JP6679360B2 - Information providing apparatus and information providing method - Google Patents

Description

  The present invention relates to an information providing device and an information providing method.

  BACKGROUND ART Conventionally, there is known a technique of searching or generating information related to the input information based on an analysis result of the input information and outputting the searched or generated information as a response. As an example of such a technique, the words, sentences, and contexts included in the input text are converted into a multidimensional vector and analyzed, and based on the analysis result, text similar to the input text or the input text is input. There is a known natural language processing technique in which a text following a text is analogized and the analogy result is output.

JP, 2015-28625, A

"FastSLAM: A Factored Solution to the Simultaneous Localization and Mapping Problem", Michael Montemerlo, Sebastian Thrun, Daphne Koller, Ben Wegbreit, ", [online], [March 9, 2016 search], internet <http: // ai.stanford.edu/~koller/Papers/Montemerlo+al:AAAI02.pdf>

  However, when the above-mentioned conventional technique is applied as an aid to a conference, the conference may not be able to proceed efficiently.

  For example, a technique is conceivable in which a utterance corresponding to a utterance at a conference or the like is output from a robot or the like by using the above-described conventional technique. However, in the above-mentioned conventional technique, the information similar to the input text, the text following the input text, and the like that the user can predict are merely output. Therefore, for example, in the above-described related art, there is a possibility that the conference cannot be efficiently advanced.

  The present application has been made in view of the above, and an object thereof is to efficiently advance a conference.

  An information providing apparatus according to the present application includes an acquisition unit that acquires the content of statements of a plurality of users in a conference, a projection unit that projects the content of statements of the users acquired by the acquisition unit onto a vector space, and the vector space. A specifying unit that specifies the direction in which the conference is guided based on the position of the history of the comment contents projected above and the position of the new comment contents projected on the vector space, and the specifying of the meeting. And an output unit that outputs a response for guiding in the specified direction.

  According to one aspect of the embodiment, the conference can be efficiently advanced.

FIG. 1 is a diagram illustrating an example of an information providing device according to an embodiment. FIG. 2 is a diagram illustrating an example of a functional configuration of the information providing apparatus according to the embodiment. FIG. 3 is a diagram illustrating an example of processing in which the information providing apparatus according to the embodiment specifies a direction in which a conference is guided. FIG. 4 is a diagram illustrating an example of a route by which the information providing apparatus according to the embodiment guides a conference. FIG. 5 is a diagram illustrating an example of deep reinforcement learning executed by the information providing apparatus according to the embodiment. FIG. 6 is a flowchart illustrating the flow of learning processing executed by the information providing apparatus according to the embodiment. FIG. 7 is a diagram illustrating an example of the hardware configuration.

  Hereinafter, modes (hereinafter, referred to as “embodiments”) for carrying out the information providing apparatus and the information providing method according to the present application will be described in detail with reference to the drawings. The information providing apparatus and the information providing method according to the present application are not limited by this embodiment. Further, in the following embodiments, the same parts are designated by the same reference numerals, and duplicated description will be omitted.

[1. Example of information providing device]
First, an example of the response process executed by the information providing apparatus 10 will be described with reference to FIG. FIG. 1 is a diagram illustrating an example of an information providing device according to an embodiment. For example, in the example shown in FIG. 1, an example of the response process is as follows. When a plurality of users are having a meeting such as brainstorming, the user's remark is acquired as input information, and based on the acquired input information. , Learning process that learns collective intelligence of users who participated in the conference by learning individual opinions participating in the conference and direction of topics in the conference by deep reinforcement learning (DQN: Deep Q-Network). , The content of the speech at the conference is mapped on the vector space, and it is explained by using the SLAM (Simultaneous Localization and Mapping) technology used in automatic driving etc., and the guidance process for guiding the direction of the topic at the conference. To do.

  More specifically, in the following description, as a learning process, the collective intelligence of the users who participated in the conference is learned, and a response for controlling the current situation in the conference is output based on the learning result. An example of processing will be described. In addition, in the following description, an example of a process of avoiding a discussion conducted in the past in the conference or guiding the contents of the conference to a topic or the like that the conference was excited in the past will be described as the guidance process.

  The information providing device 10 shown in FIG. 1 is realized by an information processing device such as a server device. The information providing apparatus 10 may be realized by a single information processing apparatus, for example, a plurality of information processing apparatuses existing on the cloud network may be realized in cooperation with each other. Such an information providing device 10 converts a user's statement into text data and analyzes the text data by natural language processing. Then, the information providing apparatus 10 generates a comment that supports the meeting and the thinking of the user based on the analysis result, and outputs the generated comment.

[1-1. Example of response processing]
Here, in the prior art, a text similar to the input text or a text following the input text is analogized by using a distributed expression represented by a multidimensional word vector forming the input text. A technique of assisting the progress of the conference by outputting a utterance corresponding to the utterance of the conference or the like from a robot or the like using such a conventional technique is conceivable. However, in the above-mentioned conventional technique, the information similar to the input text, the text following the input text, and the like that the user can predict are merely output. Therefore, for example, in the above-described conventional technology, the content of the conference may be stagnant or the content of the conference may be diverged, and the conference may not be able to proceed efficiently. Therefore, the information providing apparatus 10 executes, as response processing, a learning processing for learning collective intelligence and a guidance processing for guiding the content of the conference. In the following description, the learning process and the guidance process will be described separately, but in reality, the learning process and the guidance process will be simultaneously executed in parallel by the information providing device 10.

[1-1-1. Learning process]
First, an example of the learning process executed by the information providing device 10 will be described. First, the information providing apparatus 10 acquires the content of a user's statement in a conference. Subsequently, the information providing apparatus 10 determines a response to the input utterance content, which guides the content of other subsequent utterances so as to approach the purpose of the conference. More specifically, the information providing apparatus 10 determines a response in which the utterance subsequent to the response output for the acquired utterance content approaches the purpose of the conference. For example, the information providing apparatus 10 determines a response by using a learner that has learned the response to the utterance by deep reinforcement learning so that the utterance following a certain utterance content approaches the purpose of the conference. Then, the information providing apparatus 10 outputs the determined response as a response to the message.

  Here, deep reinforcement learning is a learning method that combines deep learning such as so-called deep learning with reinforcement learning that observes the current situation based on the previous output result and determines the action to be taken according to the observation result. is there. More specifically, the information providing apparatus 10 holds a DNN (Deep Neural Network), which is a neural network having an input layer, a plurality of intermediate layers, and an output layer, as a learning device, and inputs input information from the input layer. The response is determined based on the information output from the output layer when input. Further, the information providing apparatus 10 acquires the environment after outputting the determined response, that is, the user's evaluation and reaction to the response, the content of the meeting, and the like, acquires the acquired evaluation and the like as a reward, and acquires the acquired reward. The learner is corrected (re-learned) based on That is, the information providing apparatus 10 sets the statement of the user in the conference as the state S in the deep reinforcement learning, and determines the policy a based on the state S and the policy π. Then, the information providing apparatus 10 calculates the state value function by using the statement of the user described later as the result of the policy π in deep reinforcement learning (that is, the reward R), and updates the policy π based on the calculated state value function. I do.

  For example, the information providing apparatus 10 newly obtains the content of the user's statement in response to the output response. Then, the information providing apparatus 10 calculates a value for evaluating the response, that is, a state-value function based on the newly acquired content of the user's statement, and newly acquires the value based on the calculated state-value function. It determines the response to the user's comments.

  Here, the information providing apparatus 10 sets the value of the reward R so that the speech of the user approaches the purpose of the conference. For example, the information providing apparatus 10 increases the value of the reward R as the amount of blurring between the previously acquired statement content and the newly acquired statement content decreases. Here, the value of the reward R is, for example, based on a sentence analysis technique such as morphological analysis, the content of the previous statement is compared with the content of the new statement, and the difference between the content of the previous statement and the content of the new statement is calculated. Is realized as the value of R increases, the value of the reward R decreases.

  Further, the value of the reward R may be determined, for example, based on the evaluation value input by the users who participate in the conference after the statement. For example, the information providing apparatus 10 receives, after outputting the response, an evaluation as to whether or not the content of the meeting is close to the purpose from the users who participate in the meeting. Such an evaluation is acquired by, for example, an input device such as a slider provided to each user, a BMI (Brain-machine Interface), or the like. Then, the information providing apparatus 10 sets the value of the reward R based on the acquired evaluation, and corrects the learning device based on the set value of the reward R.

[1-1-2. Induction processing)
Here, even if a response in which the user's remark simply approaches the purpose of the conference is output, it may not be possible to appropriately guide the conference. For example, in a meeting or the like, there may be a case where the user wants to smoothly advance the meeting by causing each user to come up with an idea different from the previous one, or by leading to a topic or the like that has been excited in the past.

  Therefore, the information providing apparatus 10 executes the following guidance process. For example, the information providing apparatus 10 acquires the content of remarks made by a plurality of users in a conference. Then, the information providing apparatus 10 projects the acquired user's comment contents on the vector space. For example, the information providing apparatus 10 uses the W2V (Word 2 Vector) or the like to convert the meaning or concept of a word or a sentence into a multidimensional amount (distributed expression) by using an arbitrary technique. The meaning it has is projected onto the vector space. That is, the information providing apparatus 10 projects the utterance content so that the utterance content of the same topic falls within a predetermined range in the vector space. Then, the information providing apparatus 10 determines the direction in which the conference is guided based on the position of the history of the utterance content projected on the vector space and the position of the new utterance content projected on the vector space.

  That is, the information providing apparatus 10 replaces the relationship between the content of the past meeting and the content of the current meeting with the position on the vector space by replacing the meaning of the utterance content with the position on the vector space. Then, the information providing apparatus 10 uses a technique such as SLAM used for automatic driving or the like to specify the direction in which the conference is guided in the vector space, and outputs a response to guide the conference in the specified direction. That is, the information providing apparatus 10 projects the meaning of the utterance content on the vector space to generate an idea map in which the meaning of the utterance content and the closeness of the idea are mapped on the distance space. Then, the information providing apparatus 10 projects the latest utterance content or a combination of vectors projecting the utterance content up to that point on the vector space to specify the position indicating the current state of the conference in the idea map. , To enable a bird's eye view of the position of the center of the discussion from the overall discussion.

  After that, the information providing apparatus 10 generates and outputs a response that guides the conference in the specified direction. For example, in the learning process described above, the information providing apparatus 10 updates the learning device so that the user's utterance content goes in the specified direction on the vector space, thereby giving a response that guides the conference in the specified direction. Output. More specifically, the information providing apparatus 10 projects the new utterance content of the user following the response on the vector space, and the position where the new utterance content is projected is specified more than the position where the previous utterance content is projected. If the direction is approaching, the value of the reward R is updated to a larger value and the learning device is corrected.

  Here, the information providing apparatus 10 may specify an arbitrary direction as the direction for guiding the conference. For example, the information providing apparatus 10 identifies the direction in which the user's utterance content when the conference was excited in the past was projected, and outputs a response for guiding the conference in the identified direction. As a result, the information providing device 10 can prevent the conference from being stagnant and smoothly advance the conference. Further, in another example, the information providing apparatus 10 identifies the position where the utterance content in the past conference is projected, that is, the direction away from the position where the history of the utterance content is projected, and guides the conference in the identified direction. Output the response. As a result, the information providing apparatus 10 can prevent a situation where the conference is boring around the same topic, etc., and advance the conference in a direction according to the purpose.

  Regarding the process of identifying the direction in which the conference is guided, arbitrary settings and strategies can be applied based on the purpose of the conference, the current state of the conference, the content of the conference in the past, and the like. For example, the information providing apparatus 10 considers an area on the vector space where the user's utterance content is projected when the conference is not exciting as an obstacle and uses the SLAM technology to avoid the obstacle. The above route may be specified and the conference may be guided along the specified route.

  In addition, the information providing apparatus 10 processes a high-dimensional vector space at high speed, and thus a technique such as MDS (Multi Dimensional Scaling) (multi-dimensional distance scale) that compresses the number of dimensions to maintain the distance between points. May be used to compress the dimension of the vector space, and then the process of identifying the guiding direction may be executed.

[1-2. Example of response processing]
Next, an example of the response process executed by the information providing apparatus 10 will be described with reference to FIG. For example, in the example shown in FIG. 1, the information providing apparatus 10 receives, as an input, the statement A or the statement B of the user in the conference (step S1). More specifically, the information providing apparatus 10 converts the utterance A uttered by the user into text data, and acquires the converted text data as input information.

  In such a case, the information providing device 10 executes a response process (step S2). First, the information providing apparatus 10 projects the utterance content on the vector space (step S3). Then, the information providing apparatus 10 identifies the positional relationship between the position of the past message content and the position of the new message content in the vector space (step S4). Then, the information providing device 10 specifies the direction in which the conference is to be guided by using a technique such as SLAM based on the specified positional relationship (step S5).

  Subsequently, the information providing apparatus 10 uses the learning device that has performed deep reinforcement learning so as to approach the purpose of the conference, and generates a response to the utterance content (step S6). Then, the information providing device 10 outputs the generated response (step S7). For example, the information providing apparatus 10 guides the conference by causing the robot or the like to read the generated response as the statement C.

  Further, the information providing apparatus 10 receives the user's evaluation of the statement C (step S8). For example, a user who participates in a meeting inputs an evaluation that the statement C is useful when, for example, the statement C feels that the meeting approaches the purpose. Then, the information providing apparatus 10 performs deep reinforcement learning by using the input evaluation as a reward for the current policy (step S9). That is, the information providing apparatus 10 modifies the learning device so that a better reward can be obtained.

  In addition, in the above-mentioned example, an example of the process of outputting the execution result of the learning process and the guidance process as the response is described as the response process. However, the embodiment is not limited to this. For example, the information providing device 10 may output the response obtained by executing only the above-described guidance process, or may not output the response generated by using the learner obtained as a result of deep reinforcement learning without performing the guidance process. You may output.

[2. Configuration of information providing device]
Next, the configuration of the information providing apparatus 10 that executes the learning process and the guidance process shown in FIG. 1 will be described with reference to FIG. FIG. 2 is a diagram illustrating an example of a functional configuration of the information providing apparatus according to the embodiment. As shown in FIG. 2, the information providing device 10 is connected to the input device 30 and the output device 31. The information providing device 10 also includes a communication unit 11, a storage unit 12, and a control unit 16.

  The communication unit 11 is realized by, for example, a NIC (Network Interface Card) or the like. The communication unit 11 is connected to an input device 30 such as a microphone and a keyboard and an output device 31 such as a monitor, a printer, and a robot capable of producing a voice, and transmits and receives various information.

  The storage unit 12 is realized by, for example, a semiconductor memory device such as a RAM (Random Access Memory) or a flash memory (Flash Memory), or a storage device such as a hard disk or an optical disk. The storage unit 12 also has a model database 13 and a vector space database 14.

  In the model database 13, the model learned by the learning process, that is, the data of the learning device that has learned the response that makes the content of the speech of the subsequent user closer to the purpose of the conference by the deep reinforcement learning is registered. . For example, in the model database 13, the connection relationship between neurons included in the learning device, the connection coefficient, and the like are registered.

  In the vector space database 14, a vector space in which the user's utterance content in the conference is projected is registered. For example, in the vector space database 14, a history of user's utterance contents converted into a plurality of dimensions using W2V technology or the like is registered. Since the vector (distributed expression) indicating each utterance content is generated based on the relationship between each utterance content, the direction and distance of each utterance content indicate the meaning, concept, and co-occurrence of each utterance content. It corresponds to the degree of similarity.

  The control unit 16 is a controller, and for example, a processor such as a CPU (Central Processing Unit) or an MPU (Micro Processing Unit) causes various programs stored in a storage device inside the information providing apparatus 10 to be a RAM or the like. Is implemented as a work area. The control unit 16 is a controller, and may be realized by an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

  As shown in FIG. 2, the control unit 16 includes an acquisition unit 17, a projection unit 18, a position identification unit 19, a guidance direction identification unit 20, a response determination unit 21, a response output unit 22, an evaluation acquisition unit 23, and a model update unit. With 24. It should be noted that each of the units 17 to 20 has a functional configuration for implementing the above-described guidance process, and each of the units 21 to 24 has a functional configuration for implementing the learning process. Therefore, the information providing apparatus 10 may be realized, for example, by the guidance apparatus including the units 17 to 20 and the learning apparatus including the units 21 to 24 operating in cooperation with each other.

[2-1. Configuration example for implementing guidance processing]
The acquisition unit 17 acquires the content of a user's statement in a conference. For example, the acquisition unit 17 acquires a user's statement in the conference from the input device 30 realized by a microphone, a keyboard, or the like. In such a case, the acquisition unit 17 converts the received statement of the user into text data.

  The projection unit 18 projects the acquired speech content of the user on the vector space. For example, the projection unit 18 extracts a word group included in the text data by using a technique such as morphological analysis, and calculates the concept or meaning of the extracted word group, that is, the concept or meaning of the utterance content in multiple dimensions. By converting it into a quantity, the content of the speech is projected on the vector space. Then, the projection unit 18 registers the converted speech content in the vector space database 14.

  The position specifying unit 19 specifies the relationship between the position of the history of the utterance content projected on the vector space and the position of the new utterance content projected on the vector space. For example, the position identifying unit 19 refers to the vector space database 14 and identifies the positional relationship between the position indicating the current state of the conference and the position on the vector space where the past utterance content is projected. For example, the position identifying unit 19 indicates the position in the vector space where the last output utterance content (latest utterance content) is projected, the total vector of the utterance contents projected in the conference, and the like as the current conference state. As the position on the vector space to be shown, the relationship with the position on the vector space where the past utterance content is projected is specified.

  The guidance direction specifying unit 20 specifies the direction in which the conference is guided, based on the positional relationship specified by the position specifying unit 19. For example, the guiding direction identification unit 20 sets the position on the vector space where the last output utterance content is projected as the current position, and the position on the vector space where the past utterance content is projected as the past position, SLAM technology is used to identify the direction in which to guide the conference. For example, the guidance direction identification unit 20 identifies, as the direction in which the conference is guided, a direction away from the position where the history of the message content is projected, or a direction in which the message content is projected when the conference is excited.

  For example, FIG. 3 is a diagram illustrating an example of processing in which the information providing apparatus according to the embodiment specifies a direction in which a conference is guided. For example, as shown in FIG. 3A, the projection unit 18 projects the user's utterance content in the conference on the vector space, thereby projecting the meaning of the utterance content and the relationship between the concepts to the metric space. Generate an idea map. More specifically, as shown in FIG. 3B, the projection unit 18 projects the utterance content at a position relatively indicating the meaning or concept of the utterance content. Therefore, the guiding direction specifying unit 20 determines how to guide the conference by specifying how to move in the idea map shown in FIG. 3A using the SLAM technology. To do.

  For example, as shown in FIG. 3C, the guiding direction identification unit 20 identifies an area where the user's utterance content in a past conference is projected as an obstacle. Then, when guiding the conference in the direction of a topic different from the past topic in the conference, the guidance direction specifying unit 20 avoids obstacles as shown in FIG. The direction leading the conference is specified by specifying the above route. That is, the guidance direction identification unit 20 guides the conference to the purpose along a route different from the contents of the past conference while creating an idea map in which the contents of the statement of the user in the conference are projected.

For example, FIG. 4 is a diagram illustrating an example of a route by which the information providing apparatus according to the embodiment guides a conference. For example, when the guidance direction identifying unit 20 acquires the user's statement S ₁ as shown in (A) of FIG. 4, each user is directed to the purpose of the conference as shown in (B) of FIG. is performed remarks S ₂ to S _t, so is not performed remarks z ₁ to z _t away from desired conference, specifying a direction to induce speech of each user. Further, the guidance direction specifying section 20, speech theta ₁ away from the target of the meeting, when the θ _{2, u} 1 ~u _t like is performed, the direction in which speech _s 2 ~s _t of each user is performed Is the direction in which the speech of each user is guided.

Here, an example of a method in which the guidance direction identification unit 20 determines the direction in which the conference is guided by using the SLAM technology will be described. For example, the guidance direction identification unit 20 sets the position in the vector space at the time “t” of the conference to be guided as “s ^t ”. For example, the guidance direction specifying unit 20 sets the sum of vectors obtained by projecting all the utterance contents in the conference on the vector space or the vector obtained by projecting the last utterance contents as "s ^t ". In addition, the guidance direction identifying unit 20 sets the direction in the vector space to be guided to “θ”. In addition, the guidance direction identification unit 20 sets the vector position of the message content mapped on the vector space as “z ^t ” (corresponding to the metric), and the synthetic vector of the message content of the user within the predetermined period in the conference is “u ^t ”. (Corresponding to control), and the moving distance in the vector space is “ ^nt ”. In such a case, the guide direction specifying unit 20 specifies the guide direction at the next time based on the SLAM technique using the following formula (1).

  Note that, for example, the guidance direction specifying unit 20 may clarify the position of each user who participates in the conference based on the position of the message content projected on the vector space. In addition, the guiding direction identification unit 20 may determine whether or not the user's utterance contents are simply saying different things by projecting the utterance contents on the vector space, It may be determined whether the discussion at the conference is looping. And the guidance direction specific | specification part 20 may specify the direction which guides a meeting based on a determination result.

[2-2. Configuration example for realizing learning process using DQN]
Returning to FIG. 2, the description will be continued. The response determination unit 21 determines the response to the acquired message content by using the learning device that has learned the response that makes the message content of the subsequent user closer to the purpose of the conference by deep reinforcement learning. For example, the response determination unit 21 acquires a learning device on which deep reinforcement learning has been performed from the model database 13, inputs the acquired user's utterance content to the learning device, and according to the output of the learning device, Determine the response to the user's statement. Then, the response output unit 22 outputs the response determined by the response determination unit 21 from the output device 31 such as a speaker.

  As will be described later, the learning device registered in the model database 13 is sequentially updated by the model updating unit 24 based on the evaluation acquired by the evaluation acquisition unit 23 each time a response is output. Therefore, when the new utterance content is acquired, the response determination unit 21 outputs the new response using the learning device updated based on the response corresponding to the previous utterance content. .

  The evaluation acquisition unit 23 acquires an evaluation for the response output by the response output unit 22 via the input device 30 having a predetermined interface such as a slider or BMI. The evaluation acquisition unit 23 may acquire the user's evaluation of the response, for example, by acquiring the content of the user's remark when the response is output and analyzing the acquired content of the remark. That is, the evaluation acquisition unit 23 acquires the state of the conference after the response is output.

  The evaluation acquisition unit 23 determines whether or not the meeting is close to the purpose based on the content of the user's statement acquired before the response and the content of the user's statement acquired after the response. You may acquire the evaluation according to a result. For example, the evaluation acquisition unit 23 compares the meaning of the remark content of the user acquired before the response with the meaning of the remark content of the user acquired after the response, and when the meaning of each remark content is not blurred. Alternatively, when the direction is specified by the guide direction specifying unit 20, that is, when the direction is the target direction of the conference, the evaluation that the response is useful may be acquired.

  The model updating unit 24 updates the learning device based on the state of the conference after the response is output. That is, the model updating unit 24 learns the response that makes the content of the speech of the subsequent user closer to the purpose of the conference based on the state of the conference after the response is output by the learner. Perform deep reinforcement learning. Specifically, the model update unit 24 performs deep reinforcement learning of the learning device based on the evaluation acquired by the evaluation acquisition unit 23 after the response output unit 22 outputs the previous response.

For example, FIG. 5 is a diagram illustrating an example of deep reinforcement learning executed by the information providing apparatus according to the embodiment. For example, in a conference, the user's utterance is “s”, the measure indicating what response the learner outputs to the utterance is “π”, and the user's utterance “s” is the measure. When the response output by “π” is described as “π (s)”, the message and response of the user in the conference can be schematically shown in the diagram shown in FIG. More specifically, when there is a statement “S ₀ ” of the user A, the information providing apparatus 10 outputs a response “π (s ₀ )”. When the statement “S ₁ ” of the user B is made in response to such a response “π (s ₀ )”, the information providing apparatus 10 outputs the response “π (s ₁ )” and the response “π (s ₁ When there is a statement “S ₂ ” of the user C for “s ₁ )”, the response “π (s ₂ )” is output.

When there is such a statement and response, the information providing apparatus 10 determines the state action value function “Q ^π (s, a) based on the evaluation with respect to the policy“ π ”, as shown in FIG. 5B. ) ”Is calculated. Then, the information providing apparatus 10 updates the policy “π” based on the state action value function “Q ^π (s, a)”. More specifically, the information providing apparatus 10 updates the value of the policy “π” so as to maximize the value of the state action value function “Q ^π (s, a)”. The information providing apparatus 10 realizes the deep reinforcement learning of the learner by repeatedly executing the output of such a response and the update of the policy based on the evaluation of the output.

More specifically, when the information providing apparatus 10 considers the value of the response “π (s)” to the policy is “a”, as shown in the following equation (2), the following equation (3) As shown in, the next policy “π ^* (s)” according to the reward is determined based on the maximization of the value of the state action value function. That is, the information providing apparatus 10 calculates the state action value function so that the value of the reward that will be obtained next is maximized according to the reward, and based on the calculated state action value function, the policy Update.

Here, the state action value function can be obtained by the following equation (4). Here, R (S ₀ , a ₀ , s ₁ ) is a variable indicating the reward for the previous response. The subscripts of "s" and "a" are values indicating the policy and the response time (order) to the policy. For example, “s _t ” is the value of “s” at time “t”, and “s _{t + 1} ” is the value of “s” at time “t + 1”.

  As described above, the information providing apparatus 10 updates the learning device based on the evaluation of the user with respect to the output response, and uses the updated learning device to guide the conference in the target direction. Is output. Therefore, the information providing apparatus 10 can efficiently proceed with the conference.

[3. Flow of response processing executed by the information providing apparatus 10]
Next, the flow of the response process executed by the information providing apparatus 10 will be described with reference to FIG. FIG. 6 is a flowchart illustrating the flow of response processing executed by the information providing apparatus according to the embodiment. The information providing apparatus 10 repeatedly executes the processes of steps S101 to S108 described below.

  First, the information providing apparatus 10 acquires the content of a user's statement (step S101). Then, the information providing apparatus 10 projects the acquired utterance content on the vector space (step S102), and specifies the relationship between the position of the history of the utterance content and the current position of the utterance content (step S103). Then, the information providing apparatus 10 uses the SLAM technology to identify the direction in which the conference is to be guided (step S104).

  Further, the information providing apparatus 10 determines the response to the message content by using the learning device that has learned the response for guiding the user's message content in the target direction of the conference by deep reinforcement learning (step S105). , And outputs the determined response (step S106). The information providing apparatus 10 also sets a reward based on the user's evaluation of the response (step S107), and executes deep reinforcement learning based on the set reward (step S108).

[4. Modification)
In the above, an example of the process executed by the information providing apparatus 10 has been described using the mode illustrated in FIG. 1. However, the embodiment is not limited to this. Hereinafter, variations of the extraction process executed by the information providing device 10 will be described.

[4-1. Execution mode of processing]
In the above-described example, the information providing apparatus 10 projects the utterance content on the vector space in order to efficiently advance the conference, specifies the direction in which the conference is guided based on the positional relationship between the utterance contents, and A response was generated and output using a learner that learned by deep reinforcement learning a response that induces a person's remark in the direction of the meeting. However, the embodiment is not limited to this.

  For example, the information providing device 10 may specify the direction in which the conference is to be guided as a result of the above-described guidance process, and output information indicating the specified direction. Moreover, the information providing apparatus 10 may generate and output a response by using the learning device described above. That is, the information providing device 10 may be a device that independently executes the learning process and the guidance process described above and outputs the execution result.

[4-2. Other]
Further, of the processes described in the above embodiment, all or part of the processes described as being automatically performed may be manually performed, or the processes described as manually performed may be performed. All or part of the process can be automatically performed by a known method. In addition, the processing procedures, specific names, and information including various data and parameters shown in the above documents and drawings can be arbitrarily changed unless otherwise specified. For example, the various information shown in each drawing is not limited to the information shown.

  Further, each component of each device shown in the drawings is functionally conceptual, and does not necessarily have to be physically configured as shown. That is, the specific form of distribution / integration of each device is not limited to that shown in the figure, and all or a part of the device may be functionally or physically distributed / arranged in arbitrary units according to various loads and usage conditions. It can be integrated and configured.

  Further, the respective embodiments described above can be appropriately combined within a range in which the processing content is not inconsistent.

[4-3. About hardware configuration]
Further, the information providing device 10 according to the above-described embodiment is realized by, for example, a computer 1000 configured as shown in FIG. 7. FIG. 7 is a diagram illustrating an example of the hardware configuration. The computer 1000 is connected to an output device 1010 and an input device 1020, and an arithmetic device 1030, a primary storage device 1040, a secondary storage device 1050, an output IF (Interface) 1060, an input IF 1070, and a network IF 1080 are connected by a bus 1090. Have.

  The arithmetic unit 1030 operates based on a program stored in the primary storage device 1040 or the secondary storage device 1050, a program read from the input device 1020, or the like, and executes various processes. The primary storage device 1040 is a memory device such as a RAM that temporarily stores data used by the arithmetic device 1030 for various calculations. The secondary storage device 1050 is a storage device in which data used by the arithmetic device 1030 for various calculations and various databases are registered, and is realized by a ROM (Read Only Memory), an HDD, a flash memory, or the like.

  The output IF 1060 is an interface for transmitting information to be output to an output device 1010 that outputs various kinds of information such as a monitor and a printer. For example, a USB (Universal Serial Bus) or a DVI (Digital Visual Interface), It is realized by a connector of a standard such as HDMI (registered trademark) (High Definition Multimedia Interface). The input IF 1070 is an interface for receiving information from various input devices 1020 such as a mouse, a keyboard, and a scanner, and is realized by, for example, USB.

  The input device 1020 is, for example, an optical recording medium such as a CD (Compact Disc), a DVD (Digital Versatile Disc), and a PD (Phase change rewritable Disk), a magneto-optical recording medium such as an MO (Magneto-Optical disk), and a tape. It may be a device that reads information from a medium, a magnetic recording medium, a semiconductor memory, or the like. The input device 1020 may also be an external storage medium such as a USB memory.

  The network IF 1080 receives data from another device via the network N and sends the data to the arithmetic device 1030, and also transmits the data generated by the arithmetic device 1030 via the network N to another device.

  The arithmetic device 1030 controls the output device 1010 and the input device 1020 via the output IF 1060 and the input IF 1070. For example, the arithmetic device 1030 loads a program from the input device 1020 or the secondary storage device 1050 onto the primary storage device 1040, and executes the loaded program.

  For example, when the computer 1000 functions as the information providing device 10, the arithmetic device 1030 of the computer 1000 implements the function of the control unit 16 by executing the program loaded on the primary storage device 1040.

[5. effect〕
As described above, the information providing apparatus 10 acquires the utterance content of a plurality of users in the conference and projects the acquired utterance content of the user on the vector space. Then, the information providing apparatus 10 specifies the direction in which the conference is guided, based on the position of the history of the utterance content projected on the vector space and the position of the new utterance content projected on the vector space, Outputs a response to guide the meeting in the specified direction. For this reason, the information providing apparatus 10 projects the relevance between the utterance contents up to that point and the current conference on the positional relationship in the vector space, and identifies the direction in which the conference is guided based on the positional relationship in the vector space. As a result, the conference can be guided in an appropriate direction, so that the conference can be efficiently advanced.

  Further, the information providing apparatus 10 uses the SLAM technology to identify the direction in which the conference is guided in the vector space. For example, the information providing apparatus 10 identifies, as the direction in which the conference is guided, the direction in which the user's utterance content is projected when the conference is excited or the direction away from the position in which the utterance content history is projected. For this reason, the information providing apparatus 10 can prevent a situation in which the conference is excited, the same topic is repeated, and the like, and the conference can be efficiently advanced.

  Further, the information providing apparatus 10 projects a plurality of utterance contents belonging to the same topic into a predetermined area in the vector space. Therefore, the information providing device 10 can appropriately specify the direction in which the conference is guided.

  As described above, some of the embodiments of the present application have been described in detail based on the drawings, but these are examples, and various modifications based on the knowledge of those skilled in the art, including the modes described in the section of the disclosure of the invention, It is possible to implement the present invention in other forms with improvements.

  Further, the above-mentioned "section (module, unit)" can be read as "means" or "circuit". For example, the control unit can be read as a control unit or a control circuit.

10 information providing device 11 communication unit 12 storage unit 13 model database 14 vector space database 16 control unit 17 acquisition unit 18 projection unit 19 position identification unit 20 guidance direction identification unit 21 response determination unit 22 response output unit 23 evaluation acquisition unit 24 model update Part 30 Input device 31 Output device

Claims (5)

An acquisition unit that acquires the content of statements of multiple users in a meeting,
By using a conversion method for converting a sentence having a similar concept into a similar vector, the word group included in the user's statement acquired by the acquisition unit is vectorized and registered in a predetermined storage device. A projection unit that projects the concept of remarks on a vector space,
Of the vectors registered in the storage device, from the position on the vector space where the concept of the past utterance is projected, to the position on the vector space where the concept of the new utterance that is the last output utterance is projected. Based on the similarity of the direction to the position and the direction to the position in the vector space where the concept of the predetermined utterance based on the utterances in the past meetings is projected , the new utterance deviates from the purpose of the meeting. If it is determined that the new utterance has deviated from the purpose of the conference, the predetermined utterance concept is projected from the position on the vector space where the new utterance concept is projected. A direction to the position in the vector space, a specifying unit that specifies the direction to guide the conference,
As a response for guiding the conference in the direction specified by the specifying unit , the user's utterance after outputting a response is a reward of a larger value as it approaches the direction in which the conference is guided, and, The smaller the difference between the position on the vector space onto which the concept of the past utterance is projected and the position on the vector space onto which the new utterance concept is projected after outputting the response, the smaller the reward is set. And an output unit that outputs a response generated by the model in which deep reinforcement learning is performed . The specifying unit uses a SLAM (Simultaneous Localization and Mapping) technique to project a concept that the user wants to make a new statement from a position in the vector space where the new concept of the statement is projected. The information providing apparatus according to claim 1, wherein a direction to a position in the vector space is specified. The specifying unit specifies a direction in which a concept of a predetermined utterance is projected as a utterance of the user when the conference is excited as a direction for guiding the conference. Alternatively, the information providing device described in 2. The said specific | specification part specifies the direction which deviates from the position where the concept of the speech acquired last time was projected as a direction which guides the said meeting. The claim | item 1 characterized by the above-mentioned. Information providing device. An information providing method executed by an information providing device, comprising:
An acquisition process for acquiring the content of remarks of multiple users in a meeting,
By using a conversion method for converting a sentence having a similar concept into a similar vector, the word group included in the user's statement acquired in the acquisition step is vectorized and registered in a predetermined storage device. A projection process for projecting the concept of speech on a vector space,
Of the vectors registered in the storage device, from the position on the vector space where the concept of the past utterance is projected, to the position on the vector space where the concept of the new utterance that is the last output utterance is projected. Based on the similarity of the direction to the position and the direction to the position in the vector space where the concept of the predetermined utterance based on the utterances in the past meetings is projected , the new utterance deviates from the purpose of the meeting. If it is determined that the new utterance has deviated from the purpose of the conference, the predetermined utterance concept is projected from the position on the vector space where the new utterance concept is projected. A step of identifying the direction to the position in the vector space as a direction guiding the conference,
As a response for guiding the conference in the direction specified by the specifying step , the user's utterance after outputting a response is a reward of a larger value as it approaches the direction in which the conference is guided, and, The smaller the difference between the position on the vector space onto which the concept of the past utterance is projected and the position on the vector space onto which the new utterance concept is projected after outputting the response, the smaller the reward is set. And an output step of outputting a response generated by the model for which deep reinforcement learning has been performed .

Images