Dialogue system, program, and control method

Abstract

To improve response speed of bot utterances, which are responses to user utterances.SOLUTION: A dialogue system includes: a prediction utterance generation unit that, between determination of a first bot utterance and the end of a first user utterance, uses a language model to generate at least one pair consisting of a first predicted user utterance and a second predicted bot utterance predicted as a response to the first predicted user utterance; a similarity measurement unit that, after the end of the first user utterance, measures similarity between each generated first predicted user utterance and the first user utterance: and an utterance identification unit that identifies the second predicted bot utterance paired with the first predicted user utterance when the maximum similarity was measured as a second bot utterance, if the maximum similarity among the measured similarities exceeds a predetermined threshold.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a dialogue system for dialogue with a user. [Background technology]

[0002] With the advancement of large-scale language models and the development of high-performance Speech2Text (Non-Patent Document 1) and Text2Speech models (Non-Patent Document 2), systems capable of natural voice dialogue are becoming possible. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever, "Robust speech recognition via large-scale weak supervision.", Proceedings of the 40th International Conference on Machine Learning,ICML'23.JMLR.org, 2023, [Non-patent document 2] Chenshuang Zhang, Chaoning Zhang, Sheng Zheng, Mengchun Zhang, Maryam Qamar, Sung-Ho Bae, and In So Kweon, "A survey on audio diffusion models: Text to speech synthesis and enhancement in generative AI", [online], 2023, [Retrieved February 1, 2024], Internet<URL https: / / arxiv.org / pdf / 2303.13336.pdf> Summary of the Invention [Problem to be solved by the invention]

[0004] Dialogue systems can be broadly divided into rule-based and generative model-based systems. Generative model-based systems have the problem of taking a long time to generate responses, as the larger the model, the longer the calculation time. Japanese speakers generally tend to begin responding to what the other person is saying within one second, but GPT-3.5-turbo, made by OpenAI in the US, for example, can only generate around 14 tokens per second, making it difficult to complete all generation and respond within one second.

[0005] One existing solution is to output the generated tokens as speech one after another before the language model has finished generating all the sentences. However, this method of outputting speech as speech before all the tokens have been generated makes it difficult to apply filtering to reject harmful sentences, and there is a possibility that the sentence may have already been spoken halfway by the time harmful sentences are discovered.

[0006] One aspect of the present invention aims to improve the response speed of bot utterances, which are responses to user utterances. [Means for solving the problem]

[0007] In order to solve the above problem, a dialogue system according to one embodiment of the present invention is a dialogue system equipped with a chatbot that dialogues with a user, the dialogue system comprising: a predicted utterance generation unit that generates a second bot utterance as a response to a first user utterance that is a response to a first bot utterance, and that uses a language model to generate at least one pair of a first predicted user utterance that is a prediction of the first user utterance and a second predicted bot utterance that is a prediction of the second bot utterance predicted as a response to the first predicted user utterance from the time the first bot utterance is determined until the first user utterance is completed; a similarity measurement unit that measures the similarity between each of the generated first predicted user utterances and the first user utterance after the first user utterance is completed; and an utterance identification unit that, when the maximum similarity among the measured similarities exceeds a predetermined threshold, identifies the second predicted bot utterance that was paired with the first predicted user utterance when the maximum similarity was measured as the second bot utterance.

[0008] In order to solve the above problem, a program according to one embodiment of the present invention is a program for causing a computer to function as a dialogue system equipped with a chatbot that dialogues with a user, wherein the dialogue system causes the computer to function as a predicted utterance generation unit that generates a second bot utterance as a response to a first user utterance that is a response to a first bot utterance, and that uses a language model to generate at least one pair of a first predicted user utterance that is a prediction of the first user utterance and a second predicted bot utterance that is a prediction of the second bot utterance predicted as a response to the first predicted user utterance from the time the first bot utterance is determined until the first user utterance is completed; a similarity measurement unit that measures the similarity between each of the generated first predicted user utterances and the first user utterance after the first user utterance is completed; and an utterance identification unit that, if the maximum similarity among the measured similarities exceeds a predetermined threshold, identifies the second predicted bot utterance that was paired with the first predicted user utterance when the maximum similarity was measured as the second bot utterance.

[0009] In order to solve the above problem, a control method according to one embodiment of the present invention is a control method for a dialogue system equipped with a chatbot that dialogues with a user, wherein the dialogue system generates a second bot utterance as a response to a first user utterance that is a response to a first bot utterance, and includes a predicted utterance generation step of generating at least one pair of a first predicted user utterance that is a prediction of the first user utterance and a second predicted bot utterance that is a prediction of the second bot utterance predicted as a response to the first predicted user utterance after the first user utterance is completed, a similarity measurement step of measuring the similarity between each of the generated first predicted user utterances and the first user utterance after the first user utterance is completed, and an utterance identification step of identifying, if the maximum similarity among the measured similarities exceeds a predetermined threshold, the second predicted bot utterance that was paired with the first predicted user utterance when the maximum similarity was measured as the second bot utterance. [Effects of the Invention]

[0010] According to the present invention, it is possible to improve the response speed of bot utterances, which are responses to user utterances. [Brief explanation of the drawings]

[0011] [Figure 1] 1 is a diagram showing an example of a dialogue between a voice dialogue system according to an embodiment of the present invention and a user; [Figure 2] FIG. 2 is a block diagram showing a configuration example of the voice dialogue system. [Figure 3] FIG. 10 is a diagram showing a specific example of a prompt when GPT-3.5-turbo is used as the language model. [Figure 4] 10 is a flowchart illustrating an example of a flow of processing executed in the voice dialogue system. [Figure 5] FIG. 10 is a diagram showing experimental results. [Figure 6] FIG. 10 is a diagram showing experimental data in which the maximum similarity was highest. DETAILED DESCRIPTION OF THE INVENTION

[0012] As an example of an interactive system according to one aspect of the present invention, a voice interactive system that interacts with a user through voice will be described. Note that the present invention is not limited to voice interactive systems, and may also be applied to, for example, a text interactive system equipped with a chatbot that interacts with a user through text.

[0013] [Outline of the Spoken Dialogue System] A voice dialogue system 1 according to one embodiment of the present invention is a system equipped with a chatbot that dialogues with a user through voice. In this specification, an utterance from the chatbot is referred to as a "bot utterance," and an utterance from the user is referred to as a "user utterance." A pair of a bot utterance and a user utterance that is a response to the bot utterance is referred to as a "turn," and the t-th turn in a dialogue is referred to as a "turn." t " (t is an integer greater than or equal to 1). Also, turn t The bot utterance and user utterance in t " and "User utterance u t " should be written as ".

[0014] The target language for the voice dialogue system 1 is assumed to be Japanese. Japanese has the characteristic that predicates are placed at the end of a sentence, so the meaning cannot be accurately determined unless the sentence is listened to until the end. On the other hand, Japanese speakers generally tend to start responding to the other person's utterance within one second. In order to deal with these unique circumstances of the Japanese language, it is often difficult for conventional systems to complete the generation of all bot utterances, which are responses to a user utterance, and output them as voice within one second of the user utterance being completed. This is because conventional systems require a turn t In bot utterance b t After the speech output of t Furthermore, in the conventional system, the user utterance u t After the end of t+1 Bot utterance b t+1 This will cause a delay in the start of the response.

[0015] In response to this, an example of a dialogue between the voice dialogue system 1 and a user will be described with reference to Fig. 1. The voice dialogue system 1 is premised on outputting a bot utterance after generating the bot utterance, and incorporates a technology for predicting utterances to enable the system to predict the turn. t In bot utterance b t While outputting the voice, the user speaks u t and then predict the user utterance u t Bot utterance b for t+1 Specifically, in the voice dialogue system 1, the turn that occurred in the conventional system is generated. t The idle time in the

[0016] [Process 1] N patterns (N is an integer greater than or equal to 1) of user utterances u t Hereafter, we will predict the N patterns of user utterances u t Each of these is called a "predicted user utterance u' k t " (k is an integer between 1 and N). Note that Figure 1 shows an example where N=2, and the predicted user utterance u' 1 t and predicted user utterance u' 2 t is predicted.

[0017] [Process 2] Predicted user utterance u' k t The predicted turn in response to each of t+1 Bot utterance b t+1 The generation process of bot utterance b t+1 "Predicted bot utterance b' k t+1 In the example shown in Figure 1, the predicted bot utterance b' 1 t+1 and predicted bot utterance b' 2 t+1 is generated.

[0018] In the voice dialogue system 1, the user utterance ut After the end of t The predicted user utterance u' that is most similar to k t (matching process), and the predicted user utterance u' is identified. k t bot utterance b' predicted as a response to k t+1 bot utterance b t+1 In the example shown in FIG. 1, the matching process generates a predicted user utterance u' 1 t is identified, resulting in a predicted bot utterance b' 1 t+1 bot utterance b t+1 It states that:

[0019] As described above, the voice dialogue system 1 is characterized by the look-ahead generation process and the matching process, which provides a significant effect not found in conventional techniques of improving the response speed of bot utterances that are responses to user utterances.

[0020] [Configuration of the Spoken Dialogue System] An example configuration of the voice dialogue system 1 will be described with reference to Fig. 2. The voice dialogue system 1 includes a control unit 2 and a storage unit 3. As shown in the figure, the voice dialogue system 1 may further include a voice recognition unit 4 and a voice synthesis unit 5.

[0021] The control unit 2 controls various functions including the voice dialogue function in an integrated manner. The control unit 2 is, for example, a processor or a CPU (Central Processing Unit). The control unit 2 may be realized by, for example, a logic circuit. The control unit 2 reads out a program for causing the computer to function as a control unit that executes various functions including the voice dialogue function, for example, from the storage unit 3, and loads it into a RAM (Random Access Memory) (not shown) for execution.

[0022] The control unit 2 includes a look-ahead generation processing unit 21, a matching processing unit 22, and an utterance generation unit .

[0023] The look-ahead generation processing unit 21 predicts a user utterance based on the dialogue history up to that point before the user utterance is finished, and generates a bot utterance that is a response to the user utterance. The look-ahead generation processing unit 21 includes a predicted utterance generation unit 211 and an embedding unit 212.

[0024] The predicted utterance generation unit 211 t In this case, the user utterance u is calculated from the dialogue history. t Predicted user utterance u', which is a prediction of k t The predicted utterance generation unit 211 predicts N patterns (that is, at least one) of the predicted user utterance u'. k t The predicted bot utterance b' is predicted as a response to each of k t+1 That is, the predicted utterance generation unit 211 generates the predicted user utterance u' k t and predicted bot utterance b' k t+1 Generate N patterns of pairs with .

[0025] N may be a predetermined fixed value or a variable value that changes depending on the context or turn. As an example of a case where N is a variable value, for example, a user utterance u that can be taken in the context t In situations where the number of possible user utterances is expected to be limited, N may be changed to a relatively small value. t In situations where it is expected that there will be many such patterns, N may be changed to a relatively large value.

[0026] Predicted user utterance u' by the predicted utterance generation unit 211 k t Prediction processing and predicted bot utterance b' k t+1 The generation process of bot utterance b t is determined (specifically, the utterance determination unit 223, which will be described later, determines the bot utterance b t is specified, or the utterance generation unit 23 (to be described later) generates a bot utterance b tis generated), and the user utterance u t It is preferable that the user utterance u be completed within the time limit. t When the user utterance u' is completed, k t If the prediction is correct, the predicted bot utterance b' is generated. k t+1 Turn t+1 In bot utterance b t+1 The audio can be output immediately as

[0027] Predicted user utterance u' k t Prediction of and predicted bot utterance b' k t+1 The predicted utterance generation unit 211 may use, for example, a known language model to generate the predicted user utterance u'. The language model used by the predicted utterance generation unit 211 is generated by inputting a dialogue history. k t and predicted bot utterance b' k t+1 Any type of model can be used as long as it can output the following, but a large-scale language model is preferable. t The dialogue history input to the language model in t , as well as previous bot utterances and user utterances. A specific example of a prompt when GPT-3.5-turbo is used as the language model is shown in FIG. 3. The language model used by the predictive utterance generation unit 211 may be subjected to few-shot learning. The location of the language model used by the predictive utterance generation unit 211 is not limited, and may be stored in the storage unit 3 or an external device capable of communicating with the voice dialogue system 1.

[0028] The embedding unit 212 embeds the predicted user utterance u' k t Each of these is represented by an embedding eu' k tFor example, Sentence BERT may be used to convert the user utterance u into an embedded representation. t It is preferable that the process is completed by the time the end of

[0029] As described above, the look-ahead generation processing unit 21 basically uses a known language model to generate a user utterance u t Predicted user utterance u', which is a prediction of k t and predicted user utterance u' k t Bot utterance b, predicted as a response to t+1 Predicted bot utterance b' is a prediction of k t+1 and generate at least one pair of predicted user utterance u' k t Each of these is represented by an embedding eu' k t In this process, the pre-reading generation processing unit 21 preferably converts the bot utterance b t After the bot utterance b starts to be output, t A user utterance u is a response to t By the time the turn is over (in other words, t In order to complete the process within the period, N may be adjusted to a relatively small value, for example, to reduce the processing time taken by the predictive utterance generation unit 211.

[0030] The look-ahead generation processing unit 21 performs the embedded eu' k t and predicted bot utterance b' k t+1 It is preferable to cache the predicted user utterance u' in a memory or the like (not shown) for subsequent use by the matching processing unit 22. k t may be cached.

[0031] The matching processing unit 22 processes the actual user utterance u tAfter the completion of the above, the predicted user utterance u' predicted by the look-ahead generation processing unit 21 is k t is the actual user utterance u t If the predicted bot utterance is similar to the utterance generated by the look-ahead generation processing unit 21, the matching processing unit 22 identifies the generated predicted bot utterance as the actual bot utterance. The matching processing unit 22 includes an embedding unit 221, a similarity measurement unit 222, and an utterance identification unit 223.

[0032] The embedded portion 221 is a turn t User utterance u in t After the end of t The embedding representation is the embedding eu t Convert to.

[0033] The similarity measurement unit 222 measures the embedded eu' converted by the embedding unit 212. k t For each of the embedded eu converted by the embedding unit 221, t Cosine similarity score with k Hereafter, we will measure the cosine similarity score k The value of k when is the maximum value (i.e., the value closest to 1) is denoted as max. In other words, the cosine similarity score max is the cosine similarity score k is the maximum similarity among

[0034] The utterance identification unit 223 calculates the cosine similarity score max If exceeds a threshold T (for example, T = 0.75), the look-ahead is considered successful and the embedding eu' max t The original predicted user utterance u' is converted to max t The predicted bot utterance b' that has already been generated by the look-ahead generation processing unit 21 as a pair with max t+1 bot utterance b t+1 Identify as:

[0035] The threshold T may be a predetermined fixed value or a variable value that changes depending on the context or the turn. As an example of when the threshold T is a variable value, the threshold T may be changed to a relatively high value in a situation where the context requires a highly accurate response (e.g., financial transactions, etc.), and conversely, the threshold T may be changed to a relatively low value in a situation where the context does not necessarily require a highly accurate response (e.g., casual conversation, etc.).

[0036] As described above, in short, the matching processing unit 22 calculates the user utterance u t After the end of the predicted user utterance u' k t and user utterance u t Specifically, the similarity measurement unit 222 measures the similarity between the predicted user utterance u′ and the predicted user utterance u′. k t The embedding eu' is an embedding representation of k t and user utterance u t The embedding eu is an embedding representation of t Cosine similarity score with k Then, the matching processing unit 22 measures the cosine similarity score in the utterance identification unit 223. max If exceeds the threshold T, the cosine similarity score max Predicted user utterance u' when measuring max t The paired predicted bot utterance b' max t+1 bot utterance b t+1 Identify as:

[0037] User utterance u by the embedding unit 221 t Embed from eu t the cosine similarity score by the similarity measurement unit 222. k and a process of measuring the predicted bot utterance b' by the utterance identification unit 223. max t+1 The total processing time for the process of identifying the user utterance u (e.g., about 0.01 seconds) is calculated using a known language model. t bot utterance bt+1 This is extremely short compared to the processing time (for example, several seconds to several tens of seconds) required by conventional techniques to generate a response. Therefore, according to the present invention, it is possible to realize a response that is significantly faster than the conventional techniques while taking advantage of the language model's ability to generate a response according to the context.

[0038] The utterance generation unit 23 calculates the cosine similarity score max If T is equal to or smaller than the threshold T, a method similar to that of the prior art is used to convert the user utterance u t bot utterance b t+1 This generates the predicted user utterance u' k t The language model used by the utterance generation unit 23 is a language model for the user utterance u. t By inputting bot utterance b t+1 Although any type of language model can be used as long as it can output the above, a large-scale language model is preferable. The location of the language model used by the utterance generation unit 23 is not limited, and it may be stored in the storage unit 3 or in an external device that can communicate with the voice dialogue system 1.

[0039] The storage unit 3 stores various data and programs used by the voice dialogue system 1, and is, for example, a semiconductor drive such as a solid state drive or a magnetic disk such as a hard disk. The data stored in the storage unit 3 is, in particular, the dialogue history (the history of user utterances and the history of bot utterances).

[0040] The speech recognition unit 4 recognizes the user's speech input to a speech input unit 6 such as a microphone, and outputs the user's speech as the recognition result. The speech input unit 6 is provided in, for example, a terminal device used by the user that can communicate with the speech dialogue system 1.

[0041] The speech synthesis unit 5 performs speech synthesis using the bot utterance as input, and outputs the synthesis result to a speech output unit 7 such as a speaker. The speech output unit 7 is provided, for example, in a terminal device used by a user that can communicate with the speech dialogue system 1.

[0042] [Example of processing flow executed by the voice dialogue system 1] FIG. 4 is a flowchart showing an example of the flow of processing (a control method for the voice dialogue system 1) executed by the control unit 2 of the voice dialogue system 1.

[0043] First, turn t Bot utterance b t When the predicted utterance generation unit 211 determines the predicted user utterance u' using a known language model, k t and predicted bot utterance b' k t+1 Then, the embedding unit 212 generates N patterns of pairs of the predicted user utterance u′ and the predicted user utterance u′ (S1, predicted utterance generation step). k t Embed each of eu' k t (S2) The steps up to step S2 are converted into the user utterance u t It is preferable that the process is completed by the time the end of

[0044] Next, user utterance u t After the completion of the above, the embedding unit 221 embeds the user utterance u t Embed EU t Then, the similarity measurement unit 222 converts the embedded eu' k t For each of the embeddings, t Cosine similarity score with k (S4, similarity measurement step). Then, the cosine similarity score max exceeds the threshold T (Yes in S5), the utterance identification unit 223 determines the predicted bot utterance b' max t+1 bot utterance b t+1 (S6, utterance identification step).

[0045] On the other hand, the cosine similarity score max If is equal to or smaller than the threshold T (No in S5), the utterance generation unit 23 uses a known language model to generate the user utterance u t bot utterance b t+1 is generated (S7).

[0046] Then, the control unit 2 performs the bot utterance b identified in step S6. t+1 Or, bot utterance b generated in step S7 t+1 Finally, the speech synthesis unit 5 sends the bot utterance b t+1 The speech synthesis is performed using the input, and the synthesis result is output to the speech output unit 7 (S9). [Experiment and Evaluation] As a result of experiments conducted by the applicants, it was confirmed that in the voice dialogue system 1, the look-ahead generation processing unit 21 can look ahead to approximately 20% of user utterances with a high degree of similarity, and that a bot utterance can be returned in a natural response time consisting only of the processing time of the matching processing unit 22. It was also confirmed that look-ahead prediction is more likely to be successful when the bot utterance immediately before the user utterance is in a context that can narrow down the user utterance. These details are described below.

[0047] [Experimental data] The experimental data used were dialogue logs collected by the applicants at the Conversational Robot Competition 2022. This dialogue was a task-oriented dialogue in which the bot was a travel agency clerk and the user was a customer visiting the travel agency, and the dialogue lasted for five minutes to determine the user's travel plans. This data was selected because it was based on an actual conversation between a bot and a human, and is effective for measuring the practicality of the present invention. There were a total of 71 user utterances in the seven dialogue logs, and the present invention was applied to all of them and evaluated.

[0048] [Evaluation indicators] To evaluate the improvement in response generation speed due to lookahead, the applicants used three evaluation indices. The first index relates to the time required for lookahead. Specifically, the evaluation indices examined were the "look ahead duration," which measures the time it takes from the start of lookahead generation to the completion of N patterns, and the "matching duration," which measures the time it takes to match the user's utterance with the lookahead. As a reference index, the experimental data was manually annotated, and the average bot speech time was 20.77 seconds, and the average speech time including the user's utterance was 23.90 seconds. Furthermore, due to the nature of the task, explanations about tourist spots sometimes took more than 30 seconds. When these explanations were omitted and only relatively short utterances were included, the average bot speech time was 11.21 seconds, and the average time including the user's utterance was 14.05 seconds.

[0049] The second index is an index related to the accuracy of prediction. This index is divided into two parts. The first is the degree to which the user utterance u is based on the dialogue history up to that point. t The second index is the ability to accurately predict the predicted user utterance u'. This index is measured by the number of data (number of matches) that exceed the threshold T and the average maximum similarity for each data. k t and user utterance u t This is the degree of agreement, which indicates how accurately the bot can judge whether a pre-read utterance matches the actual user utterance. If an inaccurate pre-read utterance is deemed to match the actual user utterance, a bot utterance in the wrong context will be returned as a response. To evaluate the frequency of this occurrence, we manually evaluated the degree to which pre-read utterances whose maximum similarity exceeded the threshold T matched the actual user utterance on a five-point scale from 1 (no match at all) to 5 (complete match), and the average was taken as the "degree of agreement."

[0050] The third metric is how well the pre-generated bot response is in line with the context up to that point. max t is the user utterance u t Even if the bot could accurately predict the next utterance b',max t+1 If the utterance is out of context, the benefit of predicting the future is lost. Therefore, based on the context up to that point, the predictive bot utterance b' max t+1 The naturalness of the sentences was assessed manually on a two-point scale: 0: unnatural, 1: natural, and the average of these scores was used to determine the naturalness of the sentences. Since the degree of agreement and naturalness are subjective assessments, we report the average of the scores of three annotators.

[0051] [Experimental results] Figure 5 shows the experimental results. First, we will explain the index related to the time required for look-ahead. Approximately one second of generation time is required for each pattern, and the more predicted patterns there are, the longer the time required for look-ahead. On the other hand, the time required for matching remains almost the same even if the number of patterns increases. Furthermore, this time is negligible compared to the time required for look-ahead. As mentioned above, the average utterance time, including the user's utterance, is 14.05 seconds, and on average, the present invention can be suitably applied as long as N is at least around 20. When there are many utterances of 5 to 10 seconds, an N of around 3 to 5 is considered useful.

[0052] Next, we will explain the prediction accuracy. First, looking at the number of matches, in all settings, 16 out of 71 user utterances, or approximately 20% or more, were predicted with a similarity above the threshold T, demonstrating a certain degree of effectiveness.

[0053] Next, we will explain the average maximum similarity. We can see that the similarity increases as the number of patterns increases. This suggests that increasing the number of patterns makes it possible to cover a greater number of possible answers, but there is no strong correlation between the average maximum similarity and the number of matches, and we can see that predicting high similarities does not necessarily increase just because the number of patterns is increased. The score for the match is roughly close to 4, and for matches that were determined to be matches, the match determination was basically correct.

[0054] Finally, regarding the naturalness of responses, we can see that a fairly high percentage of participants were able to give natural responses when N=3.

[0055] Figure 6 shows the two with the highest maximum similarity out of the 71 pieces of experimental data. The ones with the highest maximum similarity can be said to be contexts that narrow down the user's response candidates to a certain extent. The responses returned were what could be called typical examples of answers, and this led to high prediction accuracy.

[0056] [Software implementation example] The functions of the voice dialogue system 1 (hereinafter referred to as the "device") can be realized by a program that causes a computer to function as the device, and a program that causes a computer to function as each control block of the device (particularly each part included in the control unit 2).

[0057] In this case, the device includes a computer having at least one control device (e.g., a processor) and at least one storage device (e.g., a memory) as hardware for executing the program. The control device and storage device execute the program, thereby realizing the functions described in each of the above embodiments.

[0058] The program may be non-transitory and may be recorded on one or more computer-readable recording media. The recording media may or may not be included in the device. In the latter case, the program may be supplied to the device via any wired or wireless transmission medium.

[0059] Furthermore, some or all of the functions of the control blocks can be realized by logic circuits. For example, an integrated circuit in which a logic circuit that functions as each of the control blocks is formed is also included in the scope of the present invention. In addition, the functions of the control blocks can also be realized by, for example, a quantum computer.

[0060] Furthermore, each process described in each of the above embodiments may be executed by AI (Artificial Intelligence). In this case, the AI ​​may run on the control device or on another device (for example, an edge computer or a cloud server).

[0061] [Additional Notes] The application of the technical idea of ​​the present invention is not limited to voice dialogue systems, but may also be applied to, for example, a text dialogue system equipped with a chatbot that converses with a user via text.

[0062] The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present invention.

[0063] 〔summary〕 A voice dialogue system according to a first aspect of the present invention is a voice dialogue system equipped with a chatbot that verbally interacts with a user, the voice dialogue system comprising: a predicted utterance generation unit that generates a second bot utterance as a response to a first user utterance that is a response to a first bot utterance; and that uses a language model to generate at least one pair of a first predicted user utterance that is a prediction of the first user utterance and a second predicted bot utterance that is a prediction of the second bot utterance predicted as a response to the first predicted user utterance from the time the first bot utterance is determined until the first user utterance is completed; a similarity measurement unit that measures the similarity between each of the generated first predicted user utterances and the first user utterance after the first user utterance is completed; and an utterance identification unit that, if the maximum similarity among the measured similarities exceeds a predetermined threshold, identifies the second predicted bot utterance that was paired with the first predicted user utterance when the maximum similarity was measured as the second bot utterance.

[0064] A voice dialogue system according to a second aspect of the present invention is based on the first aspect, and the similarity measurement unit may measure the similarity between an embedded expression of the first predicted user utterance and an embedded expression of the first user utterance.

[0065] The voice dialogue system of aspect 3 of the present invention may further include, in aspect 1 or 2 above, an utterance generation unit that uses a language model to generate the second bot utterance using the first user utterance as input when the maximum similarity is less than or equal to the threshold.

[0066] The voice dialogue system according to aspect 4 of the present invention may further include a voice recognition unit that outputs a user utterance as a recognition result of the user's voice in any of aspects 1 to 3 above, and a voice synthesis unit that performs voice synthesis using a bot utterance as input.

[0067] A program according to a fifth aspect of the present invention is a program for causing a computer to function as a dialogue system equipped with a chatbot that dialogues with a user, wherein the dialogue system causes the computer to function as a predicted utterance generation unit that generates a second bot utterance as a response to a first user utterance that is a response to a first bot utterance, and that uses a language model to generate at least one pair of a first predicted user utterance that is a prediction of the first user utterance and a second predicted bot utterance that is a prediction of the second bot utterance predicted as a response to the first predicted user utterance from the time the first bot utterance is determined until the first user utterance is completed; a similarity measurement unit that measures the similarity between each of the generated first predicted user utterances and the first user utterance after the first user utterance is completed; and an utterance identification unit that, if the maximum similarity among the measured similarities exceeds a predetermined threshold, identifies the second predicted bot utterance that was paired with the first predicted user utterance when the maximum similarity was measured as the second bot utterance.

[0068] A control method according to aspect 6 of the present invention is a control method for a dialogue system equipped with a chatbot that dialogues with a user, wherein the dialogue system generates a second bot utterance as a response to a first user utterance that is a response to a first bot utterance, and includes a predicted utterance generation step of using a language model to generate at least one pair of a first predicted user utterance that is a prediction of the first user utterance and a second predicted bot utterance that is a prediction of the second bot utterance predicted as a response to the first predicted user utterance after the first user utterance has ended, a similarity measurement step of measuring the similarity between each of the generated first predicted user utterances and the first user utterance after the first user utterance has ended, and an utterance identification step of identifying, if the maximum similarity among the measured similarities exceeds a predetermined threshold, the second predicted bot utterance that was paired with the first predicted user utterance when the maximum similarity was measured, as the second bot utterance. [Explanation of symbols]

[0069] 1. Spoken dialogue system 4. Voice Recognition Unit 5. Speech synthesis unit 21 Read-ahead generation processing unit 211 Predictive speech generation unit 22 Matching processing section 222 Similarity measurement part 223 Utterance Identification Unit 23 Speech generation unit b t turn t Bot utterances in u t turn t User utterances in

Claims

1. A dialogue system equipped with a chatbot that dialogues with a user, the dialogue system generates a second bot utterance in response to a first user utterance that is a response to the first bot utterance; a predicted utterance generation unit that generates, using a language model, at least one pair of a first predicted user utterance that is a prediction of the first user utterance and a second predicted bot utterance that is a prediction of the second bot utterance that is predicted as a response to the first predicted user utterance, from when the first bot utterance is determined until when the first user utterance is completed; a similarity measurement unit that measures a similarity between each of the generated first predicted user utterances and the first user utterance after the first user utterance is completed; an utterance identification unit that, when a maximum similarity among the measured similarities exceeds a predetermined threshold, identifies the second predicted bot utterance paired with the first predicted user utterance when the maximum similarity is measured as the second bot utterance, Dialogue system.

2. the similarity measurement unit measures a similarity between the embedded representation of the first predicted user utterance and the embedded representation of the first user utterance. The dialogue system according to claim 1 .

3. and an utterance generation unit that generates the second bot utterance using a language model based on the first user utterance when the maximum similarity is equal to or less than the threshold. The dialogue system according to claim 1 .

4. the dialogue is a voice dialogue, a speech recognition unit that outputs a user utterance as a recognition result of the user's speech; and a speech synthesis unit that performs speech synthesis using a bot utterance as an input. The dialogue system according to claim 1 .

5. A program for causing a computer to function as an interactive system equipped with a chatbot that interacts with a user, the dialogue system generates a second bot utterance in response to a first user utterance that is a response to the first bot utterance; a predicted utterance generation unit that generates, using a language model, at least one pair of a first predicted user utterance that is a prediction of the first user utterance and a second predicted bot utterance that is a prediction of the second bot utterance that is predicted as a response to the first predicted user utterance, from the time the first bot utterance is determined until the time the first user utterance is completed; a similarity measurement unit that measures a similarity between each of the generated first predicted user utterances and the first user utterance after the first user utterance is completed; and A program for causing a computer to function as an utterance identification unit that, when the maximum similarity among the measured similarities exceeds a predetermined threshold, identifies the second predicted bot utterance paired with the first predicted user utterance when the maximum similarity is measured as the second bot utterance.

6. A control method for a dialogue system equipped with a chatbot that interacts with a user, comprising: the dialogue system generates a second bot utterance in response to a first user utterance that is a response to the first bot utterance; a predicted utterance generation step in which a predicted utterance generation unit uses a language model to generate at least one pair of a first predicted user utterance that is a prediction of the first user utterance and a second predicted bot utterance that is a prediction of the second bot utterance that is predicted as a response to the first predicted user utterance, from the time the first bot utterance is determined until the time the first user utterance is completed; a similarity measurement step in which, after the first user utterance is completed, a similarity measurement unit measures a similarity between each of the first predicted user utterances generated and the first user utterance; and an utterance identification step of, when a maximum similarity among the measured similarities exceeds a predetermined threshold, identifying the second predicted bot utterance paired with the first predicted user utterance when the maximum similarity is measured as the second bot utterance. Control method.

Images