Dialogue device, dialogue method, and dialogue program

Abstract

Based on the user's past utterances, it can provide accurate responses to the user's utterances. [Solution] The system includes: an inference unit that, upon detecting a user's utterance, uses a first trained language model to infer the topic of the current utterance from utterance data that includes at least summaries or tags created in relation to the user's past utterances; a creation unit that, if the inference unit is able to infer the topic, creates a prompt to input the topic into the current utterance to the second trained language model; and, if the inference unit is unable to infer the topic, creates a prompt based on the current utterance.

Description

Technical Field

[0001] The present disclosure relates to an interaction device, an interaction method, and an interaction program.

Background Art

[0002] There is known a voice input device that outputs an interaction history before interruption at the restart of voice input and provides it to the speaker.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] By including the past conversation history in the prompt input to the trained language model, the prior art realizes an interaction that takes into account the content of past conversations. However, if all past conversations are included in the prompt, it will cause limitations in the amount of prompts that can be input to the trained language model, an increase in response time, an increase in usage fees, etc.

[0005] The present disclosure has been made in view of the above, and an object thereof is to provide an interaction device, an interaction method, and an interaction program capable of performing an interaction that captures the user's intention.

Means for Solving the Problems

[0006] To solve the above-mentioned problems and achieve the objectives, the dialogue device according to this disclosure is characterized by comprising: an inference unit that, when it detects a user's utterance, uses a first trained language model to infer a topic indicating what the current utterance is about from utterance data in which at least summaries or tags created in relation to the user's past utterances are recorded; a creation unit that, when the inference unit is able to infer the topic, creates a prompt to input the current utterance containing the topic to a second trained language model, and when the inference unit is unable to infer the topic, creates the prompt based on the current utterance.

[0007] Furthermore, the dialogue method relating to this disclosure is a dialogue method performed by a dialogue device, and is characterized by including: an inference step in which, when a user's utterance is detected, the device uses a first trained language model to infer a topic indicating what the current utterance is about from utterance data in which at least summaries or tags created in relation to the user's past utterances are recorded; a creation step in which, if the topic can be inferred by the inference step, the device creates a prompt to input the current utterance into a second trained language model by including the topic; and if the topic cannot be inferred by the inference step, the device creates the prompt based on the current utterance.

[0008] Furthermore, the dialogue program relating to this disclosure is characterized in that, upon detecting a user's utterance, it causes a computer to execute an inference procedure that uses a first trained language model to infer the topic of the current utterance from utterance data in which at least summaries or tags created in relation to the user's past utterances are recorded; if the topic can be inferred by the inference procedure, it creates a prompt to input the topic into the current utterance and input it into a second trained language model; and if the topic cannot be inferred by the inference procedure, it creates a prompt based on the current utterance. [Brief explanation of the drawing]

[0009] [Figure 1]Figure 1 shows an example of how the dialogue device is used. [Figure 2] Figure 2 is a block diagram showing an example configuration of the dialogue device according to Embodiment 1. [Figure 3] Figure 3 shows an example of speech data according to Embodiment 1. [Figure 4] Figure 4 shows an example of the processing content for determining whether the subject of the user's utterance is clear according to Embodiment 1. [Figure 5] Figure 5 shows an example of the processing content for inferring the topic of a user's utterance according to Embodiment 1. [Figure 6] Figure 6 shows an example of the processing content for creating a prompt and responding, including the user's utterance and the presumed topic, according to Embodiment 1. [Figure 7] Figure 7 is a flowchart showing an example of the dialogue processing flow of the dialogue device according to Embodiment 1. [Figure 8] Figure 8 shows an example of the processing content for inferring the topic of the user's utterance and confirming it with the user, according to Embodiment 2. [Figure 9] Figure 9 shows an example of the processing content for creating a prompt and responding to the user's utterance, including the confirmed topic, according to Embodiment 2. [Figure 10] Figure 10 is a diagram showing an example of the processing content for inferring the topic of the user's utterance and confirming it with the user according to Embodiment 2. [Figure 11] Figure 11 shows an example of the processing content for creating a prompt and responding to a user's utterance, including the results of confirmation, according to Embodiment 2. [Figure 12] Figure 12 is a diagram showing an example of the processing content for inferring the topic of the user's utterance and confirming it with the user according to Embodiment 2. [Figure 13] Figure 13 shows an example of the processing content for creating a prompt and responding to a user's utterance, including the results of confirmation, according to Embodiment 2. [Figure 14] Figure 14 is a flowchart showing an example of the dialogue processing flow of the dialogue device according to Embodiment 2. [Figure 15]FIG. 15 is a diagram showing an example of processing content inferred from the latest summary among related summaries according to Embodiment 3. [Figure 16] FIG. 16 is a diagram showing an example of processing content inferred from summaries with a common speech environment among related summaries according to Embodiment 4. [Figure 17] FIG. 17 is a diagram showing an application example of the dialogue device. [Figure 18] FIG. 18 is a hardware configuration diagram showing an example of a computer that realizes the functions of the dialogue device according to Embodiment 1, Embodiment 2, Embodiment 3, Embodiment 4, and Embodiment 5.

Embodiments for Carrying out the Invention

[0010] Hereinafter, embodiments of the dialogue device, dialogue method, and dialogue program according to the present disclosure will be described in detail based on the drawings. Note that the present disclosure is not limited to the embodiments described below. Also, the respective embodiments can be appropriately combined within a non - contradictory range.

[0011] 〔Embodiment 1〕 FIG. 1 is a diagram showing an example of the usage situation of the dialogue device. As shown in FIG. 1, a user who is a driver or a passenger activates the in - vehicle dialogue device 1 and speaks to the dialogue device 1. Then, the dialogue device 1 responds to the user's current speech.

[0012] The dialogue device 1 is realized by a computer or the like mounted on the vehicle. For example, the dialogue device 1 may be a dedicated navigation device built - in or loaded on the vehicle. For example, the dialogue device 1 may be a single device having a navigation function, a display function, and a dialogue function.

[0013] In addition, a user who boards a vehicle can also substitute a portable terminal device (e.g., a smartphone, a tablet terminal, a notebook PC, a desktop PC, a PDA, etc.) that the user uses daily by introducing a predetermined application, and use this as the dialogue device 1. For example, when the portable terminal device is utilized as the dialogue device 1, it is installed, for example, on the dashboard of the vehicle during driving.

[0014] By the way, regarding a new utterance, when the theme indicating what the user wants to know is unclear, a conventional dialogue device has given a general answer. For example, when the user made an utterance about a company, the general answer was to provide an overview of the company that was spoken about. However, since the conventional dialogue device gives a general answer, a situation has occurred where an answer that deviates from the user's theme is given.

[0015] Also, a technique of using a trained language model to create a more accurate answer is also known. For example, a conventional dialogue device creates a prompt that includes all past dialogues in the user's utterance and inputs it into the trained language model, and then accurately answers the user's utterance by inputting it into the trained language model.

[0016] However, in order for a conventional dialogue device to create a more accurate answer, it is necessary to create a prompt that includes a new utterance and all the user's past utterances, and the amount of the prompt increases. Along with that, in a conventional dialogue device, problems such as a limit on the amount of the prompt input to the trained language model, an increase in the response time of the trained language model, and an increase in the usage fee of the trained language model occur.

[0017] Therefore, in the present disclosure, when the user's utterance content does not include a theme, a dialogue device 1 that can conduct a dialogue on the premise of the user's past utterances by inferring the theme from the past dialogue content using a trained language model will be described.

[0018] Specifically, when the dialogue device 1 of this disclosure detects a user's utterance, it uses a first trained language model 63 to infer the topic of the current utterance from utterance data 61, which contains at least summaries or tags created in relation to the user's past utterances. If the dialogue device 1 is able to infer the topic, it creates a prompt to input the topic into the second trained language model 64; if it is unable to infer the topic, it creates a prompt based on the current utterance. The dialogue device 1 then inputs the created prompt into the second trained language model 64 and responds to the user's current utterance.

[0019] For example, a user might say to dialogue device 1, "What about Chinese food?" Dialogue device 1 detects the utterance and uses the first pre-trained language model 63 to infer the topic from the summary or tags related to Chinese food contained in the utterance data 61. As a result, dialogue device 1 infers that the topic is "Restaurants near your current location." Dialogue device 1 then creates a prompt that includes the inferred topic in the current utterance and inputs it into the second pre-trained language model 64. Dialogue device 1 then responds, "The Chinese restaurants near your current location are as follows..."

[0020] Thus, even when the topic of the user's utterance is unclear, the dialogue device 1 can infer the topic of the user's utterance and respond appropriately to the user's utterance.

[0021] (Configuration of Dialogue Device 1) Next, an example of the functional configuration of the dialogue device 1 according to Embodiment 1 will be described. Figure 2 is a block diagram showing an example of the configuration of the dialogue device 1 according to Embodiment 1.

[0022] As shown in Figure 2, the dialogue device 1 consists of a communication unit 2, an audio input unit 3, an output unit 4, a control unit 5, a storage unit 6, and the like.

[0023] The communication unit 2, under the control of the control unit 5, connects to the server via the network and sends and receives various types of information. For example, if speech data or trained language models contained in the memory unit 6 of the dialogue device 1 are stored in an external device such as a server, the communication unit 2 sends and receives data via the network.

[0024] The voice input unit 3 includes a microphone or the like for capturing sound, and generates voice information based on the captured sound. For example, under the control of the control unit 5, the voice input unit 3 generates voice information when it detects speech from a user, such as a driver or passenger. The voice input unit 3 may also generate voice information for speech environments other than those of the user. The voice input unit 3 then outputs the generated voice information to the control unit 5.

[0025] The output unit 4 outputs information under the control of the control unit 5. For example, the output unit 4 includes an audio output device and outputs the information input from the control unit 5 as audio. The output unit 4 also includes a display device and, in addition to audio output, may display characters, symbols, etc., on an in-vehicle display or the like of the dialogue device 1.

[0026] The memory unit 6 stores various programs executed by the control unit 5 (dialogue programs according to this embodiment), as well as data necessary for the control unit 5 to perform processing. As shown in Figure 2, the memory unit 6 includes speech data 61, a trained language model for judgment 62, a first trained language model 63, a second trained language model 64, and the like.

[0027] The speech data 61 stores summaries of utterances previously made by the user. Figure 3 shows an example of the speech data 61. As shown in Figure 3, the speech data 61 is information associated with "date and time," "summary," and "tags." Here, the date and time records the date and time when the user's utterance was detected. The summary records a summary of what the user said to the dialogue device 1. The tags record keywords from the summary of what the user said to the dialogue device 1.

[0028] In the example in Figure 3, it is shown that the user spoke to the dialogue device 1 at the date and time "10 / 17 14:13", and a summary of the utterance, "The user is talking about the stock price of Company A...", and the tag "Company A stock price..." were recorded in the utterance data 61.

[0029] The pre-trained language model 62 for determination is a language model trained to determine whether the topic of the user's utterance is clear or unclear. For example, the pre-trained language model 62 for determination determines whether the topic of the input utterance is clear or unclear and outputs the determination result to the determination unit 51.

[0030] The first trained language model 63 is a language model trained to search for data related to the user's utterance in accordance with the input prompt, and to infer and output the topic of the utterance based on the search results. For example, the first trained language model 63 searches for a summary or tag related to the current utterance from the utterance data 61, and infers the topic of the current utterance from the searched summary or tag, and the summary and tag, and outputs it to the inference unit 52.

[0031] The second trained language model 64 is a language model trained to create and output a response according to an input prompt. For example, if the second trained language model 64 receives a prompt that is designed to include a predicted topic in the current utterance, it will create a response according to the prompt's instructions and output it to the output unit 4.

[0032] Furthermore, the pre-trained language model 62 for judgment and the first pre-trained language model 63 may be the same, and the pre-trained language model 62 for judgment and the second pre-trained language model 64 may be the same. Also, the first pre-trained language model 63 and the second pre-trained language model 64 may be the same, and the pre-trained language model 62 for judgment, the first pre-trained language model 63, and the second pre-trained language model 64 may be the same.

[0033] For example, the pre-trained language model 62 for judgment, the first pre-trained language model 63, and the second pre-trained language model 64 can be large language models (LLMs). Furthermore, the pre-trained language model 62 for judgment, the first pre-trained language model 63, and the second pre-trained language model 64 may be a group of language models containing multiple language models. Note that the configuration and type of the pre-trained language model 62 for judgment, the first pre-trained language model 63, and the second pre-trained language model 64 are not particularly limited, as long as they are machine learning models trained to handle the processing described later.

[0034] The control unit 5 is implemented by a controller such as a CPU (Central Processing Unit) or MPU (Micro Processing Unit) executing various programs stored in the memory unit 6, thereby controlling the operation of the entire interactive device 1. Note that the control unit 5 may be composed of integrated circuits other than a CPU or MPU, such as an ASIC (Application Specific Integrated Circuit) or FPGA (Field Programmable Gate Array).

[0035] As shown in Figure 2, this control unit 5 includes a determination unit 51, an estimation unit 52, and a creation unit 53, etc.

[0036] When the voice input unit 3 detects human speech, the determination unit 51 uses a pre-trained language model 62 to determine whether the topic of the current utterance is clear. Figure 4 illustrates the process by which the determination unit 51 determines the clarity of the topic. Figure 4 shows the process for determining whether the topic of the user's utterance is clear.

[0037] Figure 4 shows the input process to the trained language model 62 for determination and the output process from the trained language model 62 for determination in order to determine the clarity of the subject of a new utterance according to Embodiment 1. For example, as shown in Figure 4, first the determination unit 51 receives the user's utterance, "What about company B?", from the voice input unit 3.

[0038] Next, as shown in Figure 4A, the determination unit 51 creates a prompt in which it sets the question content in "Q:" to indicate a question and the content of the current utterance in "Current Utterance:" to indicate that it is the current utterance, in order to determine whether the detected utterance is clear or unclear. In the example in Figure 4A, the question content indicating instructions for the trained language model 62 for determination is set to "Is the subject of the current utterance below clear?". The content of the current utterance is set to "What about company B?". The determination unit 51 then inputs the created prompt into the trained language model 62 for determination.

[0039] Next, as shown in Figure 4B, the trained language model 62 for judgment outputs the content of the judgment result, "The question 'What about company B?' does not have a clear topic. Specifically..." to "A:", which indicates the answer. The judgment unit 51 then determines that the topic is not clear and instructs the inference unit 52 to infer the topic. If the topic is determined to be clear, the judgment unit 51 instructs the creation unit 53 to create a prompt that includes a question indicating instructions for the second trained language model 64 based on the current utterance.

[0040] Returning to Figure 2, if the subject is not clear from the determination result of the determination unit 51, the inference unit 52 uses the first trained language model 63 to infer the subject of the utterance from the relevant summaries or tags, summaries and tags included in the utterance data 61.

[0041] Here, we will explain the process of inferring the topic of an utterance using Figure 5. Figure 5 is a diagram illustrating the process of inferring the topic of an utterance.

[0042] As shown in Figure 5C, the inference unit 52 creates a prompt by adding the detected utterance "What is Company B?" to the question "Q: Obtain a summary or tag related to the current utterance from the speech data and infer the topic of the current utterance," which is an instruction to the first trained language model 63, and inputs it into the first trained language model 63.

[0043] Next, as shown in Figure 5D, the first trained language model 63 outputs the relevant summary and related tags obtained from the speech data 61 along with the prediction result to "A: Obtain the following summary and tags related to Company B...". Note that either the relevant summary or the related tags may be included. Then, the prediction unit 52 outputs the prediction result to the creation unit 53.

[0044] Returning to Figure 2, the creation unit 53 creates a prompt to be input to the second trained language model 64. For example, if the inference unit 52 is able to infer the topic of the utterance, the creation unit 53 creates a prompt that includes the current utterance and the inferred topic. Also, if the topic of the utterance is clear, or if the topic of the utterance is unclear and cannot be inferred, the creation unit 53 creates a prompt based on the current utterance.

[0045] Here, using Figure 6, we will explain the process by which the inference unit 52 infers the topic, creates a prompt that includes the inferred topic in the current utterance, and responds to the user. Figure 6 is a diagram showing the process of creating a prompt and responding to the user.

[0046] Specifically, the creation unit 53 creates a prompt to be input to the second trained language model 64, which will respond to the current utterance. For example, as shown in Figure 6E, the creation unit 53 creates a prompt for the second trained language model 64 that includes the topic inferred from the current utterance, "User: What is the stock price of Company B?". Next, the creation unit 53 inputs the created prompt into the second trained language model 64.

[0047] Next, as shown in Figure 6, the second trained language model 64, following the prompt, obtains the stock price of Company B from the internet or other source, generates the response "The current stock price of Company B is 8,208 yen," and outputs it to the output unit 4. The output unit 4 responds to the user's utterance by outputting "The current stock price of Company B is 8,208 yen" in voice. The second trained language model 64 may also display "The current stock price of Company B is 8,208 yen" as text on the in-vehicle display, either along with or instead of voice.

[0048] (Processing flow of Dialogue Device 1) Next, an example of the processing procedure by the dialogue device 1 according to Embodiment 1 will be described using Figure 7. Figure 7 is a flowchart showing an example of the dialogue processing flow of the dialogue device 1 according to Embodiment 1. Note that each step in the flowchart shown in Figure 7 can be executed in a different order, and additional or omitted processes may be included.

[0049] First, the dialogue device 1 detects the user's utterance (S11). Next, the dialogue device 1 determines whether the topic of the utterance detected is clear or not (S12). If the topic of the utterance detected is not clear (S12; NO), the dialogue device 1 infers the topic of the utterance detected from the utterance data 61 (S13). If the topic of the utterance detected is clear (S12; YES), the dialogue device 1 creates a prompt based on the utterance detected (S15).

[0050] If the dialogue device 1 can guess the topic of the utterance it has detected (S13; YES), it creates a prompt that includes the utterance it has detected and the guessed topic (S14). If the dialogue device 1 cannot guess the topic of the utterance it has detected (S13; NO), it creates a prompt based on the utterance it has detected (S15).

[0051] The dialogue device 1 inputs the created prompt into the second trained language model 64 (S16). The dialogue device 1 responds to the user based on the response created by the second trained language model 64 (S17), and then terminates the process.

[0052] (effect) According to the embodiment 1 described above, the following effects are achieved. When the topic of an utterance is unclear, the dialogue device 1 infers the topic from a summary or tags of the user's past utterances, or both, and creates a prompt that includes the inferred topic. As a result, even for utterances with unclear topics, the dialogue device 1 can provide a response that aligns with the user's intent, rather than a general response.

[0053] By inferring the topic of the utterance, the dialogue device 1 can maintain or improve the quality of prompts without including all of the user's past utterances in the prompts, thereby reducing the amount of prompts. Consequently, the dialogue device 1 can shorten the response time of the trained language model, improving user convenience. Furthermore, the dialogue device 1 can reduce the usage fees of the trained language model.

[0054] [Embodiment 2] In Embodiment 1, when the topic of the user's utterance is unclear, the process of inferring the topic and, if successful, creating a prompt that includes the inferred topic in the utterance was described. However, the disclosed dialogue device 1 can ask the user whether the inferred topic is appropriate or not.

[0055] Therefore, Embodiment 2 describes an example in which, when the topic of the user's utterance is unclear, the device infers the topic, and if it is able to infer it, confirms the inferred topic with the user before creating a prompt. As a result, the dialogue device 1 of Embodiment 2 can provide a more accurate response to the user's utterance.

[0056] Note that a detailed explanation of the dialogue device 1 according to Embodiment 2, which has the same configuration and processing as the dialogue device 1 according to Embodiment 1, will be omitted. Below, the creation unit 53, which has a different configuration and processing content from Embodiment 1, will be described in detail.

[0057] The creation unit 53 asks the user a confirmation question if the inference unit 52 has been able to infer the topic of the user's utterance. The creation unit 53 then receives the user's answer to the question, creates a prompt including the utterance, the question, and the answer to the question, and inputs it into the second trained language model 64.

[0058] Here, using Figures 8 and 9, we will explain the process of confirming the guessed topic and handling cases where the answer is positive. Figure 8 is a diagram showing the process of confirming the guessed topic with the user, and Figure 9 is a diagram showing the process of creating a prompt including the user's utterance, question, and answer and then responding.

[0059] Figure 8 shows the process of confirming the predicted topic with the user, based on the prediction results from Figures 4 and 5.

[0060] For example, as shown in Figure 8, first, the creation unit 53 creates a question, "Are you asking about the stock price of Company B?", which includes the user's utterance, "What about Company B?", and the stock price estimated by the estimation unit 52, and sends the question to the user through the output unit 4. When the user answers "yes" to the question, the creation unit 53 receives the user's answer "yes" through the voice input unit 3.

[0061] Next, as shown in Figure 9F, the creation unit 53 creates a prompt that includes the utterance detected this time, "User: What about Company B?", the question from the creation unit 53 to the user, "System: Are you talking about Company B's stock price?", and the user's answer, "User: Yes", in response to the question "Q: Please answer based on the following conversation" which instructs the second trained language model 64. The creation unit 53 then inputs the created prompt into the second trained language model 64.

[0062] Next, as shown in Figure 9, the second trained language model 64, following the prompt's instructions, obtains the stock price of Company B from the internet or other sources, and generates the response, "The current stock price of Company B is 8,208 yen," which it outputs to the output unit 4. The output unit 4 then responds to the user's current utterance.

[0063] Next, using Figures 10 and 11, we will explain how to confirm the guessed topic and what to do if the answer is negative. Figure 10 is a diagram showing the process of confirming the guessed topic with the user, and Figure 11 is a diagram showing the process of creating a prompt that includes the user's utterance, question, and answer, and then responding.

[0064] Similar to Figure 8, Figure 10 shows the process of confirming the inferred topic with the user, based on the inference results from Figures 4 and 5.

[0065] For example, as shown in Figure 10, the creation unit 53 creates a question, "Are you asking about the stock price of Company B?", which includes the user's utterance, "What about Company B?", and the "stock price" estimated by the estimation unit 52, and sends the question to the user through the output unit 4. When the user answers "No" to the question, the creation unit 53 receives the user's answer "No" through the voice input unit 3.

[0066] Next, as shown in Figure 11G, the creation unit 53 creates a prompt that includes the utterance detected this time, "User: What about Company B?", the question from the creation unit 53 to the user, "System: Are you talking about Company B's stock price?", and the user's answer, "User: No", in response to the question "Q: Please answer based on the following conversation" which instructs the second trained language model 64. Then, the creation unit 53 inputs the created prompt into the second trained language model 64.

[0067] Next, the second trained language model 64 generates a general response because the prompt contains information that denies the stock price of Company B, but does not specify what the user wants to know about Company B. As shown in Figure 11, the second trained language model 64 obtains general information about Company B, such as a company profile, from the internet, and generates the response, "Company B is a globally renowned Japanese game company...", which it outputs to the output unit 4. The output unit 4 then responds to the user's current utterance.

[0068] Next, using Figures 12 and 13, we will explain the process of confirming the guessed topic and handling cases where the answer indicates a different topic. Figure 12 is a diagram showing the process of confirming the guessed topic with the user, and Figure 13 is a diagram showing the process of creating a prompt including the user's utterance, question, and answer and then responding.

[0069] Similar to Figure 8, Figure 12 shows the process of confirming the predicted topic with the user, based on the prediction results from Figures 4 and 5.

[0070] For example, as shown in Figure 12, the creation unit 53 creates a question, "Are you asking about the stock price of Company B?", which includes the user's utterance, "What about Company B?", and the "stock price" estimated by the estimation unit 52, and sends the question to the user through the output unit 4. When the user responds to the question with "I want to know about recently released game software," the creation unit 53 receives the user's response, "I want to know about recently released game software," through the voice input unit 3.

[0071] Next, as shown in Figure 13, the creation unit 53 creates a prompt that includes the utterance detected this time, "User: What about Company B?", the question from the creation unit 53 to the user, "System: Are you talking about Company B's stock price?", and the user's answer, "User: A recently released video game." The creation unit 53 then inputs the created prompt into the second trained language model 64.

[0072] Next, the second trained language model 64 creates a response that includes the information identified in the prompt, which includes "recently released game software," indicating what the user wants to know about Company B. As shown in Figure 13, the second trained language model 64 obtains information about Company B's recently released game software from the internet, etc., and creates a response saying, "Recently, Company B has released the following titles..." and outputs it to the output unit 4. The output unit 4 then responds to the user's current utterance.

[0073] (Processing flow of Dialogue Device 1) Next, an example of the processing procedure by the dialogue device 1 according to Embodiment 2 will be described using Figure 14. Figure 14 is a flowchart showing an example of the dialogue processing flow of the dialogue device 1 according to Embodiment 2. Note that each step in the flowchart shown in Figure 14 can be executed in a different order, and additional or omitted processes may be included.

[0074] First, the dialogue device 1 detects the user's utterance (S21). Next, the dialogue device 1 determines whether the topic of the utterance detected is clear or not (S22). If the topic of the utterance detected is not clear (S22; NO), the dialogue device 1 infers the topic of the utterance detected from the utterance data 61 (S23). If the topic of the utterance detected is clear (S22; YES), the dialogue device 1 creates a prompt based on the utterance detected (S26).

[0075] If the dialogue device 1 can guess the topic of the utterance it has detected (S23; YES), it asks the user whether the guessed topic is what the user intended (S24). If the dialogue device 1 receives a response from the user to the question confirming whether the guessed topic is what the user intended (S24; YES), it creates a prompt that includes the utterance it has detected, the question, and the answer to the question (S25). If the dialogue device 1 cannot guess the topic of the utterance it has detected (S23; NO), or if it does not receive a response from the user to the question confirming whether the guessed topic is what the user intended (S24; NO), it creates a prompt based on the utterance it has detected (S26).

[0076] The dialogue device 1 inputs the created prompt into the second trained language model 64 (S27). The dialogue device 1 responds to the user based on the response created by the second trained language model 64 (S28), and then terminates the process.

[0077] Thus, the dialogue device 1 of Embodiment 2 includes a process to confirm the inferred topic with the user, enabling it to respond to utterances where the user's topic is ambiguous, while still understanding the user's intent.

[0078] [Embodiment 3] Incidentally, the disclosed dialogue device 1 can engage in dialogue based on the most recent conversation by inferring from the latest summary when there are multiple summaries related to the user's utterance. Specifically, when there are multiple summaries or tags related to the user's utterance, the dialogue device 1 infers the subject of the user's utterance from the latest summary or tag, the summaries, and the tags.

[0079] Therefore, Embodiment 3 describes the process of inferring from the latest summary when there are multiple summaries related to the user's utterance. Note that a detailed explanation of the configuration and processing of the dialogue device 1 according to Embodiment 3 is omitted as it is the same as that of the dialogue device 1 according to Embodiment 1. Below, the inference unit 52, which differs from that of Embodiment 1, will be described in detail.

[0080] The inference unit 52 uses the first trained language model 63 to infer the topic of the current utterance if there are multiple summaries or tags associated with the utterance data 61.

[0081] Figure 15 shows an example of the process of inferring from the latest summary or tags. Note that the "#" in "#Company A #Stock Price..." in Figure 15 indicates a tag.

[0082] The inference unit 52 creates a prompt containing wording that instructs the first trained language model 63 to infer the topic from the most recent summary if the utterance data 61 has multiple related summaries or tags. The inference unit 52 then inputs the created prompt into the first trained language model 63.

[0083] For example, as shown in Figure 15, I, the speech data 61 has multiple related summaries or tags. In this case, the first trained language model 63 infers the topic of the current utterance from the most recent summary, which is the newest summary from the date and time the utterance was detected. For example, as shown in Figure 15, the first trained language model 63 infers the topic of the current utterance from the most recent summary, "No.1 10 / 17 14:13 About the stock price of Company A...".

[0084] As described above, the dialogue device 1 can provide a response based on the most recent utterance by inferring the topic from the most recent summary related to the user's utterance. For example, the process described in Embodiment 3 is particularly useful when the user and the dialogue device 1 are having a continuous conversation.

[0085] [Embodiment 4] Embodiment 1 described an example in which the dialogue device 1 infers the topic from relevant summaries or tags, summaries and tags, included in the speech data 61. However, depending on the speaker and the speech environment, there may be summaries or tags that are relevant to the speech content but should not be used for inference. Therefore, Embodiment 4 describes a process in which the dialogue device 1 includes a speech environment indicating the environment at the time of utterance in the speech data 61, and infers the topic of the utterance from summaries or tags, summaries and tags, of the speech environment common to the current utterance.

[0086] Note that a detailed explanation of the configuration and processing of the dialogue device 1 according to Embodiment 4, which is the same as that of the dialogue device 1 according to Embodiment 1, will be omitted. Below, the speech data 61 and the prediction unit 52, which differ from those in Embodiment 1, will be described in detail.

[0087] The speech data 61 stores summaries of utterances previously made by the user. In addition to the "date and time," "summary," and "tags" described in the embodiment, the speech data 61 also includes information associated with the "speech environment." Here, the speech environment records the in-car environment in which the user is speaking.

[0088] For example, speech data 61 stores a summary of what the user said to the dialogue device 1 at the date and time "10 / 17 14:13": "Regarding the stock price of Company A...", and the tag "Company A...". Furthermore, the speech data 61 is processed by the voice input unit 3 of the dialogue device 1 to determine the user who is the conversation partner based on the type of voice, and information such as the number of users and the user's attributes (father, mother) is recorded. For example, if the father was the only passenger in the car when the user was speaking, the number of passengers "1" and the passenger's attribute "father" will be recorded as the speaking environment. The user may also set the passenger user to the in-vehicle device 1 when they get in the car. Alternatively, the speaking user may be estimated by general voice analysis, etc.

[0089] Figure 16 shows the process for obtaining summaries that are common to the speech environment. For example, as shown in J in Figure 16, for the current utterance, the speech environment is "1 crew member, attribute: father," and the utterance content is "What about company B?".

[0090] The inference unit 52 creates a prompt that instructs the first trained language model 63 to infer the topic of the current utterance from the speech data 61, based on summaries or tags that are related to the current utterance and are common to the current utterance and the speech environment. The inference unit 52 inputs the created prompt into the first trained language model 63.

[0091] Then, as shown in Figure 16K, the first trained language model 63 infers the subject of the current utterance from the summary "No.1 10 / 17 14:13 About the stock price of Company A..." which contains a speech environment common to the speech environment of the current utterance, "1 crew member, attribute: father," among the relevant summaries or tags.

[0092] As described above, in the dialogue device 1 according to Embodiment 4, by including the speech environment in the speech data 61, it becomes possible to infer the topic of the user's utterance based on summaries or similar information that are in a similar situation. This makes it possible to infer the topic of an utterance from summaries or similar information of utterances by the same person, and to provide an even more accurate response to the user's utterance.

[0093] [Embodiment 5] Now, while embodiments of this disclosure have been described, this disclosure may be implemented in various other forms besides those described above.

[0094] (Examples of application) By the way, although the above embodiment was described using an example where the dialogue device 1 is an in-vehicle device, it is not limited to this. The dialogue device 1 can be used not only in in-vehicle devices but also in various other devices that can interact with a user. Here, we will illustrate some examples of applications for the dialogue device 1.

[0095] Figure 17 shows an example of the application of the dialogue device 1. For example, as shown in Figure 17(1), the dialogue device 1 can be used with a smart speaker. In this case, the user activates a smart speaker that has the same functions as the dialogue device 1 and speaks to the smart speaker. Upon receiving this speech, the smart speaker executes the processes described in the above embodiment and responds to the user's speech.

[0096] Furthermore, as shown in Figure 17(2), for example, the dialogue device 1 can be used as a dialogue application for a smartphone. In this case, the user launches a dialogue application with functions equivalent to those of the dialogue device 1 and speaks to the smartphone. The dialogue application, upon receiving this utterance, executes the processes described in the above embodiment and responds to the user's utterance. Note that the dialogue device 1 is not limited to smartphones; it can also be applied to mobile phones, tablets, and other devices.

[0097] Furthermore, as shown in Figure 17(3), for example, it can be used as the dialogue device 1 in a personal computer. In this case, the user launches dialogue software or the like that has the same functionality as the dialogue device 1 and speaks to the personal computer. The dialogue software or the like that receives this utterance executes the processes described in the above embodiment and responds to the user's utterance. Note that the dialogue device 1 is not limited to personal computers, but can be similarly applied to various computers such as game consoles with internet communication capabilities.

[0098] (modified version) The trained language models 62, 63, and 64 for judgment related to the dialogue device 1 can determine the content of an instruction from the context and respond, even without inputting the instruction "Q:".

[0099] The inference unit 52 of the dialogue device 1 can also use the first trained language model 63 to infer the representative tag most relevant to the current utterance from the tags included in the utterance data 61. In this case, the creation unit 53 asks the user a confirmation question about the representative tag inferred by the inference unit 52. The creation unit 53 then receives the user's answer to the question, creates a prompt including the current utterance, the question, and the answer to the question, and inputs it into the second trained language model 64.

[0100] For example, the inference unit 52 creates a prompt by adding the detected utterance "What is Company B?" to the question "Please obtain a representative tag related to this utterance from the utterance data 61," which is an instruction to the first trained language model 63, and inputs it into the first trained language model 63. The inference unit 52 infers "stock price," which is the output result of the first trained language model 63, as the representative tag.

[0101] Next, the creation unit 53 creates a question, "Are you talking about stock prices?", which includes the representative tag "stock price", and confirms it with the user through the output unit 4. The creation unit 53 receives the user's answer, "Yes", through the voice input unit 3. Then, the creation unit 53 creates a prompt including the current utterance, "What about company B?", the question, "Are you talking about stock prices?", and the answer to the question, "Yes", and inputs it into the second trained language model 64.

[0102] [Hardware configuration] Furthermore, the dialogue device 1 according to Embodiments 1 to 5 described above is realized by a computer 1000 having a configuration such as that shown in Figure 18. The following explanation will use the dialogue device 1 as an example. Figure 18 is a hardware configuration diagram showing an example of a computer that realizes the functions of the dialogue device 1 of Embodiments 1 to 5. The computer 1000 has a CPU 1100, RAM 1200, ROM 1300, HDD 1400, communication interface (I / F) 1500, input / output interface (I / F) 1600, and media interface (I / F) 1700.

[0103] The CPU 1100 operates based on programs stored in the ROM 1300 or HDD 1400, controlling various components. The ROM 1300 stores boot programs executed by the CPU 1100 when the computer 1000 starts up, as well as programs that depend on the computer 1000's hardware.

[0104] The HDD1400 stores programs executed by the CPU1100, as well as data used by such programs. The communication interface1500 receives data from other devices via a predetermined communication network and sends it to the CPU1100, and transmits data generated by the CPU1100 to other devices via the predetermined communication network.

[0105] The CPU 1100 controls output devices such as displays and printers, and input devices such as keyboards and mice, via the input / output interface 1600. The CPU 1100 acquires data from input devices via the input / output interface 1600. The CPU 1100 also outputs the generated data to output devices via the input / output interface 1600.

[0106] The media interface 1700 reads a program or data stored in the recording medium 1800 and provides it to the CPU 1100 via the RAM 1200. The CPU 1100 loads the program from the recording medium 1800 onto the RAM 1200 via the media interface 1700 and executes the loaded program. The recording medium 1800 is, for example, an optical recording medium such as a DVD (Digital Versatile Disc) or PD (Phase Change Rewritable Disk), a magneto-optical recording medium such as an MO (Magneto-Optical disk), a tape medium, a magnetic recording medium, or a semiconductor memory.

[0107] For example, when computer 1000 functions as an interactive device 1 according to Embodiments 1 to 5, the CPU 1100 of computer 1000 realizes the functions included in the control unit 5 of the interactive device 1 by executing programs loaded on RAM 1200. The CPU 1100 of computer 1000 reads and executes these programs from the recording medium 1800, but as another example, these programs may be obtained from other devices via a predetermined communication network.

[0108] 〔others〕 Furthermore, among the processes described in each of the above 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 by known methods. In addition, the processing procedures, specific names, and information including various data and parameters shown in the above document and drawings can be changed at will unless otherwise specified. For example, the various information shown in each figure is not limited to the information shown.

[0109] Furthermore, the components of each illustrated device are functionally conceptual and do not necessarily need to be physically configured as shown. In other words, the specific forms of distribution and integration of each device are not limited to those shown, and all or part of them can be functionally or physically distributed and integrated in any unit according to various loads and usage conditions.

[0110] Furthermore, the above embodiments can be combined as appropriate, provided that the processing content is not contradictory.

[0111] Although some embodiments of the present invention have been described in detail above with reference to the drawings, these are illustrative examples, and the present disclosure can be implemented in various modified and improved forms based on the knowledge of those skilled in the art, starting with the embodiments described in the disclosure section.

[0112] Furthermore, the terms "section," "module," and "unit" mentioned above can be replaced with "means" or "circuits." For example, the inference unit can be replaced with inference means or inference circuit. [Explanation of Symbols]

[0113] 1. Dialogue device 2 Communications Department 3. Voice input section 4 Output section 5. Control Unit 51 Judgment section 52 Guessing part 53 Creation Section 6 Memory section 61 Utterance Data 62 Pre-trained language models for decision-making 63. First pre-trained language model 64. Second pre-trained language model

Claims

1. When a user's utterance is detected, an inference unit uses a first trained language model to infer the topic of the current utterance from utterance data that includes at least summaries or tags created in relation to the user's past utterances, and If the prediction unit is able to predict the topic, it creates a prompt to input the topic into the current utterance and into the second trained language model; if the prediction unit is unable to predict the topic, it creates a prompt based on the current utterance. A dialogue device characterized by comprising the following features.

2. When a user's utterance is detected, an inference unit uses a first trained language model to infer the topic of the current utterance from utterance data that includes at least summaries or tags created in relation to the user's past utterances, and If the prediction unit has been able to predict at least one of the themes, the creation unit asks the user whether the predicted at least one of the themes is the theme the user intended, and if a response is received from the user, the creation unit creates a prompt to input the current utterance, the question, and the answer into a second trained language model. A dialogue device characterized by comprising the following features.

3. The dialogue device according to claim 1 or 2, wherein, if the utterance data has a plurality of related summaries or tags, the inference unit infers the topic from the most recent related summary or tag among the plurality of related summaries using the first trained language model.

4. The aforementioned speech data includes information about the speech environment in which the user previously made speech, The dialogue device according to claim 1 or 2, characterized in that the inference unit infers the topic using the first trained language model from the related summaries or tags that are common to the information regarding the speech environment of the current utterance and the information regarding the speech environment in which utterances were made in the past.

5. A method of dialogue performed by a dialogue device, When a user's utterance is detected, an inference step is performed using a first trained language model to infer the topic of the current utterance from utterance data that includes at least summaries or tags created in relation to the user's past utterances; If the subject can be inferred by the inference step, a prompt is created to input the subject into the second trained language model, and if the subject cannot be inferred by the inference step, a creation step is made to create the prompt based on the current utterance. A dialogue method characterized by including

6. A method of dialogue performed by a dialogue device, When a user's utterance is detected, an inference step is performed using a first trained language model to infer the topic of the current utterance from utterance data that includes at least summaries or tags created in relation to the user's past utterances; If at least one topic can be inferred through the inference step, the creation step involves asking the user whether the inferred at least one topic is the topic the user intends to discuss, and if a response is received from the user, creating a prompt to input into a second trained language model, including the current utterance, the question, and the answer. A dialogue method characterized by including

7. When a user's utterance is detected, an inference procedure is performed using a first trained language model to infer the topic of the current utterance from utterance data that includes at least summaries or tags created in relation to the user's past utterances, and If the subject can be inferred by the inference procedure, a prompt is created to input the subject into the second trained language model, and if the subject cannot be inferred by the inference procedure, a creation procedure is provided to create the prompt based on the current utterance. An interactive program characterized by causing a computer to execute something.

8. When a user's utterance is detected, an inference procedure is performed using a first trained language model to infer the topic of the current utterance from utterance data that includes at least summaries or tags created in relation to the user's past utterances, and If at least one of the themes can be inferred by the inference procedure, the procedure involves asking the user whether the inferred at least one of the themes is the theme the user intends to discuss, and if a response is received from the user, the procedure involves creating a prompt to input into a second trained language model, including the current utterance, the question, and the answer. An interactive program characterized by causing a computer to execute something.

Images