example of the dialogue related to the inquiry task will be described in order.

  In step S 101, the agent 2001 gives (provides) an interactive action “clear ()” representing a command for emptying the stack 50 storing the search keyword for the search processing of the knowledge base 206 to the back end 2005. . Note that it is determined that the back end 2005 is the providing destination if the dialogue action provided by the agent 2001 is lowercase, and the front end 2002 is the providing destination if the uppercase is uppercase.

  In step S102, the agent 2001 receives (acquires), from the back end 2005, an interactive action “pass ()” indicating that, for example, the processing according to the command is completed.

  In step S <b> 103, the agent 2001 provides the front end 2002 with an interactive action “What1st ()” representing a question “What have you done today?”. As a result, the front end 2002 asks the user, for example, a natural language “How did you do today?”. In this example, in response to this question, the user makes an answer indicating that his / her head is dizzy.

  In step S <b> 104, the agent 2001 acquires, from the front end 2002, an interactive action “Provide (headache)” obtained by converting the user's answer of “I'm dizzy” at the front end 2002.

  Following step S104, the agent 2001 determines the dialogue action type “Confirm” according to the dialogue action type “Provide” in the obtained dialogue action “Provide (headache)”, and the “Confirm” and the obtained dialogue action are determined. The provided dialogue action “Confirm (headache)” is generated by synthesizing the accompanying information “headache”.

  In step S <b> 105, the agent 2001 provides the front end 2002 with the dialogue action “Confirm (headache)”. Accordingly, the front end 2002 asks the user a confirmation question using the natural language “Is it a headache?” Converted based on the dialogue action “Confirm (headache)”. In this example, the user corrects the headache and answers “I am dizzy” to this question.

  In step S <b> 106, the agent 2001 acquires from the front end 2002 an interactive action “Correct” obtained by converting the user's answer “that is dizziness” (that is, an answer that corrects the confirmation content) by the front end 2002.

  Subsequent to step S106, the agent 2001 determines the dialogue action type “push” according to the dialogue action type “Correct” in the obtained dialogue action “Correct”, the “push” and the obtained dialogue action. The provided dialogue action “push” is generated by synthesizing “vertigo” which is the incidental information in.

  In step S <b> 107, the agent 2001 provides the back end 2005 with the dialogue action “push”. In response to a command represented by the dialogue action “push” (for example, a command for setting a search keyword), the stack 50 of the back end 2005 stores a designated keyword indicating “vertigo” in the accompanying information.

  In step S <b> 108, the agent 2001 acquires, from the back end 2005, for example, an interactive action “pass ()” indicating that the process according to the command is completed.

  Subsequent to step S108, the agent 2001 determines the dialogue action type “WhatElse” according to the dialogue action type “pass” in the acquired dialogue action “pass ()”, and generates the dialogue action “WhatElse ()”.

  In step S109, the agent 2001 provides the front end 2002 with the interactive action “WhatElse ()”. Note that, in the agent 2001, the belief state referred to when determining the dialogue action type is changed depending on the series of dialogue action types in the sequentially obtained dialogue actions. Even different providing interactive actions can be determined and provided. The front end 2002 asks the user further questions using the natural language “Do you care about others?” Converted based on the interactive action “WhatElse ()”. In this example, in response to this question, the user answers “I also palpitate” indicating further symptoms.

  In step S <b> 110, the agent 2001 acquires from the front end 2002 an interactive action “Provide” obtained by converting the user's answer “I will also palpate” in the front end 2002.

  Subsequent to step S110, the agent 2001 determines the dialogue action type “push” according to the dialogue action type “Provide” in the obtained dialogue action “Provide”, the “push” and the obtained dialogue action. The provided dialogue action “push” is generated by synthesizing “palpit” which is incidental information in.

  In step S111, the agent 2001 provides the back end 2005 with the dialogue action “push”. In response to a command represented by the dialogue action “push” (for example, a command for setting a search keyword), the stack 50 of the back end 2005 further stores a designated keyword indicating “palpit” of the accompanying information. .

  In step S112, the agent 2001 acquires, from the back end 2005, for example, an interactive action “pass ()” indicating that the process according to the command is completed.

  Following step S112, the agent 2001 determines the dialogue action type “WhatElse” in accordance with the dialogue action type “pass” in the acquired dialogue action “pass ()”, and generates the dialogue action “WhatElse ()”.

  In step S113, the agent 2001 provides the front end 2002 with the interactive action “WhatElse ()”. The front end 2002 asks the user further questions using the natural language “Do you care about others?” Converted based on the interactive action “WhatElse ()”. In this example, the user answers “No” to this question indicating that there are no other symptoms.

  In step S <b> 114, the agent 2001 acquires from the front end 2002 an interactive action “Negative ()” obtained by converting the user's answer “No” in the front end 2002.

  Subsequent to step S114, the agent 2001 determines the dialogue action type “solve1st” according to the dialogue action type “Negative” in the acquired dialogue action “Negative ()”, and provides the providing dialogue action “solve1st ()”. Generate.

  In step S115, the agent 2001 provides the interactive action “solve1st ()” to the back end 2005. In response to a command (for example, a command for search execution) represented by the dialogue action “solve1st ()”, the back end 2005 performs a search process of the knowledge base 206. The back end 2005 determines whether or not the search processing result has been sufficiently narrowed down. For example, the back end 2005 determines whether or not the number of diseases as a result of the search process has been narrowed down to a predetermined number (for example, one, several, etc.) or less. When the search process result is not sufficiently narrowed down, the back end 2005 uses the knowledge base 206 to identify and propose an effective search keyword to narrow down the search process result. In this example, “heat generation” is proposed as a search keyword effective for narrowing down without narrowing down the search processing results to a predetermined number or less. At this time, in this example, a search keyword indicating “fever” in a state where the designated keyword or the excluded keyword is not determined is accumulated in the stack 50 of the back end 2005.

  In step S <b> 116, the agent 2001 acquires from the back end 2005 an interactive action “suggest (fever)” representing suggesting a question of a “fever” symptom.

  Following step S116, the agent 2001 determines the dialogue action type “Check” according to the dialogue action type “suggest” in the obtained dialogue action “suggest (fever)”, and the “Check” and the obtained dialogue action are determined. The provisional interactive action “Check (fever)” is generated by synthesizing “fever” which is incidental information in FIG.

  In step S117, the agent 2001 provides the front end 2002 with an interactive action “Check (heat generation)”. The front end 2002 asks the user about the symptom of fever by the natural language “Is there fever?” Converted based on the dialogue action “Check (fever)”. In this example, in response to this question, the user answers “No” indicating that no symptoms of fever have appeared.

  In step S <b> 118, the agent 2001 acquires from the front end 2002 an interactive action “Negative ()” obtained by converting the user's answer “No” in the front end 2002.

  Following step S118, the agent 2001 determines the dialogue action type “set_negative” in accordance with the dialogue action type “Negative” in the obtained dialogue action “Negative ()”, and provides the dialogue action “set_negative ()” for provision. Generate.

  In step S119, the agent 2001 provides the interactive action “set_negative ()” to the back end 2005. In response to a command represented by the interactive action “set_negative ()” (for example, a command for executing a search using a list of search keywords after setting the search keyword last stacked on the stack as an exclusion keyword) The knowledge base 206 is searched using “heat generation” in the stack 50 as an exclusion keyword. Then, the back end 2005 determines whether or not the result of the search process has been sufficiently narrowed down. In this example, the result of the search process is not yet narrowed down to a predetermined number or less, and “cough” is further proposed as a search keyword effective for narrowing down. At this time, in this example, a search keyword indicating “cough” is accumulated in the stack 50 of the back end 2005.

  In step S <b> 120, the agent 2001 obtains an interactive action “suggest” representing suggesting a question of a “cough” symptom from the back end 2005.

  Following step S120, the agent 2001 determines the dialogue action type “Check” according to the dialogue action type “suggest” in the obtained dialogue action “suggest”, and the “Check” and the obtained dialogue action are determined. The provided interactive action “Check (cough)” is generated by synthesizing “cough” which is incidental information.

  In step S121, the agent 2001 provides the front end 2002 with an interactive action “Check (cough)”. The front end 2002 asks the user about cough symptoms using the natural language “Do you have a cough?” Converted based on the interactive action “Check (cough)”. In this example, in response to this question, the user answers “I will come out” indicating that cough symptoms are present.

  In step S122, the agent 2001 acquires from the front end 2002 an interactive action “Positive ()” obtained by converting the user's answer “Exit” by the front end 2002.

  Subsequent to step S122, the agent 2001 determines the dialogue action type “set_positive” according to the dialogue action type “Positive” in the acquired dialogue action “Positive ()”, and provides the providing dialogue action “set_positive ()”. Generate.

  In step S123, the agent 2001 provides the back end 2005 with the dialogue action “set_positive ()”. In response to a command represented by the interactive action “set_positive ()” (for example, a command for executing a search using a list of search keywords after setting a search keyword last loaded on the stack as a specified keyword) The knowledge base 206 is searched using “cough” in the stack 50 as a designated keyword. Then, the back end 2005 determines whether or not the result of the search process has been sufficiently narrowed down. In this example, the search processing result is narrowed down to “disease C”, which is one less than a predetermined number.

  In step S <b> 124, the agent 2001 acquires a dialogue action “conclude (disease C)” representing “disease C” as a conclusion from the back end 2005.

  Following step S124, the agent 2001 determines the dialogue action type “Present” according to the dialogue action type “conclude” in the obtained dialogue action “conclude (disease C)”, and the “Present” and the obtained dialogue are obtained. The provided dialogue action “Present (disease C)” is generated by synthesizing “disease C” which is incidental information in the action.

  In step S125, the agent 2001 provides the front end 2002 with the dialogue action “Present (illness C)”. The front end 2002 provides a conclusion to the user by the natural language “Like Sick C” converted based on the interactive action “Present (Sick C)”. In this example, the user answers “Okay” to provide this conclusion.

  In step S <b> 126, the agent 2001 acquires from the front end 2002 an interactive action “Acknowledge ()” obtained by converting the user's answer “OK” by the front end 2002. As a result, the agent 2001 ends the dialogue processing, and the dialogue sequence is finished.

  In this way, the interaction between the front end 2002 and the back end 2005 is performed by the agent 2001.

  In the above-described example, the sequence of dialogue action that is information exchanged between the agent 2001 and the front end 2002 is the following dialogue action sequence 1.

  Dialogue action sequence 1 = [What1st (), Provide (headache), Confirm (headache), Correct (vertigo), WhatElse (), Provide (palpitations), WhatElse (), Negative (), Check (fever), Negative () , Check (cough), Positive (), Present (disease C), Acknowledge ()]

  Further, a sequence obtained by adding a dialogue act exchanged between the agent 2001 and the back end 2005 to the dialogue act sequence 1 is the following dialogue act sequence 2.

  Dialogue action sequence 2 = [clear (), pass (), What1st (), Provide (headache), Confirm (headache), Correct (vertigo), push (vertigo), pass (), WhatElse (), Provide (palpitations) , Push (palpitations), pass (), WhatElse (), Negative (), solve1st (), suggest (fever), Check (fever), Negative (), set_negative (), suggest (cough), Check (cough), Positive (), set_positive (), conclude (disease C), present (disease C), Acknowledge ()]

  In the dialogue action sequence 2, the odd-numbered dialogue act corresponds to the action of the agent 2001, and the even-numbered dialogue act corresponds to the observation obtained by the agent 2001 from the front end 2002 or the back end 2005 as the environment. Looking at the accompanying information in parentheses, there is a match between the observation and subsequent actions. That is, the agent 2001 (dialogue agent) determines the dialogue action type as “Confirm” for “Provide”, “push” for “Correct”, etc., but it merely mediates the incidental information. The operation of determining an action according to such observation is realized by the determination unit 200 (that is, the separation unit 202, the action determination unit 203, and the synthesis unit 204) of the agent 2001.

(Consideration on dialogue state)
Hereinafter, the reduction of the number of dialog states that transition in the dialog by the dialog system 10 described above will be discussed with reference to FIGS. 6A to 6D.

  First, using the dialogue action sequence 1 according to the example of FIG. 5 described above, a dialogue system in which only the front end associated with the user is used as an environment is assumed, and the front end and the agent (here, knowledge processing) in the assumed dialogue system are assumed. Consider an example of interaction with a system that includes the functions of

  FIG. 6A is a state transition diagram illustrating an example of a state transition created so as to include the dialogue action sequence 1. In the POMDP model, if the action a is taken when the state is s, the transition probability T (s ′ | s, a) when the state transitions to s ′ and the result state obtained by taking the action a The relationship between the behavior, the observation, and the state is expressed by the observation probability O (z | a, s ′) at which the observation z is obtained at s ′. Here, the above probabilities in the POMDP model are subjected to threshold processing, and the relationship between behavior, observation and state is represented by a state transition diagram. Each state transition diagram of FIGS. 6B to 6D described below is also the same as FIG. 6A. For example, an arrow line in the state transition diagram represents a route in which the state transition probability exceeds a certain threshold. Further, in FIGS. 6A to 6D, the names of the states are written in circles representing the respective nodes. An arrow representing an edge between nodes describes a dialogue act as an action from the agent to the environment. In addition, each node (state) describes an interactive action as an environment observation by an agent. As a result, the state transition diagram indicates that the state of the environment changes depending on the behavior, and observation is performed under the changed state.

  Here, the dialogue action sequence 1 is considered by dividing it into two parts. First, the preceding parts [What1st (), Provide (headache), Confirm (headache), Correct (vertigo), WhatElse (), Provide (palpitations), WhatElse (), Negative ()] are open questions. This is a part where user-centered dialogues are mainly conducted (open questions). This part is referred to as patient-centered interview in the interview task. On the other hand, the following parts [Check (fever), Negative (), Check (cough), Positive (), Present (disease C), Acknowledge ()] are mainly closed questions. This is the part where the system (dialogue system) is engaged. This part is called a doctor-centered interview in the interview task.

  With respect to the preceding part (user-centered dialogue part), the state (node) of this part is arranged in a straight line in FIG. 6A. The sequence of dialogue actions corresponding to this part is a variable length sequence that ends with the observation Negative (). In order to reduce the number of states of this part, it seems that “s1a” to “s3a” are degenerated to form a loop structure. Loops can handle variable length sequences.

  Further, in FIG. 6A, the state (node) of the subsequent part (system-based dialogue part) is arranged in a tree structure. In order to trace the dialogue part of the system subject of the dialogue action sequence 1, it is only necessary to follow the path of the state series [s6a, s9a, s13a] following “s4a”. The part of the tree structure as a whole corresponds to the question tree illustrated in FIG. 4B so that different conclusions can be reached according to the question and the answer. This system-centered dialogue part can increase with the size of the knowledge base, for example. That is, if there are 1000 conclusions, the number of states becomes 1000 or more. The query tree in FIG. 4B is a simple example and has the same length, but the path length of the system-based dialogue part is variable. It seems that this part should also have a loop structure.

  FIG. 6B is a state transition diagram showing an example of state transition having two loop structures based on the above consideration.

  6B is similar to FIG. 6A, the user-directed dialogue part of the dialogue action sequence 1 [What1st (), Provide (headache), Confirm (headache), Correct (vertigo), WhatElse (), Provide (palpitations), WhatElse (), Negative ()] and system-oriented dialogue parts [Check (fever), Negative (), Check (cough), Positive (), Present (disease C), Acknowledge ()]. However, in FIG. 6B, “headache”, “disease C”, and the like, which are supplementary information of the dialogue action, are generalized by “*”. Regardless of the generalized part of the incidental information, it is considered that the correspondence with the state does not change.

  When attention is paid to the state “s1b” in FIG. 6B, the next state is “s1b” or “s2b”. In other words, after the loop that goes out of “s1b” and returns to “s1b” is repeated zero or more times, the loop exits and goes to “s2b”. Different observations are written in “s1b” and “s2b”, and when a dialogue action “Negative ()” is observed with respect to the dialogue action “WhatElse ()” as an action, the process exits the loop. .

  When attention is paid to the state “s3b”, the next state is “s3b” or “s4b”. In other words, after the loop that goes out of “s3b” and returns to “s3b” is repeated zero or more times, the loop exits the loop and goes to “s4b”. Since “s3b” and “s4b” have the same observation (“Negative ()” and “Positive ()”) and the action “Check (*)” is also the same, the condition for exiting the loop Is not clear. In such a case, how many times the loop exits the loop is determined by the ratio of the transition probability that traverses the loop and the transition probability that exits the loop. However, considering an interrogation task in which an illness is identified through an interrogation, the termination is determined by the doctor according to the contents of the interrogation, not the number of times. Or it may be sufficient to change from “s3b” to “s5b” by the action “Present (*)” except for “s4b” (not shown). However, there is no information for action selection. Here, whether the state and observation are the same even if the policy obtained by reinforcement learning is used to determine whether to go around the loop by the action “Check (*)” or to exit the loop by the action “Present (*)” The same behavior is always chosen. In other words, you can only enter an infinite loop or pass through the loop without ever going through it. This occurs when only the front end associated with the user is used as an environment, and the state is degenerated as much as possible without using the incidental information of the dialogue action. With this, the task cannot be performed.

  Next, consider the dialogue system 10 using the front end 2002 and the back end 2005 as environments using the dialogue action sequence 2 according to the example of FIG. 5 described above.

  FIG. 6C is a state transition diagram illustrating an example of a state transition created corresponding to the dialogue action sequence 2. The state transition diagram of FIG. 6C has two loop structures and corresponds to two environments.

  The existence of two loops throughout is similar in FIGS. 6B and 6C. In FIG. 6C, regarding the arrangement of each node, the left side corresponds to the interactive action exchanged between the front end 2002 and the agent 2001, and the right side corresponds to the interactive action exchanged between the agent 2001 and the back end 2005. The nodes are arranged to do so.

  Here, like the above-described dialogue action sequence 1, the dialogue action sequence 2 is divided into two parts (parts) for consideration. In other words, the user-centered conversation parts [clear (), pass (), What1st (), Provide (headache), Confirm (headache), Correct (vertigo), push (vertigo), pass (), WhatElse (), Provide (powder), push (powder), pass (), WhatElse (), Negative ()] parts, and subsequent parts that are system-oriented dialog parts [solve1st (), suggest (fever), Check (fever), Negative (), set_negative (), suggest (cough), Check (cough), Positive (), set_positive (), conclude (disease C), Present (disease C), Acknowledge () Divide.

  A first loop exists in the user-centered dialogue part in FIG. 6C, and the end condition of this loop is given by observation from the front end 2002 (specifically, “Negative ()”). The end condition of the first loop is the same as the end condition of the first loop shown in FIG. 6B. On the other hand, a second loop exists in the system-based dialogue part in FIG. 6C, and the loop termination condition is given by observation from the back end 2005 (specifically, “conclude (*)”). In other words, the action to be taken in “s5c” is “Check (*)”, and there are three transition destinations. However, since the state is set for each next assumed observation, the observation probability is multiplied. And the probability of a state corresponding to the observation increases. In other words, the end condition of the second loop is different from that shown in FIG. 6B. Since the dialogue act given as an observation is reflected in the probability of the state and can be used as a criterion for action determination, acquisition of an appropriate policy can be achieved by reinforcement learning. It becomes possible.

  In the first loop in FIG. 6C, the state transition corresponding to the sequence [Confirm (headache), Correct (vertigo)] is omitted, and this will be discussed with reference to FIG. 6D.

  FIG. 6D is a state transition diagram illustrating a state transition in which the state of “s2c” in FIG. 6C is divided and expanded to cope with the uncertainty. In FIG. 6D, the nodes except for “s2d lev0”, “s2d lev1”, “s2d posi”, and “s2d corr” correspond to the nodes having the same name in FIG. 6C.

  The first extension is a state division according to the reliability of the language understanding result obtained from the input processing unit 201 of the front end 2002. Each node of “s2d lev0” and “s2d lev1” is separated by a magnitude comparison with a predetermined reliability threshold, and “s2d lev0” is greater than the threshold, and if it is less than the threshold “S2d lev1”. In this example, the number of divisions is two, but the number of divisions may be increased several times as necessary. For example, as a method of setting each threshold value in the case of dividing into four, a method of setting a threshold value so that a range of reliability can be divided into four equal parts, a method of obtaining a frequency distribution and setting a quartile as a threshold value Etc.

  The second extension is the state “s2d posi” corresponding to the case where the agent 2001 confirms the front end 2002 before mediating the information obtained from the front end 2002 to the back end 2005 to increase the certainty. And “s2d corr”.

  According to the [Provide, Confirm (headache), Correct, push (vertigo), pass (), WhatElse (), Provide (palpation), push (palpitation)] part of dialogue action sequence 2 Following the state transition, it is assumed that the state is “s2d lev1” when the reliability when “Provide (headache)” is observed is relatively low, and the action “Confirm (headache)” is decided Then, the state transitions to “s2d corr”, and “Correct” is observed. Next, the action “push (vertigo)” is determined, transition is made to “s3c”, and “pass ()” is observed. Next, when the action “WhatElse ()” is determined and “Provide” is observed with a relatively high degree of reliability, it is assumed that the user is in “s2d lev0”. Here, without confirming, the process transitions to “s3c” with the action “push”.

  In this manner, the state can be divided based on the reliability of observation from the front end 2002, and different actions can be associated with each divided state. In the example of FIG. 6D, when the reliability level (that is, the number of divisions at the threshold) is two and the reliability is high, confirmation is not performed, and when the reliability is low, confirmation is performed. It is easy to manually set a policy for determining an action according to such reliability if the level of reliability is two. In the case where the level of reliability is set in multiple stages, trial and error may be required to set whether to perform confirmation for each level.

  In the description of the example of FIG. 5, it has been described that the incidental information matches between the dialog action of observation and the dialog action of the subsequent action in the dialog action sequence 2, but in FIG. 6D, [s2d lev1 , S2d posi, s3c ", for example, state transition corresponding to the sequence [Provide (headache), Confirm (headache), Positive (), push (headache)] is also possible. As described above, since the agent 2001 holds the latest acquired one of the incidental information separated from the dialogue action type in the acquired dialog action, the agent 2001 needs the latest incidental information held. Accordingly, it is possible to cope with the state transition of FIG.

(Comparison experiment result)
Hereinafter, a comparative experiment for quantitatively showing the effect of the dialogue system 10 will be described.

  A dialogue simulation experiment as an inquiry task was conducted using the dialogue system 10. FIG. 7 is a diagram illustrating a result of a simulation experiment using the dialogue system 10.

  In this simulation experiment, a knowledge base 206 containing 39 kinds of symptoms and 854 illnesses was used. Moreover, the dialogue action type in the dialogue action contained in the observation 2003, the observation 2006, the action 2004, and the action 2007 (see FIG. 2) was used as the dialogue action type. The patient simulator as a user who interacts with the front end 2002 of the dialogue system 10 in the dialogue simulation experiment performs a response of a symptom associated with the specified disease (an answer corresponding to the dialogue action “Provide ()”) once. Thereafter, the question of the dialogue system 10 is answered.

  Other main conditions are as follows.

-The functional part of language understanding in the front end 2002 generates false symptoms at a rate of 25%.
The number of divisions according to the reliability is 4, and the three threshold values for division are 0.2, 0.5, and 0.8.
The back end 2005 (knowledge processing unit 207) searches for a symptom-related search keyword to identify one disease (disease candidate) or a keyword (symptom) for narrowing down the disease. When there are no more candidates, the disease identified by the search is output as a conclusion (provided to the agent 2001). In other cases, the back end 2005 proposes (provides to the agent 2001) effective keywords for narrowing down.
-If the number of illnesses obtained as a conclusion is within 3 and the illness specified for the patient simulator is included, the answer is correct.

  FIG. 7 summarizes the experimental results in three types of models, “No confirmation strategy”, “Hand-crafted”, and “With confirmation strategy (reinforcement learning)” under the above conditions. Yes. Here, “no confirmation strategy” is a model corresponding to the state transition diagram of FIG. 6C. In addition, “hand-crafted” does not check for two states with a reliability of 0.5 or higher, but checks for two states with a reliability of less than 0.5. This is a model corresponding to the state transition diagram of FIG. 6D. “With confirmation strategy (reinforcement learning)” is a model corresponding to the type of state transition diagram in FIG. 6D. In this model, the actions to be taken by the four states according to the reliability are confirmed. Whether to “push” afterwards or to “push” without confirming is to follow the policy obtained by reinforcement learning in the POMDP model. In an environment, the agent determines the action to be taken by observing the current state. In reinforcement learning, the agent learns the policy that can obtain the most reward from the environment by performing the action selected by the agent sequentially. To do. As for the reward in the reinforcement learning of the model with “confirmation strategy (reinforcement learning)”, the path through which the dialogue action appears in the training dialogue corpus is 5; “push (*)” from “s2d corr” is 100, “push (*)” from “s2d posi” is −1, and it is highly valuable if it is corrected by checking for error correction. Confirmation when there is no need is set to be low value.

  The average number of turns in FIG. 7 is the number of turns counted only for the interaction between the front end 2002 and the user (patient simulator). The average reward is also calculated under the same conditions for a model that does not use reinforcement learning, in addition to “with confirmation strategy (reinforcement learning)”.

  As a result of the experiment, as shown in FIG. 7, the correct answer rate is the highest with “confirmation strategy (reinforcement learning)”, the next highest with “hand-crafted”, and the lowest with “no confirmation strategy” Was confirmed. It was also confirmed that the average number of turns was the smallest with “confirmation strategy (reinforcement learning)”, the second smallest with “hand-crafted” and the most common “no confirmation strategy”. It was also confirmed that the average reward obtained was the highest for “with confirmation strategy (reinforcement learning)”, the next highest for “hand-crafted”, and the lowest for “no confirmation strategy”.

  The time taken for the learning result by the POMDP solver used for the reinforcement learning of the model with “confirmation strategy (reinforcement learning)” to converge is less than 1 second, but the model corresponding to the state transition diagram of FIG. 6A Then, the time required for the POMDP solver was 30 days or more.

(Other embodiments, etc.)
As described above, the first embodiment has been described as an example of the technique according to the present invention. However, it goes without saying that the above-described embodiment is merely an example, and various modifications, additions, omissions, and the like are possible.

  In the above-described embodiment, an example in which the back end 2005 performs the search processing of the knowledge base 206 has been described. However, the information processing in the back end 2005 may not be the search processing but may be any processing. The back end 2005 may communicate with external devices other than the knowledge base 206. Further, the knowledge base 206 does not necessarily need to exist outside the back end 2005 and may be included in the back end 2005.

  In the above-described embodiment, an example in which there is one front end 2002 and one back end 2005 has been described. However, each of the front end 2002 and the back end 2005 is not limited to one, and a plurality of front ends 2002 and back ends 2005 may be provided. Good.

  In the above-described embodiment, an example in which the incidental information is a keyword has been shown, but the incidental information may be information other than a character string. For example, the incidental information may be a mark, a figure, an image, audio data, or the like. Further, for example, the incidental information may be not only a keyword but also a set of a slot and a keyword representing a keyword type.

  In the above-described embodiment, the action determining unit 203 determines the dialogue action type as the action based on the POMDP model. However, the action determining unit 203 does not necessarily have to be based on the POMDP model. For example, the action determining unit 203 may determine a dialogue action type based on a predetermined policy and a series (history) of dialogue action types input in the past, regardless of the belief state. .

  In the above-described embodiment, the example in which the providing unit 205 determines whether the providing destination of the interactive action is the front end 2002 or the back end 2005 based on whether the first character in the interactive action is uppercase or lowercase. However, the value of the dialogue action type in the dialogue action (dialogue action type value) is predetermined as a type (first type) that is predetermined for the front end 2002 or predetermined for the back end 2005. Since it suffices to distinguish according to the type (second type), the distinction may be realized by a method other than the distinction based on whether it is uppercase or lowercase. The providing unit 205 may have a correspondence table in which the interactive action type value is associated with the first type or the second type (that is, the type for distinguishing the providing destination). For example, the providing unit 205 provides the destination of the interactive action based on whether the interactive action type in the providing interactive action determined by the determining unit 200 represents the first type or the second type of interactive action type value. Can be selected to provide an interactive action to the selected provider.

  In the above-described embodiment, the case where the interactive system 10 is used for an inquiry has been described as an example. However, the above-described dialogue action type is not limited to an inquiry, but is common to tasks that specify information through dialogue.

  Therefore, the dialogue system 10 is not limited to the example of inquiry, and can be applied to the performance of a task for specifying information through dialogue with the user in various fields. As a specific example, the dialogue system 10 is a video content in fields such as recipe search in the field of cooking, travel plan search in the field of tourism, real estate property search in the field of real estate sales, television broadcasting, and the like. It can be applied to search. In this case, in the dialog system 10 (see FIG. 2), for example, the knowledge base 206 holds information useful for a predetermined field (application field of the dialog system 10, etc.), and the front end 2002 Regarding the field, information expressed in a natural language may be exchanged with the user. Note that the dialogue processing performed by the dialogue system 10 is the same as the processing illustrated in FIG.

  When the dialog system 10 is used for recipe search, the knowledge base 206 holds information related to, for example, ingredients, seasonings, cooking, cooking methods, and the like. What is necessary is just to acquire the information regarding foodstuff etc. by the dialog with a user. Further, when the dialog system 10 is used for travel plan search, the knowledge base 206 holds information on destinations, number of people, number of nights, budget, etc., and the front end 2002, for example, travel plans desired by the user. Information on the destination and the number of people may be acquired through dialogue with the user. Further, when the dialogue system 10 is used for real estate property search, the knowledge base 206 holds information related to, for example, a residential area, floor plan, building age, distance from the nearest station, rent, etc., and the front end 2002 Then, what is necessary is just to acquire the information regarding the residence area, floor plan, etc. of the real estate property which a user desires by dialog with a user. Further, when the interactive system 10 is used for video content search, the knowledge base 206 holds, for example, information on genre, performers, broadcast stations, broadcast date and time, and the front end 2002 displays, for example, a user's view. What is necessary is just to acquire the information regarding the genre of a program, a performer, etc. by dialogue with a user. In any task, the type, order of acquisition, and total number of information acquired from the user through dialogue with the user do not need to be predetermined constant values, and the information desired by the user is sufficiently narrowed down. It may be determined dynamically so as to satisfy the condition of being specified.

  In each of the above examples, the separation unit 202 in the agent 2001 separates the interactive action acquired from either the front end 2002 or the back end 2005 into the interactive action type and the incidental information, but specifies the information as described above. In the task, since the dialogue action type is common regardless of the field, the value of the dialogue action type can be the same as the example of the inquiry.

  Note that a dialogue action type may be added as appropriate according to a specific application field. For example, in the recipe search, in addition to the interactive action type exemplified in the embodiment, information on a recommended recipe (recommended recipe) in response to a request for information from the user is presented to the agent 2001 from the front end 2002. Dialogue action type “RequestRecommend” for recommendation information request, dialogue action type “seek_recommend” for recommendation information search from agent 2001 to back end 2005, dialogue action type “inform_recommend” for providing recommendation information from back end 2005 to agent 2001, agent Recommendation information presentation “PresentRecommend” from 2001 to the front end 2002, recommendation information, and the like may be added. In this way, information can be efficiently identified according to the field.

  In addition, some or all of the components (agent 2001, front end 2002, back end 2005, etc.) in the above-described embodiment are configured by one system LSI (Large Scale Integration). It is good. The system LSI is an ultra-multifunctional LSI manufactured by integrating a plurality of components on a single chip. Specifically, the system LSI is a computer system including a microprocessor, a ROM, a RAM, and the like. . A computer program is recorded in the RAM. The system LSI achieves its functions by the microprocessor operating according to the computer program. In addition, each part of the constituent elements constituting each of the above devices may be individually made into one chip, or may be made into one chip so as to include a part or all of them. Although the system LSI is used here, it may be called IC, LSI, super LSI, or ultra LSI depending on the degree of integration. Further, the method of circuit integration is not limited to LSI's, and implementation using dedicated circuitry or general purpose processors is also possible. An FPGA (Field Programmable Gate Array) that can be programmed after manufacturing the LSI, or a reconfigurable processor that can reconfigure the connection and setting of circuit cells inside the LSI may be used. Furthermore, if integrated circuit technology comes out to replace LSI's as a result of the advancement of semiconductor technology or a derivative other technology, it is naturally also possible to carry out function block integration using this technology. Biotechnology can be applied as a possibility.

  In addition, some or all of the components in the above embodiments may be configured from an IC card that can be attached to and detached from an apparatus such as a computer or a single module. The IC card or the module is a computer system including a microprocessor, a ROM, a RAM, and the like. The IC card or the module may include the super multifunctional LSI described above. The IC card or the module achieves its function by the microprocessor operating according to the computer program. This IC card or this module may have tamper resistance.

  Moreover, as one aspect of the present invention, for example, an interactive method including all or part of the procedures shown in FIGS. The dialog method used in the dialog system 10 includes, for example, a determination step in which the agent 2001 determines a dialog action for provision according to the dialog action acquired from one of the front end 2002 and the back end 2005; A providing step of providing the providing interactive action determined in the determining step to one of the front end 2002 and the back end 2005, wherein the determining step sets the interactive action type in the providing interactive action to the front end 2002 and It is determined based on a series of interactive action types, not based on a series of incidental information, in the interactive actions sequentially obtained from the set of backends 2005. For example, the determining step sequentially acquires dialogue actions from a set of the front end 2002 and the back end 2005, and separates the incidental information and the dialogue action type in the acquired dialogue actions, and the separation sub-step separates Action decision sub-step of determining the dialogue action type as an action by using the observed dialogue action type as an observation and referring to the belief state information as a reference for deciding the action corresponding to the observation based on the observation And a synthesizing substep for generating a providing dialogue action by synthesizing the dialogue action type determined in the action determining substep and the incidental information separated in the separating substep, and separating in the separating substep Of the dialogue action type determined in the action determination sub-step Including an update step of updating the belief state information in response to the actions and a column. Further, the present invention may be a computer program (for example, a program for performing a dialog control process including the determining step and the providing step) that realizes the dialog method, or may be a digital signal composed of the computer program. Good. In one embodiment of the present invention, the computer program or the digital signal can be read by a computer, such as a flexible disk, a hard disk, a CD-ROM, an MO, a DVD, a DVD-ROM, a DVD-RAM, and a BD. (Blu-ray (registered trademark) Disc), recorded on a semiconductor memory, or the like. The digital signal may be recorded on these recording media. Further, as one aspect of the present invention, the computer program or the digital signal may be transmitted via an electric communication line, a wireless or wired communication line, a network typified by the Internet, data broadcasting, or the like. Further, an aspect of the present invention may be a computer system including a microprocessor and a memory, the memory recording the computer program, and the microprocessor operating according to the computer program. . In addition, the program or the digital signal is recorded on the recording medium and transferred, or the program or the digital signal is transferred via the network or the like and executed by another independent computer system. You may do that.

  In addition to the above-described service mode (see FIG. 1), the technology related to the dialogue system 10 described in the above embodiment can be realized, for example, in the following types of cloud services. However, the application of the technology shown in the above embodiment is not limited to the type of cloud service described here.

(Service type 1: In-house data center type cloud service)
FIG. 8 is a diagram showing an overview of services provided by the interactive system in service type 1 (in-house data center type cloud service). In this type, the service provider 1120 acquires information input by the user 1010 from the device 1101 and provides the service to the user 1010 via the device 1101. In this type, the service provider 1120 has a function of a data center operating company. That is, the service provider 1120 has a cloud server that manages big data. In this category, there is no data center operating company. In this type, the service provider 1120 has a cloud server 1203 as a data center, and manages an operating system (OS) 1202 and application programs (applications) 1201. The service provider 1120 provides services by communicating with the device 1101 through the cloud server 1203 using the OS 1202 and the application 1201.

(Service type 2: Cloud service using IaaS)
FIG. 9 is a diagram illustrating an overview of services provided by the interactive system in service type 2 (IaaS-based cloud service). Here, IaaS is an abbreviation for infrastructure as a service, and is a cloud service provision model that provides a base for constructing and operating a computer system as a service via the Internet.

  In this type, the data center operating company 1110 operates and manages the data center (cloud server) 1203. The service provider 1120 manages the OS 1202 and the application 1201. The service provider 1120 provides a service using the OS 1202 and the application 1201 managed by the service provider 1120.

(Service type 3: Cloud service using PaaS)
FIG. 10 is a diagram showing an overall image of services provided by the interactive system in service type 3 (PaaS use type cloud service). Here, PaaS is an abbreviation for Platform as a Service, and is a cloud service provision model that provides a platform serving as a foundation for constructing and operating software as a service via the Internet.

  In this type, the data center operating company 1110 manages the OS 1202 and operates and manages the data center (cloud server) 1203. The service provider 1120 manages the application 1201. The service provider 1120 provides a service using the OS 1202 managed by the data center operating company 1110 and the application 1201 managed by the service provider 1120.

(Service type 4: Cloud service using SaaS)
FIG. 11 is a diagram illustrating an overview of services provided by the interactive system in service type 4 (SaaS-based cloud service). Here, SaaS is an abbreviation for software as a service. The SaaS-based cloud service is, for example, an application provided by a platform provider who owns a data center (cloud server), and a user such as a company or individual who does not have a data center (cloud server) uses the Internet. This is a cloud service provision model that has functions that can be used via other networks.

  In this type, the data center operating company 1110 manages the application 1201, manages the OS 1202, and operates and manages the data center (cloud server) 1203. The service provider 1120 provides a service using the OS 1202 and the application 1201 managed by the data center operating company 1110.

  As described above, in any cloud service type, the service provider 1120 provides a service. Further, for example, the service provider or the data center operating company may develop an OS, an application, a big data database, or the like, or may allow a third party to develop it.

  Note that the conversation system 10 may be configured by a single device 1101 (for example, a smartphone, a PC, or the like) illustrated in FIG. 1 without depending on the type of the cloud service described above. For example, the knowledge base 206 may be stored outside the device 1101. The dialogue system 10 may be configured by the knowledge base 206 and the device 1101 arranged on a network.

  In addition, embodiments realized by arbitrarily combining the constituent elements and functions described in the above embodiments are also included in the scope of the present invention.

  The present invention can be used as a dialog system that provides information through a dialog such as an inquiry.

DESCRIPTION OF SYMBOLS 10 Dialog system 200 Determination part 201 Input processing part 202 Separation part 203 Behavior determination part 204 Composition part 205 Provision part 206 Knowledge base 207 Knowledge processing part 208 Output processing part 1010 User 1100 Group 1101 Equipment 1102 Home gateway 1110 Data center management company 1111 Cloud Server 1120 Service provider 1121 Server 1201 Application program 1202 Operating system (OS)
1203 Cloud server 2001 Agent 2002 Front end 2003, 2006 Observation 2004, 2007 Action 2005 Back end

Claims (12)

An agent device that transmits and receives a dialogue action, which is information of a predetermined dialogue action type configured by a dialogue action type and incidental information, with each of a front end that performs user interface processing and a back end that performs information processing. ,
A determination unit that determines a providing dialogue action according to a dialogue action acquired from one of the front end and the back end;
A providing unit that provides the providing interactive action determined by the determining unit to one of the front end and the back end;
The determination unit, based on a series of dialogue action types, not based on a series of incidental information in a dialogue action sequentially obtained from the set of the front end and the back end, the dialogue action type in the providing dialogue action, Determine agent device. The agent device according to claim 1, wherein the determination unit determines a dialog action type in the providing dialog action based on a series of dialog action types determined in the past. The determination unit
An action determination unit that determines a dialogue action type as an action based on a series of dialogue action types as an observation;
A separating unit that sequentially acquires dialogue actions from the set of the front end and the back end, and sequentially inputs a series of dialogue action types as observations to the action determining unit by separating incidental information from the acquired dialogue actions;
A synthesizing unit that generates a dialogue act for provision by synthesizing the dialogue act type determined by the action deciding unit and the incidental information separated by the separating unit;
The agent device according to claim 2, wherein the action determination unit performs the determination of the dialogue action type as an action based on a series of dialog action types as an action determined in the past. The agent device according to claim 3, wherein the behavior determination unit determines a behavior based on a POMDP (Partially Observable Markov Decision Process) model. The dialogue action type as the observation represents one of a plurality of predetermined dialogue action type values,
The dialogue action type as the action is a plurality of predetermined dialogues that are classified into a first kind of dialogue action type value for the front end and a second kind of dialogue action type value for the back end. Represents one of the action type values,
The providing unit includes the front end and the back end based on whether the interactive action type in the providing interactive action determined by the determining unit represents a first or second interactive action type value. The agent device according to claim 3 or 4, wherein a provision destination of the dialogue action is selected from among the two, and the dialogue action is provided to the selected provision destination. The agent device according to claim 1, wherein the incidental information represents a keyword. The agent device according to any one of claims 1 to 6,
The front end for exchanging information expressed in natural language with a user;
An interactive system comprising the back end for providing information based on a knowledge base search. The incidental information represents keywords that can be used for searching the knowledge base,
The back end performs a search process of the knowledge base using a keyword represented by incidental information in a dialog action acquired from the agent device, and provides the agent device with a dialog action generated based on a result of the search process. Item 8. The dialogue system according to Item 7. The back end determines whether or not the result of the knowledge base search process has been sufficiently narrowed down. When it is determined that the result has been sufficiently narrowed down, the back end provides the agent device with an interactive action indicating the result, The dialog system according to claim 8, wherein if it is determined that the search result has not been narrowed down, a dialog act representing a keyword effective for narrowing down the result of the search process is provided to the agent device. A front end for exchanging information expressed in natural language with users;
A backend that provides information based on knowledge-based searches;
A dialogue method used in a dialogue system including each of the front end and the back end, and an agent device that exchanges a dialogue act that is information of a predetermined dialogue act type including a dialogue act type and incidental information. And
A determination step in which the agent device determines a providing interactive action according to an interactive action acquired from one of the front end and the back end;
A providing step in which the agent device provides the providing interactive action determined in the determining step to one of the front end and the back end;
In the determining step, a dialogue action type in the providing dialogue action is determined based on a dialogue action type sequence not based on a series of incidental information in a dialogue action sequentially obtained from the set of the front end and the back end. How to interact. The determining step includes
Separation sub-step of sequentially obtaining dialogue actions from the set of the front end and the back end, and separating incidental information and dialogue action types in the obtained dialogue actions;
The dialogue action type separated in the separation substep is set as an observation, and based on the observation, the belief state information as a reference for determining the action corresponding to the observation is referred to, and the dialogue action type as the action is determined. An action decision sub-step to decide;
A synthesizing substep for generating a providing dialogue act by synthesizing the dialogue act type determined in the action determining substep and the incidental information separated in the separating substep;
An update step of updating the belief state information according to an observation that is a series of dialogue action types separated in the separation substep and an action that is a series of dialogue action types determined in the action determination substep. The interaction method according to claim 10. A dialogue control process for exchanging a dialogue action, which is information of a predetermined dialogue action format composed of each of a front end for performing user interface processing and a back end for information processing, and a dialogue action type and incidental information, A program for causing a computer to execute,
The dialogue control process includes:
A determining step of determining a providing interactive action according to an interactive action obtained from one of the front end and the back end;
Providing the interactive action determined in the determining step to one of the front end and the back end; and
In the determining step, a dialogue action type in the providing dialogue action is determined based on a dialogue action type sequence not based on a series of incidental information in a dialogue action sequentially obtained from the set of the front end and the back end. Program.

Images