Response generation method

Abstract

To improve response times in voice dialogue systems. [Solution] The response generation method includes the steps of: performing speech recognition; obtaining a first object from RAG using the result of speech recognition as a query; creating a response by the generation AI using the first object as a prompt; estimating the degree of relevance between the response and the first object; deleting first objects from the storage medium if the estimated degree of relevance is less than a predetermined threshold value, and designating first objects that are equal to or greater than the threshold value as second objects; inputting the response as a query into RAG to obtain a third object; and deleting the duplicate portions of the second and third objects from the storage medium. The contents stored in the storage medium with the duplicate portions of the second and third objects deleted are then carried over to the first object in the next dialogue.

Description

Technical Field

[0001] The present disclosure relates to a response generation method for a spoken dialogue system using generative AI (Artificial Intelligence).

Background Art

[0002] In recent years, spoken dialogue systems using generative AI have come into widespread use. For example, in a spoken dialogue system using such generative AI, RAG (Retrieval Augmented Generation), a natural language processing technology that combines a retrieval function with generative AI, is used as a method for generating responses based on specific knowledge.

[0003] In this spoken dialogue system, an utterance from a user is used as a retrieval query, a plurality of pieces of highly relevant information (objects) are retrieved from RAG in order of relevance, and the retrieved information is embedded in a prompt for generative AI, thereby enabling a response to be generated.

[0004] Patent Document 1 discloses that retrieval performance is improved by expanding an input query using a machine learning model for each document domain and performing retrieval while also including related terms.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Summary of the Invention

Problem to be Solved by the Invention

[0006] One of the challenges with RAG is its inability to consider dialogue history. Specifically, in cases of consecutive dialogues on the same topic, questions using demonstrative pronouns or lacking subjects cannot be answered correctly, preventing the generative AI from providing a response. One approach is to leverage the generative AI's ability to retain dialogue history to pre-modify RAG's search queries and have it infer keywords, thereby obtaining correct search results. However, this requires using the generative AI multiple times for a single response, resulting in a delay in the response.

[0007] This disclosure provides a response generation method that improves the response in a voice dialogue system. [Means for solving the problem]

[0008] The response generation method according to this disclosure comprises the steps of: speech recognition of speech uttered by a user; inputting the result of the speech recognition as a query to obtain a first object from RAG; creating a response by a generation AI using the first object as a prompt; storing the first object in a storage medium; estimating the degree of relevance between the created response and the first object; deleting the first object from the storage medium if the estimated degree of relevance is less than a predetermined threshold value, and storing the first object that is equal to or greater than the threshold value as a second object in the storage medium; inputting the response as a query to RAG to obtain a third object and storing it in the storage medium; and deleting any overlaps between the second object and the third object from the storage medium, wherein the contents of the second object and the third object stored in the storage medium with the overlaps removed are carried over to the first object in the next dialogue. This allows for response generation that takes dialogue history into account, without having to use the generation AI multiple times. [Effects of the Invention]

[0009] This disclosure provides a response generation method that improves the response in a voice dialogue system. [Brief explanation of the drawing]

[0010] [Figure 1] This diagram shows the configuration of the voice dialogue system related to this disclosure. [Figure 2] This flowchart illustrates an example of the operation of the voice dialogue system described herein. [Modes for carrying out the invention]

[0011] The response generation method for the voice dialogue system according to this embodiment will be described below with reference to the drawings. Figure 1 is a diagram showing an example of the configuration of the voice dialogue system 1.

[0012] The voice dialogue system 1 comprises a computer 11, a microphone 12, and a speaker 13.

[0013] Computer 11 typically uses a main memory, auxiliary memory, and arithmetic logic unit to perform calculations and control operations, and to produce output corresponding to the input. Typically, computer 11 is also connected to a computer network and can send and receive information to and from a server device (not shown). Hereafter, main memory is a temporary storage medium and may be referred to as temporary memory.

[0014] Computer 11 can perform speech recognition of the audio input from microphone 12, search RAG, generate answers using a generation AI, and convert the generated answers into audio data.

[0015] Furthermore, the processes of speech recognition, RAG search, response generation by the AI, and conversion of the generated responses into speech can be performed on the server device. In this case, computer 11 can send information for executing the processes to the server device and receive the processing results from the server device.

[0016] The microphone 12 receives audio input from a user's utterance. Typically, the microphone 12 is wired-connected to the computer 11, and the audio input via the microphone 12 is transmitted to the computer 11.

[0017] The speaker 13 can output audio converted from audio data by the computer 11. This allows the user to hear a voice response via the speaker 13.

[0018] In this manner, the voice interactive system 1 enables voice interaction by receiving a user's voice input via the microphone 12, generating response content by the computer 11, and outputting the response voice via the speaker 13.

[0019] Next, a response generation method in the voice interactive system 1 will be described with reference to FIG. 2.

[0020] First, voice recognition is performed, and the result is obtained as text (step S1). That is, the user's utterance is acquired by the microphone 12, and the result obtained through voice recognition by the computer 11 is converted into text. Note that the computer 11 can store the resulting text in a temporary memory.

[0021] The computer 11 inputs the resulting text as a search query to RAG, and acquires a plurality of highly relevant search result objects (step S2). At this time, the computer 11 may use relevance rescoring of results through Re-Rank. In this case, although the computer 11 improves the accuracy of relevance measurement, a slight delay will occur. The computer 11 additionally stores the plurality of acquired search result objects in the temporary memory. The object stored in this temporary memory is defined as a first object.

[0022] The computer 11 embeds and adds the object stored in the temporary memory into the prompt (step S3). That is, the computer 11 uses the first object stored in the temporary memory in step S1 and step S2 as the prompt. At this time, the inherited search result to be described later is also combined and embedded into the prompt.

[0023] The computer 11 inputs the content of step S3 to the generative AI and generates an answer (step S4). In other words, the computer 11 creates an answer generated by the generative AI using the first object as the prompt. At this time, the computer 11 can create an answer by the generative AI by adding a combination of the inherited search result to be described later to the prompt.

[0024] The computer 11 utters the answer generated by the generative AI in step S4 as a voice (step S5). That is, the computer 11 delivers the answer to the user by voice via the speaker 13.

[0025] In parallel with step S5, the computer 11 estimates the relevance between the answer generated in step S4 and each object input in step S3 (step S6). Here, the relevance estimation can be performed by using an existing Re-Rank model or the like.

[0026] The computer 11 deletes, from the temporary memory, objects whose relevance score obtained in step S6 is equal to or lower than an arbitrary threshold (step S7). At this time, the computer 11 may predetermine the maximum number of objects to be inherited, and if more than the maximum number of objects remain in the temporary memory, delete the objects in order from the one with the lowest score.

[0027] In this way, the computer 11 estimates the degree of relevance between the generated response and each first object. First objects whose estimated degree of relevance is less than a predetermined threshold are deleted from the storage medium, while first objects whose degree of relevance is equal to or greater than the threshold are stored in temporary memory as second objects.

[0028] On the other hand, computer 11 inputs the response of the generating AI in step S4 as a search query into the same RAG as in step S2, and obtains multiple objects of highly relevant search results (step S8). At this time, computer 11 may use rescoring of the relevance of the results by Re-Rank. Computer 11 improves the accuracy of relevance, but it will cause a slight delay. Computer 11 adds the multiple obtained search result objects to temporary memory for storage. These objects are referred to as the third object.

[0029] If, in the results of steps S7 and S8, there are duplicate objects, the computer 11 deletes one of them from temporary memory (step S9).

[0030] Computer 11 inherits the objects in temporary memory for subsequent conversations (step S10). This allows computer 11 to combine the objects remaining in temporary memory in step S9 with the search results from step S2 to execute step S3 in subsequent responses.

[0031] In other words, computer 11 can carry over the contents of the second and third objects, which are stored in temporary memory with the duplicates removed, to the first object in the next interaction.

[0032] Through the above process, the voice dialogue system 1 can generate a response for subsequent conversations by the user using the object related to the current question sentence in step S2, the object that appears to have been used in the previous answer in steps S6 and S7, and the object that appears to be newly related from the previous answer in step S8.

[0033] In this way, the response generation method of the voice dialogue system 1 described herein allows the generating AI to understand the context from the dialogue history and provide an answer based on the inherited knowledge, even if the subject is missing or demonstrative pronouns are used on the same topic. Furthermore, the object will be subjected to relevance estimation in subsequent loops, and even if the topic is repeated three or four times in a row, the system can continue to provide answers while inheriting relevant objects. In other words, the voice dialogue system 1 can provide an accurate answer based on this knowledge and the dialogue history held by the generating AI, even if the question contains missing subjects or demonstrative pronouns.

[0034] On the other hand, if the topic changes, the response generation method can generate an answer based on the results of a new search. In that case, objects inherited from the previous conversation are not inherited because their relevance to the current answer decreases, and in subsequent conversations, the objects newly added in the current conversation will be carried over.

[0035] As a result, the voice dialogue system 1 can respond to both continuous topics and topic changes using appropriate knowledge objects.

[0036] Furthermore, in the voice dialogue system 1 using the response generation method described herein, preparation for the next response is performed simultaneously with the speech utterance. Therefore, compared to a system that does not introduce the response generation method described herein, the quality of the response can be improved without any factors that increase the delay until the response is generated. In particular, in the voice dialogue system 1, the processing from step S6 onward in Figure 2 can be processed in parallel with the speech utterance of the response generation word, thus avoiding delays until the response is uttered.

[0037] Based on the above, the voice dialogue system 1 can accurately answer questions by carrying over knowledge related to the user's questions and answers to the next question, based on the carried-over knowledge and the dialogue history of the generating AI. In this case, since the voice dialogue system 1 does not use the generating AI multiple times, it can suppress the delayed speech of answers.

[0038] It should be noted that the present invention is not limited to the embodiments described above, and can be modified as appropriate without departing from the spirit of the invention. In other words, the above description has been omitted and simplified as appropriate for the sake of clarity, and those skilled in the art can easily change, add, and modify each element of the embodiments within the scope of the present invention.

[0039] In Figure 2, the voice dialogue system 1 executes steps S6 and S7 and step S8 in parallel and employs both, but it is also possible to use only one of them.

[0040] Furthermore, depending on the data structure within RAG, the output of steps S6 and S7 may be almost identical to the output of step S8. This can occur, for example, when topics across objects have little to no relation to each other and are independent of one another. In such cases, even if a search is performed in step S8 based on the generated AI's answer, no related objects may be found other than those used in past conversations inherited in steps S6 and S7.

[0041] In such cases, the speech dialogue system 1 can simplify its algorithm by adopting either the output of steps S6 and S7 or the output of step S8.

[0042] Furthermore, while the above explanation described the use of main memory as a temporary storage medium, this is not the only possible use. [Explanation of Symbols]

[0043] 1. Voice Dialogue System 11 Computer 12 microphones 13 speakers

Claims

[Claim 1] The steps include: a step of speech recognition of the voice spoken by the user, The steps include: inputting the result of the aforementioned speech recognition as a query and obtaining a first object from RAG; The steps include: creating an AI response using the first object as a prompt; The steps include saving the first object to a storage medium, The steps include: estimating the degree of relevance between the created answer and the first object; The steps include: deleting the first objects from the storage medium whose estimated correlation is less than a predetermined threshold value, and storing the first objects whose correlation is equal to or greater than the threshold value as second objects in the storage medium; The steps include: inputting the aforementioned answer as a query into RAG to obtain a third object and storing it in the storage medium; The process includes the step of deleting the duplicate portions of the second object and the third object from the storage medium, The contents of the second and third objects, stored on the storage medium with duplicates removed, are passed on to the first object in the next interaction. Response generation method.

Images