Dialogue device, dialogue control method and dialogue program

Abstract

In the present invention, a speech information acquisition unit (11) acquires speech information for collaborative workers that are performing collaborative work to achieve a goal, through dialogue. A dialogue control unit (13): acquires information on progress leading to achievement of a goal through the performance of the collaborative work through dialogue by a plurality of workers; and, on the basis of the speech information acquired by the speech information acquisition unit (11), the information on progress, and subjective results information indicating interim results according to host devices of the collaborative work, generates a system dialogue using an extrapolative model. An output unit (14) outputs the system dialogue generated by the dialogue control unit (13) to the collaborative workers.

Description

[Technical Field]

[0001] The present invention relates to a dialogue device, a dialogue control method, and a dialogue program. [Background technology]

[0002] In a dialogue system, humans interact with a computer to gather various information and fulfill requests. There are also dialogue systems that not only accomplish specific tasks but also engage in everyday conversations. These dialogue systems enable humans to maintain mental stability, satisfy their desire for recognition, and build trusting relationships. As such, there are many different types of dialogue systems.

[0003] Current chat dialogue systems are mainly question-and-answer systems that select and generate system utterances based on information from the previous user utterance, without retaining state information built up from previous dialogues. Because question-and-answer chat dialogue systems have difficulty with dialogues that go beyond complex question-and-answer formats, they face the problem of low user satisfaction in chats that require complex exchanges. One way to solve this problem is to equip the system with information called a common foundation. A common foundation is information such as knowledge and beliefs shared between participants in a dialogue, also known as mutual beliefs.

[0004] Although common ground is an important concept in dialogue modeling, there are currently few studies that analyze the process by which common ground is constructed. For example, there is a study that attempts to model common ground construction by collecting and analyzing spoken dialogues in which two workers accomplish a task. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Takuma Udagawa and Akiko Aizawa, A natural language corpus of common grounding under continuous and partially-observable context, In Proc. AAAI, pp. 7120-7127, 2019. Summary of the Invention [Problem to be solved by the invention]

[0006] In previous studies using common platforms, the results of tasks performed by workers were considered to be the common platform, and the relationship between the dialogue and the common platform was analyzed. However, the process leading up to the task completion was not quantitatively recorded. This makes it difficult to grasp how the common platform was constructed through dialogue, making it difficult to build a dialogue system based on the common platform.

[0007] The present invention has been made in view of the above, and has as its object to realize an advanced conversation with a user. [Means for solving the problem]

[0008] In order to solve the above problems and achieve the objectives, The dialogue device has an estimation model that receives as input a dialogue context including utterances between collaborators at a specific dialogue stage when the collaborators work together to accomplish a task through dialogue, a dialogue intention of the dialogue between the collaborators, and subjective result information indicating an intermediate result of the collaboration by the device itself, and outputs the next work result of the collaborator and the next system utterance. The speech information acquisition unit acquires information through dialogue. specified To achieve the task specified Collaborate specified Collaborator's specified The dialogue control unit acquires utterance information. A predetermined dialogue context including utterances between the predetermined co-workers at a predetermined stage of the dialogue is acquired from the predetermined utterance information, The worker communicates with the specified By working together, specified Until the task is accomplished predetermined Obtaining information on the process the predetermined dialogue context, The aforementioned specified Speech information and The aforementioned predetermined Information on the process A dialogue intention of a dialogue between the predetermined collaborators is identified based on and the above specified Showing intermediate results of collaborative work using own device predeterminedThe system utterance is generated by using the estimation model based on the subjective result information. specified Output to collaborators. [Effects of the Invention]

[0009] According to the present invention, it is possible to realize a sophisticated conversation with a user. [Brief explanation of the drawings]

[0010] [Figure 1] FIG. 1 is a block diagram of a dialogue device according to an embodiment. [Figure 2] FIG. 2 shows an example of a dialogue registered in the collaboration corpus. [Figure 3] FIG. 3 shows an example of collaborative graphic placement work. [Figure 4] FIG. 4 is a diagram showing an example of output from the utterance information acquisition unit. [Figure 5] FIG. 5 is a block diagram of the learning function of the dialogue control unit using a collaborative work corpus. [Figure 6] FIG. 6 is a flowchart of the learning process of the estimation model by the dialogue device according to the embodiment. [Figure 7] FIG. 7 is a flowchart of a dialogue control process performed by the dialogue device according to the embodiment. [Figure 8] FIG. 8 is a diagram illustrating an example of a computer that executes a dialogue program for dialogue control processing. DETAILED DESCRIPTION OF THE INVENTION

[0011] An embodiment of a dialogue device, a dialogue control method, and a dialogue program disclosed herein will be described in detail below with reference to the accompanying drawings. Note that the dialogue device, the dialogue control method, and the dialogue program disclosed herein are not limited to the following embodiment.

[0012] [Embodiment] [Device configuration] FIG. 1 is a block diagram of a dialogue device according to an embodiment. As shown in FIG. 1, the dialogue device 1 has an utterance information acquisition unit 11, a dialogue intention management unit 12, a dialogue control unit 13, an output unit 14, and an information storage unit 15. The dialogue device 1 solves a given problem by performing an independent task jointly through a dialogue with a user who is a co-worker. A task performed jointly by the dialogue device 1 and the user through a dialogue is called a collaborative task. The dialogue device 1 is connected to an utterance text output device 2 and a work terminal device 3.

[0013] The speech text output device 2 is, for example, a device that recognizes a voice utterance input to a microphone, converts it into text, and outputs it to the dialogue device 1. The speech text output device 2 may also output text information of an utterance input by a user operating an input device such as a keyboard.

[0014] The work terminal device 3 is a terminal used by a plurality of other workers when the other workers jointly perform the collaborative work performed by the dialogue device 1 together with the user. The work terminal device 3 outputs to the dialogue device 1 information on the dialogue exchanged between the other plurality of workers when they perform the collaborative work and information on the progress of the work.

[0015] The information storage unit 15 is a storage device such as a hard disk that stores various information used in the dialogue. The information storage unit 15 holds a collaborative work corpus 51, dialogue intention information 52, subjective result information 53, and common base information 54.

[0016] The collaborative work corpus 51 is information that compiles sentences representing the dialogues of multiple workers when they independently solve a problem through dialogue and the work progress. In other words, the collaborative work corpus is information on the process of multiple workers working together to accomplish a problem through dialogue. The collaborative work corpus 51 links the work progress to specific sentences along with sentences representing the dialogues. In other words, the collaborative work corpus 51 indicates what work was performed when a specific dialogue took place. For example, if the work involved moving a shape on an xy coordinate plane, the information indicating what kind of work was performed is expressed as which shape was moved to which position on the xy coordinates.

[0017] The collaborative work corpus 51 acquires and stores information on dialogues exchanged between multiple other workers when they perform collaborative work and information on the progress of the work, which information is transmitted from the work terminal device 3. FIG. 2 is a diagram showing an example of dialogues registered in the collaborative work corpus. In this embodiment, the collaborative work corpus 51 will be described using an example in which two workers perform collaborative work. ID is an identification number for each utterance. S represents the worker. Utterance is information indicating the content of the utterance made by the worker.

[0018] The dialogue shown in Figure 2 specifically represents the dialogue collected when workers A and B performed a collaborative graphic placement task. A collaborative graphic placement task is a task in which two workers independently place a graphic while engaging in a dialogue. Although not shown in Figure 2, in the collaborative task corpus 51, as mentioned above, the progress of the task is linked to some of the sentences shown in Figure 2.

[0019] Fig. 3 is a diagram showing an example of collaborative graphic placement work. Screen 101 in Fig. 3 is the work screen of worker A. Screen 102 is the work screen of worker B. Screens 101 and 102 are displayed on the work terminal device 3 operated by worker A and the work terminal device 3 operated by worker B, respectively. Worker A does not check screen 102, and conversely, worker B does not check screen 101. Workers A and B work together to complete their own picture while imagining the state of each other's work through dialogue.

[0020] The graphic arrangement screens 111 and 112 are workspaces for graphic arrangement. The chat screens 121 and 122 are spaces where each worker's speech is displayed in text. Additionally, work start and end buttons are located at the top of the screen, and the remaining time for the work, which is set at a maximum of 10 minutes, is also displayed. Since workers A and B are each given the same set of graphics in a random arrangement, they discuss how to arrange them using the chat screens 121 and 122 and decide on a common arrangement. Here, the graphic arrangement screens 111 and 121 do not allow rotation, scaling, or deletion of graphics, but allow for planar movement using a mouse. The work terminal device 3 records the start and end times of dragging and dropping graphics and their respective coordinates as an operation log. The work terminal device 3 stores this operation log in the collaborative work corpus 51 as information on the progress of the work.

[0021] For example, worker A changes the placement position of each shape displayed on shape placement screen 111 based on a conversation with worker B displayed on chat screen 121, and determines the placement of each shape based on his or her own image of the shape. Worker B changes the placement position of each shape displayed on shape placement screen 112, and determines the placement of each shape based on his or her own image of the shape. Since workers A and B determine the placement of shapes based on their respective images, it is unlikely that they will end up with the same picture, but there is a high possibility that there will be matching parts in the picture where the shapes are placed. At this time, the shape placement during work is recorded, and by regarding part of the shape placement as the common ground, it is possible to quantitatively record the common ground.

[0022] In this embodiment, two types of shapes were prepared as the shapes to be placed: simple shapes, which are the simplest shapes, and building shapes, which are thought to be able to utilize prior knowledge about shapes. The simple shapes and building shapes each consist of 10 types of shapes. The number of shapes of each type was set to 5 or 7, and the shapes were set to random sizes and positions with overlaps to create the initial placement.

[0023] For example, as shown in Figure 2, when a collaborative graphic placement task progresses, workers A and B reach an agreement on the final layout image by U15. Then, workers A and B consult with each other to place the graphics from U16 onwards. Among these, utterances related to the construction of a common foundation can be observed, such as utterances confirming each other's placement, as in U21. Therefore, the dialogue device 1 uses the collaborative task corpus 51 to learn how to generate utterances and how to infer the results of independent work and the work of collaborators, thereby constructing a system that can accomplish a collaborative graphic placement task together with users who are collaborators through dialogue.

[0024] Returning to FIG. 1, the explanation will be continued. The dialogue intention information 52 is information that is the basis for determining how to proceed with the dialogue, and is information that is set according to the purpose of the dialogue. An initial value of the dialogue intention information 52 is registered in advance by the user. In this embodiment, the dialogue intention information 52 is managed as a list of natural sentences.

[0025] For example, in the case of a collaborative figure placement task, the initial value of the dialogue intention information 52 includes, based on the task conditions, the sentence "align your drawing with the other person's through dialogue." Furthermore, through dialogue, intentions such as "make a beautiful layout" or "make Pinocchio's face" are added and updated as dialogue intention information 52 by the dialogue intention management unit 12, which will be described later.

[0026] The subjective result information 53 is information on the results of work carried out mainly by the dialogue device 1. The subjective result information 53 is information indicating the intermediate results of a collaborative work by the device itself. For example, in the case of a collaborative graphic placement task, the subjective result information 53 is a picture in which graphics created by the dialogue device 1 are arranged. The subjective result information 53 reflects the most recent dialogue content and can be considered as the understanding result of the dialogue device 1. In other words, the subjective result information 53 is an estimation result by the dialogue control unit 13. For example, in the case of a collaborative graphic placement task, the subjective result information 53 stores the graphic type and coordinates as text or numerical values.

[0027] The common foundation information 54 is information indicating the common parts between the work result of the dialogue partner and the subjective result information 53, which is the work result of the dialogue device 1. For example, in a collaborative graphic placement task, the pictures created by the dialogue partner and the dialogue device 1 can be considered as work results that reflect how each party understands the content of the conversation. Therefore, by introducing a measure that quantitatively measures the common foundation as the common foundation information 54, it is possible to mechanically handle information such as the extent to which a common foundation has been established, or, for example, in the case of a collaborative graphic placement task, which shapes can be considered as a common foundation. In this embodiment, the distance between the vectors defined between any two shapes is used as a measure that quantitatively measures the common foundation.

[0028] In this embodiment, a vector v defined between figures i and j in the figure arrangement of worker A is A,ij and a vector v defined similarly for worker B. B,ij The difference between these distances is used as a measure of whether each figure has been established as a foundation. The sum of the distances between each figure is used as a measure of the extent to which a common foundation has been established for the entire picture. The lower this value, which is used as a measure of the extent to which a common foundation has been established for the entire picture, the more a common foundation has been established.

[0029] For example, the common foundation information 54 can be used to determine the degree to which a common foundation has been established with a collaborating partner. Therefore, if this common foundation information 54 exceeds a certain value, it is possible to add control to move the dialogue in the direction of terminating. Also, by presenting this common foundation information 54 to a collaborating partner, it becomes possible to share an understanding of the extent to which the work is in agreement.

[0030] The utterance information acquisition unit 11 receives user utterances expressed in text from the utterance text output device 2. That is, the utterance information acquisition unit 11 acquires utterance information of co-workers who work together to accomplish a task through dialogue. Next, the utterance information acquisition unit 11 performs language analysis on the received user utterances. Thereafter, the utterance information acquisition unit 11 outputs the analysis results to the dialogue intention management unit 12 and the dialogue control unit 13.

[0031] For example, the utterance information acquisition unit 11 performs morphological analysis, focus word extraction to extract keywords that represent topics, proper noun extraction, evaluative expression extraction, modality extraction to extract the presence or absence of negative expressions, and dialogue act estimation. FIG. 4 is a diagram showing an example output from the utterance information acquisition unit. In the output example 130, "line" represents an input sentence. "forms" and "poses" are the results of morphological analysis by the utterance information acquisition unit 11. "names" represents the result of proper noun extraction by the utterance information acquisition unit 11. "sems" represents the result of modality extraction by the utterance information acquisition unit 11. "evals" represents the result of evaluative expression extraction by the utterance information acquisition unit 11. "cents" represents the result of focus word extraction by the utterance information acquisition unit 11. "da" represents the result of dialogue act estimation by the utterance information acquisition unit 11.

[0032] The dialog intention management unit 12, for example, references the dialogue text in the collaborative work corpus 51 and identifies utterances related to "proposal" in the dialogue acts as candidates for extracting a dialogue intention. Here, the dialogue acts are obtained as information estimated using an estimator based on the system disclosed in, for example, "Toyomi Meguro, Ryuichiro Higashinaka, Koji Dosaka, and Yasuhiro Minami, Analysis of Listener Dialogues and Construction of a Dialogue Control Unit Based on the Analysis, Transactions of the Information Processing Society of Japan, Vol. 53, No. 12, pp. 2787-2801." The dialog intention management unit 12 then extracts from the identified utterances as candidates those whose degree of similarity with the user's utterance is equal to or greater than a threshold, based on the degree of string similarity using, for example, the Levenshtein distance. The threshold for the degree of similarity can be set to, for example, 0.8 when the Levenshtein distance is used. The dialog intention management unit 12 then updates the dialog intention information 52 by adding the extracted words or sentences to the dialog intention information 52.

[0033] In this way, the dialogue intention management unit 12 identifies the dialogue intention of the dialogue with the user who is a co-worker, based on the utterance information acquired by the utterance information acquisition unit 11 and information on the process of multiple workers working together through dialogue to achieve the task.

[0034] (Learning process for estimation model) The dialogue control unit 13 learns an estimation model to be used for estimation based on the collaborative work corpus 51. Figure 5 is a block diagram showing the learning function of the dialogue control unit using the collaborative work corpus. In this embodiment, the dialogue control unit 13 performs learning using multitask learning, which simultaneously learns about multiple different outputs in parallel.

[0035] 5, the dialogue control unit 13 has a language feature extractor 31, an image feature extractor 32, a feature extractor 33, a parameter update unit 34, and an estimation model 35. The dialogue control unit 13 receives as input dialogue intention information 52, information on the subjective work result, and information on the dialogue context. Here, the dialogue control unit 13 extracts the information on the subjective work result and the information on the dialogue context from a collaborative work corpus 51 and uses them as input.

[0036] For example, in the case of a collaborative work corpus 51 in which two workers collaborated, the dialogue control unit 13 selects one of the workers. Then, the dialogue control unit 13 selects a sentence uttered by the selected worker at a specific stage of the dialogue from the dialogue text. The dialogue control unit 13 then acquires the previous utterance of the other worker in response to the selected sentence as the dialogue context. The dialogue control unit 13 also acquires the work result of the selected worker linked to the end of the selected sentence or an earlier sentence as the subjective work result.

[0037] The language feature extractor 31 converts the input dialogue intention information 52 into a vector representation, and converts it into a format that can be processed by the estimation model 35. The language feature extractor 31 can be realized by converting a sentence into a vector using, for example, BERT (Bidirectional Encoder Representations from Transformers).

[0038] The image feature extractor 32 converts the input subjective result information 53 into a vector representation and converts it into a format that can be processed by the estimation model 35. The image feature extractor 32 can be realized, for example, by converting an image into a vector using ResNet.

[0039] The feature extractor 33 converts the input dialogue context into a vector representation, which is then converted into a format that can be processed by the estimation model 35. Here, the feature extractor 33 and the estimation model 35 can be trained together as a single deep learning model.

[0040] The estimation model 35 can be implemented using a deep learning framework such as PyTorch Lightning for multitask learning. The estimation model 35 preferably has output layers prepared according to multiple pieces of output information. For example, in this embodiment, the estimation model 35 has three output layers prepared: the next subjective task result, the next partner task result, and the next system utterance.

[0041] The parameter update unit 34 acquires estimation results, which are outputs of each output layer of the estimation model 35, from the estimation model 35. The parameter update unit 34 also analyzes the collaboration corpus 51 to acquire correct labels corresponding to the estimation results. Specifically, the parameter update unit 34 acquires correct labels for the next system utterance from sentences of a dialogue included in the collaboration corpus 51. The parameter update unit 34 also acquires correct labels for the next subjective work result and the next partner's work result using the work progress linked to the sentences included in the collaboration corpus 51. The parameter update unit 34 then calculates the error between the estimation results output from the estimation model 35 and the correct labels.

[0042] 5 as the estimation results output from the estimation model 35. Next, the parameter update unit 34 analyzes the collaboration corpus 51 to obtain a correct answer 211 for the next subjective work result, a correct answer 212 for the next other party's work result, and a correct answer 213 for the next system utterance as correct answer labels for the respective estimation results. Then, the parameter update unit 34 calculates the error between the next subjective work result 201 and the correct answer 211 for the next subjective work result, the error between the next other party's work result 202 and the correct answer 212 for the next other party's work result, and the error between the next system utterance 203 and the correct answer 213 for the next system utterance.

[0043] Next, the parameter update unit 34 adjusts and updates the parameters so that each error is minimized. Thereafter, the parameter update unit 34 feeds back information about the updated parameters to the estimation model 35.

[0044] The dialogue control unit 13 repeatedly updates the parameters of the estimation model 35 until a predetermined learning termination condition is met. The learning termination condition may be, for example, when a predetermined number of updates is exceeded, or when an error reaches a predetermined error threshold. By performing such multi-task learning, the dialogue control unit 13 can proceed with learning so that appropriate values are output in all output layers of the estimation model 35. In this way, the dialogue control unit 13 updates the estimation model 35 based on information on the process of multiple workers completing the collaborative work through dialogue and the subjective result information 53.

[0045] (Estimation process) The dialogue control unit 13 updates the subjective result information 53, estimates the other party's work result, and generates a system utterance based on the input user utterance. Specifically, the dialogue control unit 13 receives input of text information of the user utterance from the utterance text output device 2. The dialogue control unit 13 also acquires dialogue intention information 52 from the information storage unit 15. The dialogue control unit 13 also acquires subjective result information 53 from the information storage unit 15.

[0046] Then, based on the acquired user utterance, dialogue intention information 52, and subjective result information 53, the dialogue control unit 13 uses the trained estimation model it holds to estimate the next subjective work result, the next work result of the other party, and the next system utterance, which is an utterance from the device itself.

[0047] Then, the dialogue control unit 13 stores the estimated result of the next subjective task as subjective result information 53 in the information storage unit 15. For example, in the case of a collaborative graphic placement task, the subjective result information 53 is a pictorial representation, and the dialogue control unit 13 stores the graphic type and coordinates as text or numerical values. The dialogue control unit 13 also outputs the estimated result of the next system utterance, which is the next utterance from the user's own device, to the output unit 14.

[0048] Here, the matching portion between the estimated result of the next task of the other person and the subjective result information 53 can be regarded as common foundation information 54. In other words, if there is an estimation model that has the function of being able to appropriately imagine the other person's result, the dialogue device 1 and the user, who is the other person in the task, can have a dialogue based on a common foundation by generating system utterances based on the estimation model. In other words, a system that can accomplish a joint figure placement task together with the user can be realized.

[0049] Therefore, the dialogue control unit 13 extracts the common parts between the estimated result of the next subjective work result and the estimated result of the next partner's work result.The dialogue control unit 13 then updates the common base information 54 to the information of the extracted common parts.That is, the dialogue control unit 13 estimates the next subjective result information 53 and the next partner's work result using an estimation model based on the utterance information, information on the process of multiple workers working together through dialogue to achieve the task, and the subjective result information 53, and identifies a common base with the user who is a co-worker based on the estimation results.

[0050] For example, in the case of a collaborative figure placement task, the dialogue control unit 13 determines the vector v defined between figures i and j in the figure placement of worker A. A,ij and a vector v defined similarly for worker B.B,ij The dialogue control unit 13 then uses this distance as a measure of whether the figure has been built into a foundation, and uses the sum of these distances as common foundation information 54, which is a measure of how well a common foundation has been built for the entire picture.

[0051] The dialogue control unit 13 then refers to the common foundation information 54 to determine to what extent a common foundation has been established between the work partner and the work. For example, when the common foundation information 54 exceeds a predetermined value, the dialogue control unit 13 may determine that a common foundation sufficient to solve the problem has been established and perform control to bring the dialogue to an end. Specifically, the dialogue control unit 13 may generate a system utterance that brings the dialogue to an end. That is, the dialogue control unit 13 generates a system utterance based on the common foundation with the co-worker. The dialogue control unit 13 may also output the common foundation information 54 to the output unit 14.

[0052] In this way, the dialogue control unit 13 acquires information on the intermediate process of multiple workers working together to accomplish the task through dialogue, and generates system utterances using an estimation model based on the utterance information acquired by the utterance information acquisition unit 11, the information on the intermediate process of multiple workers working together to accomplish the task through dialogue, and subjective result information 53 indicating the intermediate results of the collaborative work by the own device. More specifically, the dialogue control unit 13 generates system utterances based on the dialogue intention information 52 indicating the dialogue intention identified by the dialogue intention management unit 12, the utterance information, and the subjective result information 53.

[0053] Returning to FIG. 1, the explanation continues. The output unit 14 receives input of the next system utterance from the dialogue control unit 13. Then, the output unit 14 outputs the acquired system utterance. Here, if there are multiple utterance candidates, the output unit 14 selects one utterance based on a predetermined priority order, such as dictionary order, and outputs the selected utterance as the system utterance. If there are multiple system utterances, the output unit 14 outputs them together as a single system utterance. Furthermore, the output unit 14 may impart character to the system utterance or perform processing to improve the fluency of the system utterance. For example, characterization may be achieved by converting all endings to "gozaru." Fluency improvement processing may be processing such as deleting unnatural or inappropriate expressions using a dictionary.

[0054] The output unit 14 may also receive input of the common foundation information 54 from the dialogue control unit 13. In this case, the output unit 14 outputs the acquired common foundation information 54 to the person with whom the dialogue is taking place. This allows the dialogue device 1 and the user who is the person with whom the dialogue is taking place to share a common understanding of how much of a common foundation has been established until the task is completed.

[0055] [Learning procedure for estimation model] 6 is a flowchart of the learning process procedure of the estimation model by the dialogue device according to the embodiment. Next, the flow of dialogue control process by the dialogue device 1 according to the present embodiment will be described with reference to FIG.

[0056] The dialogue control unit 13 acquires the dialogue intention information 52 from the information storage unit 15 (step S11).

[0057] Next, the dialogue control unit 13 acquires the subjective work result and the dialogue context from the collaborative work corpus 51 (step S12).

[0058] The language feature extractor 31 converts the acquired dialogue intention information 52 into a vector representation, thereby converting it into a format that can be processed by the estimation model 35. Furthermore, the image feature extractor 32 converts the acquired subjective work result into a vector representation, thereby converting it into a format that can be processed by the estimation model 35. Furthermore, the feature extractor 33 converts the acquired dialogue context into a vector representation, thereby converting it into a format that can be processed by the estimation model 35 (step S13).

[0059] Next, the dialogue control unit 13 inputs the dialogue intention information 52 converted into vector representation, the subjective work result, and the dialogue context to the estimation model 35 (step S14).

[0060] The parameter update unit 34 acquires the estimation results of the next subjective task result, the next partner task result, and the next system utterance, which are outputs of each output layer of the estimation model 35 (step S15).

[0061] Next, the parameter update unit 34 analyzes the collaboration corpus 51 to obtain a correct label corresponding to the estimation result (step S16).

[0062] Then, the parameter update unit 34 calculates the error between the estimation result output from the estimation model 35 and the correct label. After that, the parameter update unit 34 adjusts the parameters of the estimation model using the calculated error (step S17).

[0063] Next, the parameter update unit 34 feeds back information about the adjusted parameters to the estimation model 35, thereby updating the estimation model 35 (step S18).

[0064] Thereafter, the dialogue control unit 13 determines whether or not the learning termination condition has been reached (step S19). If the learning termination condition has not been reached (step S19: No), the dialogue control unit 13 returns to step S12. On the other hand, if the learning termination condition has been reached (step S19: Yes), the dialogue control unit 13 ends the learning process of the estimation model 35.

[0065] [Dialogue control processing procedure] 7 is a flowchart of the dialogue control process by the dialogue device according to the embodiment. Next, the flow of the dialogue control process by the dialogue device 1 according to the present embodiment will be described with reference to FIG.

[0066] The utterance information acquisition unit 11 and the dialogue control unit 13 receive input of information about a user utterance (step S21).

[0067] Next, the utterance information acquisition unit 11 performs language analysis on the acquired user utterance (step S22). After that, the utterance information acquisition unit 11 outputs the analysis result of the user utterance to the dialogue intention management unit 12.

[0068] The dialogue control unit 13 acquires the dialogue intention information 52 and the subjective result information 53 from the information storage unit 15 (step S23).

[0069] Next, the dialogue control unit 13 inputs the user utterance, dialogue intention information 52, and subjective result information 53 into the estimation model (step S24).

[0070] Next, the dialogue control unit 13 uses the output from the estimation model and the common base information 54 to acquire the estimation results of the next subjective work result, the next partner's work result, and the next system utterance (step S25).

[0071] Then, the dialogue control unit 13 updates the subjective result information 53 with the estimated next subjective work result (step S26).

[0072] Next, the dialogue control unit 13 updates the common infrastructure information 54 (step S27).

[0073] Next, the dialogue control unit 13 outputs the estimation result of the next system utterance to the output unit 14. The output unit 14 outputs the estimated system utterance to the terminal of the other party in the dialogue or the like (step S28).

[0074] The dialogue control unit 13 also determines whether the common infrastructure information 54 has exceeded a predetermined value for the first time (step S29). If the common infrastructure information 54 has exceeded the predetermined value for the first time (step S29: Yes), the dialogue control unit 13 adds dialogue termination control to lead the dialogue in the direction of terminating (step S30). Thereafter, the dialogue control process proceeds to step S31.

[0075] On the other hand, if the common base information 54 does not exceed the predetermined value or has previously exceeded the predetermined value (step S29: No), the dialogue control process proceeds to step S31.

[0076] Next, the dialogue control unit 13 determines whether the joint work has ended (step S31). If the joint work has not ended (step S31: No), the dialogue control process returns to step S21. On the other hand, if the joint work has ended (step S31: Yes), the dialogue control unit 13 ends the dialogue control process.

[0077] [Effects of dialogue control processing] As described above, the dialogue device 1 according to this embodiment grasps the dialogue intention, estimates and updates the next subjective work result from the subjective work result and the user's utterance, and also estimates the next system utterance to engage in dialogue. This makes it possible to accomplish a task through collaborative work using dialogue together with the user. That is, a common foundation is built with the user through dialogue, and an appropriate estimation model for dialogue based on that foundation can be constructed. Furthermore, for dialogue involving complex content, it is preferable to build up an understanding of that content. In this regard, the dialogue device 1 according to this embodiment can build up the subjective work result, which corresponds to its own understanding, while building a common foundation, thereby realizing a system capable of engaging in advanced dialogue with the user of various types, such as education, discussion, or negotiation.

[0078] Although the present embodiment has been described taking as an example a collaborative graphic placement task by two people, other processes may be used as long as they involve collaborative work through dialogue to solve a specific problem. For example, the interactive device 1 according to the present embodiment can also achieve similar effects in a process of determining the layout of furniture.

[0079] [System configuration, etc.] The components of each device shown in the figure are conceptual functional units and do not necessarily have to be physically configured as shown. In other words, the specific form of distribution and integration of each device is not limited to that shown, and all or part of the devices can be functionally or physically distributed or integrated in any unit depending on various loads, usage conditions, etc. Furthermore, all or any part of the processing functions performed by each device can be realized by a CPU (Central Processing Unit) and a program analyzed and executed by the CPU, or can be realized as hardware using wired logic.

[0080] Furthermore, among the processes described in the embodiments, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically using a known method.In addition, the information including the processing procedures, control procedures, specific names, various data and parameters shown in the above documents and drawings can be changed as desired unless otherwise specified.

[0081] [program] Moreover, the alert verification device 10 can be implemented by installing a program that realizes the functions of the dialogue device 1 described in the above embodiment into a desired information processing device (computer). For example, the above program, provided as package software or online software, can be executed by the information processing device to cause the information processing device to function as the dialogue device 1. The information processing device referred to here includes desktop and notebook personal computers. In addition, the information processing device also includes mobile communication terminals such as smartphones, mobile phones, and PHS (Personal Handyphone Systems), as well as PDAs (Personal Digital Assistants). The alert verification device 10 may also be implemented on a cloud server.

[0082] Fig. 8 is a diagram showing an example of a computer that executes an interaction program for interaction control processing. As shown in Fig. 8, a computer 1000 includes, for example, a memory 1010, a CPU 1020, a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. These components are connected by a bus 1080.

[0083] The memory 1010 includes a ROM (Read Only Memory) 1011 and a RAM (Random Access Memory) 1012. The ROM 1011 stores, for example, a boot program such as a BIOS (Basic Input Output System). The hard disk drive interface 1030 is connected to a hard disk drive 1090. The disk drive interface 1040 is connected to a disk drive 1100. A removable storage medium such as a magnetic disk or optical disk is inserted into the disk drive 1100. The serial port interface 1050 is connected to, for example, a mouse 1110 and a keyboard 1120. The video adapter 1060 is connected to, for example, a display 1130.

[0084] The hard disk drive 1090 stores, for example, an OS 1091, an application program 1092, a program module 1093, and program data 1094. That is, an alert verification program that defines each process of an interaction device 1 having the same functions as the interaction device 1 is implemented as a program module 1093 in which computer-executable code is written. The program module 1093 is stored, for example, in the hard disk drive 1090. For example, a program module 1093 for executing processes similar to those of the functional configuration of the interaction device 1 is stored in the hard disk drive 1090. The hard disk drive 1090 may be replaced by an SSD (Solid State Drive).

[0085] Furthermore, setting data used in the processing of the above-described embodiment is stored as program data 1094, for example, in the memory 1010 or the hard disk drive 1090. Then, the CPU 1020 reads the program module 1093 or the program data 1094 stored in the memory 1010 or the hard disk drive 1090 into the RAM 1012 as necessary, and executes the processing of the above-described embodiment.

[0086] The program module 1093 and program data 1094 are not limited to being stored in the hard disk drive 1090, but may also be stored in, for example, a removable storage medium and read by the CPU 1020 via the disk drive 1100 or the like. Alternatively, the program module 1093 and program data 1094 may be stored in another computer connected via a network (such as a LAN (Local Area Network) or WAN (Wide Area Network)). The program module 1093 and program data 1094 may then be read by the CPU 1020 from the other computer via the network interface 1070. [Explanation of symbols]

[0087] 1. Interactive device 2. Speech-to-text output device 3. Work terminal device 11. Speech information acquisition unit 12 Dialogue Intention Management Unit 13 Dialogue control section 14 Output section 15 Information storage section 51 Collaborative Corpus 52 Dialogue Intention Information 53 Subjective result information 54 Common Infrastructure Information

Claims

1. An estimation model that receives as input a dialogue context including utterances between coworkers who are working together to accomplish a task through dialogue at a specific dialogue stage, a dialogue intention of the dialogue between the coworkers, and subjective result information indicating an intermediate result of the collaboration by a device of the coworker, and outputs the next work result of the coworker and the next system utterance; an utterance information acquisition unit that acquires predetermined utterance information of a predetermined co-worker who performs a predetermined collaborative work to accomplish a predetermined task through dialogue; a dialogue control unit that acquires, from the predetermined utterance information, a predetermined dialogue context including utterances between the predetermined collaborators at a predetermined stage of a dialogue, acquires information on predetermined intermediate steps until the predetermined task is achieved by the predetermined collaborators performing the predetermined collaborative work through dialogue, and generates a system utterance using an estimation model based on a dialogue intention of the dialogue between the predetermined collaborators identified based on the predetermined dialogue context, the predetermined utterance information, and the predetermined intermediate step information, and on predetermined subjective result information indicating an intermediate result of the predetermined collaborative work by the device itself; an output unit that outputs the system utterance generated by the dialogue control unit to the predetermined co-worker; An interactive device comprising:

2. a dialogue intention management unit that identifies a dialogue intention of a dialogue with the predetermined co-worker based on the predetermined intermediate information, The dialogue control unit generates the system utterance based on the predetermined utterance information, the dialogue intention specified by the dialogue intention management unit, and the predetermined subjective result information.

2. The interactive device according to claim 1.

3. The dialogue device according to claim 2, characterized in that the dialogue control unit acquires training data of the predetermined utterance information and the predetermined subjective result information based on the predetermined intermediate process information, and updates the estimation model using the dialogue intention of the dialogue with the predetermined co-worker and the training data.

4. The dialogue device according to any one of claims 1 to 3, characterized in that the dialogue control unit uses the estimation model to estimate the next predetermined subjective result information and the next work result of the predetermined co-worker based on the predetermined utterance information, the predetermined intermediate process information, and the predetermined subjective result information, and identifies a common base, which is information on knowledge shared with the predetermined co-worker, based on the estimation result.

5. The dialogue device according to claim 4 , wherein the dialogue control unit generates the system utterance based on the common base with the predetermined co-worker.

6. A dialogue device acquiring predetermined utterance information of a predetermined co-worker who performs a predetermined collaborative work to accomplish a predetermined task through dialogue; acquiring a predetermined dialogue context including utterances between the predetermined collaborators at a predetermined stage of the dialogue from the predetermined utterance information; a step of acquiring information on a predetermined intermediate process until the predetermined task is accomplished by the predetermined co-workers performing the predetermined co-worker through dialogue; a step of generating a system utterance by inputting a dialogue context including utterances between coworkers who perform a collaborative work to achieve a task through dialogue, a dialogue intention of the dialogue between the coworkers, and subjective result information indicating an intermediate result of the collaborative work by their own device into an estimation model that outputs a next work result of the coworker and a next system utterance, the dialogue intention of the dialogue between the predetermined coworkers specified based on the predetermined dialogue context, the predetermined utterance information, and the predetermined intermediate process information, and predetermined subjective result information indicating an intermediate result of the predetermined collaborative work by their own device; outputting the generated system utterance to the predetermined co-worker; A dialogue control method characterized by executing the above.

7. acquire predetermined utterance information of a predetermined co-worker who performs a predetermined collaborative work to accomplish a predetermined task through dialogue; acquiring a predetermined dialogue context including utterances between the predetermined co-workers at a predetermined stage of the dialogue from the predetermined utterance information; acquiring predetermined intermediate information on the predetermined process until the predetermined task is accomplished by the predetermined co-workers performing the predetermined co-work through dialogue; a dialogue context including utterances between coworkers who perform a collaborative task to achieve a task through dialogue, a dialogue intention of the dialogue between the coworkers, and subjective result information indicating an intermediate result of the collaborative task by their own device are input to an estimation model that outputs the next work result of the coworkers and a next system utterance, and the dialogue intention of the dialogue between predetermined coworkers, which is specified based on the predetermined dialogue context, the predetermined utterance information, and the predetermined intermediate process information, and predetermined subjective result information indicating an intermediate result of the predetermined collaborative task by their own device are input to generate a system utterance; outputting the generated system utterance to the predetermined co-worker; An interactive program characterized by causing a computer to execute processing.

Images