Session data processing method, device and equipment based on large model

By using the speech search model to determine the target speech in the session data processing, the problem of low participation in the conversation model in the task phase is solved, and more efficient conversation data processing and response time is achieved.

CN119025635APending Publication Date: 2024-11-26BEIJING WATERDROP TECH GRP CO LTD
View PDF 8 Cites 0 Cited by

Patent Information

Application Number
CN202410929274.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-11
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

In the prior art, the conversation model has low participation in the task stage of session data processing, and it is difficult to achieve the expected task effect.

Method used

By obtaining the output session text during user interaction, the speech search model is used to determine the target speech in the predetermined speech library. The target speech is the standard speech with the highest similarity to the output session text, and its corresponding speech stream is used as The session interaction results are sent to the user, avoid frequent calls to the audio translation model and improve response time.

Benefits of technology

It improves the task phase participation of session data processing, reduces frequent calls to voice translation, and improves the response time and intelligence of session data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119025635A_ABST
    Figure CN119025635A_ABST
Patent Text Reader

Abstract

The invention discloses a session data processing method, device and equipment based on a large model, relates to the technical field of intelligent sessions, and aims to generate a voice stream directly participating in a task stage, avoid frequent calling of voice translation in a session data processing process and prolong the response time of session data. The method comprises the steps that an output session text of a user in the interaction process is obtained, a verbal skill retrieval model is used for conducting retrieval processing on the output session text so as to determine a target verbal skill in a preset verbal skill library, standard verbal skill obtained through coding mapping of voice streams in different fields is recorded in the preset verbal skill library, and the verbal skill retrieval model is used for retrieving the target verbal skill. The target verbal skill is a standard verbal skill with the highest similarity between the text retrieval and the output session text, and sending the voice stream corresponding to the target verbal skill to the user as a session interaction result.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of intelligent conversation technology, and in particular to a conversation data processing method, device and equipment based on a large model. Background Art

[0002] With the continuous development of intelligent conversation technology, conversational robots are increasingly entering people's daily lives as intelligent question-and-answer customer service, guiding users to perform a series of operations to translate the conversation content during the interaction process into text and apply it to the processing of conversation data, so as to drive the conversation model to realize the service functions of intelligent question-and-answer customer service.

[0003] In related technologies, the conversation model can be used as an intelligent Q&A customer service in the process of processing conversation data, and then the platform can assign clues to the intelligent Q&A customer service. The intelligent Q&A customer service will then call designated customers according to the clues, introduce product details, and transfer calls from customers with higher intentions to real agents. The agents will then provide services based on the task requirements and guide users to complete the service on the platform. However, considering that the conversation model is mainly involved in the initial access stage during the entire conversation data processing process, and its participation in the task stage is relatively low, it is difficult to achieve the expected task effect. Summary of the invention

[0004] In view of this, the present application provides a conversation data processing method, device and equipment based on a large model, the main purpose of which is to solve the problem that the conversation data processing process in the prior art has low participation in the task stage and it is difficult to achieve the expected task effect.

[0005] According to a first aspect of the present application, a session data processing method based on a large model is provided, comprising: Get the user's output conversation text during the interaction process; The output conversation text is searched and processed using a speech retrieval model to determine a target speech in a predetermined speech library, wherein the predetermined speech library records standard speech obtained by encoding and mapping speech streams in different fields, and the target speech is the standard speech in the predetermined speech library with the highest similarity to the output conversation text; The voice stream corresponding to the target speech is sent to the user as a conversation interaction result.

[0006] Furthermore, obtaining the output conversation text of the user during the interaction process includes: Get the user's input conversation text during the interaction process; A pre-trained conversation model is used to perform question-answering reasoning based on the input conversation text to obtain an output conversation text, wherein the conversation model is trained using historical conversation texts.

[0007] Furthermore, the step of obtaining the input conversation text of the user during the interaction process includes: Obtain the user's input conversation voice during the interaction process; The input conversation speech is converted into text using a speech recognition model to obtain an input conversation text, wherein the speech recognition model is trained using historical speech data.

[0008] Further, the using of the speech retrieval model to perform retrieval processing on the output conversation text to determine the target speech in a predetermined speech library includes: Using a speech retrieval model, the output conversation text is compared with each standard speech in a predetermined speech library for similarity to obtain a similarity score; On the basis that the similarity score is greater than a preset threshold, the standard speech with the highest similarity score is determined as the target speech.

[0009] Furthermore, after the similarity comparison between the output conversation text and each standard speech in the predetermined speech library is performed by using the speech retrieval model to obtain a similarity score, the method further includes: On the basis that the similarity score is less than or equal to a preset threshold, using a speech classification model to perform quality assessment on the output conversation text; If the quality assessment score indicates that the output conversation text belongs to the first category, the standard speech with the highest similarity score is determined as the target speech.

[0010] Furthermore, after the quality assessment of the output conversation text is performed using the speech classification model, the method further includes: If the quality assessment score indicates that the output conversation text belongs to the second category, the audio translation model is called to perform speech conversion on the output conversation text to obtain the output conversation speech.

[0011] Furthermore, before using the speech classification model to perform quality assessment on the output conversation text, the method further includes: Get conversation texts with different reply types; The reply quality of the conversation text is used as a classification label to construct a reply sample, and the mapping relationship between the reply sample and the classification labels of different reply qualities during the training process is utilized to train the speech classification model.

[0012] According to a second aspect of the present application, a session data processing device based on a large model is provided, comprising: An acquisition unit, used to acquire the output conversation text of the user during the interaction process; A determination unit, configured to perform a search process on the output conversation text using a speech retrieval model to determine a target speech in a predetermined speech library, wherein the predetermined speech library records standard speech obtained by encoding and mapping speech streams in different fields, and the target speech is the standard speech in the predetermined speech library with the highest similarity to the output conversation text; A sending unit is used to send the voice stream corresponding to the target speech as a conversation interaction result to the user.

[0013] Furthermore, the acquisition unit includes: The first acquisition module is used to acquire the user's input conversation text during the interaction process; The reasoning module is used to perform question-answering reasoning based on the input conversation text using a pre-trained conversation model to obtain an output conversation text, wherein the conversation model is obtained by training using historical conversation texts.

[0014] Furthermore, the first acquisition module is specifically used to: Obtain the user's input conversation voice during the interaction process; The input conversation speech is converted into text using a speech recognition model to obtain an input conversation text, wherein the speech recognition model is trained using historical speech data.

[0015] Furthermore, the determining unit includes: A comparison module, for comparing the output conversation text with each standard speech in a predetermined speech library by using a speech retrieval model to obtain a similarity score; A determination module is used to determine the standard speech with the highest similarity score as the target speech based on the similarity score being greater than a preset threshold.

[0016] Furthermore, the determining unit further includes: An evaluation module is used for comparing the output conversation text with each standard speech in a predetermined speech library by using the speech retrieval model to obtain a similarity score, and then, on the basis that the similarity score is less than or equal to a preset threshold, using a speech classification model to perform a quality evaluation on the output conversation text; The determination module is further configured to determine the standard speech with the highest similarity score as the target speech if the quality assessment score indicates that the output conversation text belongs to the first category.

[0017] Furthermore, the determining unit further includes: The conversion module is used to call the audio translation model to perform speech conversion on the output conversation text to obtain the output conversation speech after the quality assessment is performed on the output conversation text using the speech classification model, if the quality assessment score indicates that the output conversation text belongs to the second category.

[0018] Furthermore, the determining unit further includes: A second acquisition module is used to acquire conversation texts of different answer types before using the speech classification model to perform quality assessment on the output conversation text; The training module is used to construct reply samples using the reply quality of the conversation text as classification labels, and to train the speech classification model by utilizing the mapping relationship between the reply samples and classification labels of different reply qualities during the training process.

[0019] According to a third aspect of the present application, a computer device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method described in the first aspect when executing the computer program.

[0020] According to a fourth aspect of the present application, a readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method described in the first aspect are implemented.

[0021] By means of the above technical scheme, the present application provides a method, device and equipment for processing conversation data based on a large model. Compared with the current prior art method of processing conversation data using a conversation model with low participation in the task stage, the present application obtains the output conversation text of the user during the interaction process, and uses the speech retrieval model to retrieve and process the output conversation text to determine the target speech in the predetermined speech library. The predetermined speech library records the standard speech obtained by encoding and mapping the speech streams in different fields. The target speech is the standard speech with the highest similarity between the text retrieval and the output conversation text, and the speech stream corresponding to the target speech is sent to the user as the result of the conversation interaction. The whole process uses the speech retrieval model to standardize and map the conversation text on the basis of the output conversation text. The standardized conversation text has the speech stream of the conversation output in the predetermined speech library. The speech stream can directly participate in the task stage of conversation data processing, avoiding the frequent call of speech translation during the conversation data processing process, and improving the response time of conversation data.

[0022] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings: Figure 1 It is a flowchart of a session data processing method based on a large model in one embodiment of the present application; Figure 2 yes Figure 2 A schematic flow chart of a specific implementation of step 101; Figure 3 yes Figure 2 A schematic flow chart of a specific implementation of step 102; Figure 4 yes Figure 2 Another specific implementation flow diagram of step 102; Figure 5 yes Figure 2 Another specific implementation flow diagram of step 102; Figure 6 It is a flowchart of session data processing based on a large model in one embodiment of the present application; Figure 7 It is a structural diagram of a session data processing device based on a large model in an embodiment of the present application; Figure 8 It is a schematic diagram of the device structure of a computer device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0024] Now the content of the present invention will be discussed with reference to several exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and thus implement the content of the present invention, rather than implying any limitation on the scope of the present invention.

[0025] As used herein, the term "including" and variations thereof are to be interpreted as open-ended terms meaning "including but not limited to." The term "based on" is to be interpreted as "based, at least in part, on." The terms "one embodiment" and "an embodiment" are to be interpreted as "at least one embodiment." The term "another embodiment" is to be interpreted as "at least one other embodiment."

[0026] In the scenario of processing conversation data, the conversation model can be applied to the conversation data processing as an intelligent question-and-answer customer service, and then the platform can assign clues to the intelligent question-and-answer customer service. The intelligent question-and-answer customer service will then call designated customers according to the clues, introduce product details, and transfer calls from customers with higher intentions to real agents. The agents will then provide services based on the task requirements and guide users to complete the service on the platform. However, considering that the conversation model is mainly involved in the initial access stage during the entire conversation data processing process, and its participation in the task stage is relatively low, it is difficult to achieve the expected task effect.

[0027] In order to solve this problem, this embodiment provides a session data processing method based on a large model, such as Figure 1 As shown, the method is applied to the platform side of the intelligent conversation, and includes the following steps: 101. Obtain the output conversation text of the user during the interaction process.

[0028] In this embodiment, the interaction process can be applied in a conversation scenario, which can be implemented through an artificial intelligence product of intelligent voice technology, which can help enterprises to perform outbound and inbound phone calls. In a conversation scenario, the robot customer service can realize information interaction through a conversation model.

[0029] It should be noted that the conversation model usually has strong multi-round conversation capabilities. The text interaction information of each round of conversation includes input conversation text and output conversation text. Here, the input conversation text is equivalent to the question text, and the output conversation text is equivalent to the answer text. Each set of question text and answer text forms a round of information interaction in the conversation scenario. Taking a round of information interaction as an example, the input text is "How much does it cost to go to Chengdu Airport" and the answer text is "40 yuan".

[0030] Specifically, in the process of historical information interaction, the conversation scenario will generate a large number of question-and-answer conversations formed by interactive tasks. Each interactive task is equivalent to an information interaction process. The conversation text generated by the interactive task includes text interaction information of multiple rounds of conversations. Here, the text interaction information contains the context information of the question and answer, which is expressed in the form of one question and one answer. The conversation scenario here can be applied to scenarios in various industries, including but not limited to second-hand transactions, house transactions, after-sales services, order customer service, appointments, and sharing scenarios.

[0031] The executor of this embodiment can be a conversation data processing device or equipment based on a large model, which can be configured on the platform end of the intelligent conversation. The conversation model is embedded through the platform end of the intelligent conversation. After receiving the input conversation text, the input conversation text is inferred through the conversation model to obtain the output conversation text of the user during the interaction process.

[0032] 102. Use a speech retrieval model to perform retrieval processing on the output conversation text to determine the target speech in a predetermined speech library.

[0033] Among them, the predetermined speech library records the standard speech obtained by encoding and mapping the voice streams in different fields. The standard speech here is equivalent to using the existing historical dialogue data to generate the best speech answer sentence for this round of conversation. The standard speech can be directly connected to the voice stream through encoding and mapping. The target speech is the standard speech in the predetermined speech library with the highest similarity to the output conversation text. Based on the high similarity, the voice stream corresponding to the target speech can be used directly as the output conversation voice.

[0034] Specifically, a text clustering method can be used to construct a predetermined speech library. The text clustering method requires specifying a speech category n, clustering the original speech, and obtaining speech clusters of n categories. The larger n is, the finer the clustering granularity is. Each speech cluster obtained by clustering can be a standard speech generated manually. The text clustering method here includes but is not limited to k-means.

[0035] In this embodiment, the speech retrieval model can perform text search in the predetermined speech according to the similarity matching rules, and perform similarity scoring for each standard text in the predetermined speech library to search for the standard speech with the highest similarity match with the output conversation text to determine as the target speech.

[0036] 103. Send the voice stream corresponding to the target speech as a conversation interaction result to the user.

[0037] It is understandable that the conversation model of the interaction process is unimodal, which means that the input and output of the conversation model are in text form. However, normal user conversation interactions are all voice streams, which requires converting the output conversation text into a voice stream. In related technologies, an audio translation model is usually used to convert the output conversation text into a voice stream. However, this frequent call to the audio translation model to achieve text-to-speech conversion process will bring great time costs.

[0038] In order to avoid this phenomenon, before converting the output conversation text into the output conversation speech, this embodiment uses the speech retrieval model to search in the predetermined speech library to obtain the target speech with the highest similarity match with the output conversation text. The predetermined speech library is used to store standard speech in different fields, and then the voice stream corresponding to the target speech with the highest phase speed match is used as the output conversation speech, without frequently calling the audio translation model, thereby improving the response speed of the conversation interaction.

[0039] The conversation data processing method based on a large model provided in the embodiment of the present application is compared with the method of processing conversation data using a conversation model with low participation in the task stage in the current prior art. The present application obtains the output conversation text of the user during the interaction process, and uses the speech retrieval model to retrieve and process the output conversation text to determine the target speech in the predetermined speech library. The predetermined speech library records the standard speech obtained by encoding and mapping the voice streams in different fields. The target speech is the standard speech with the highest similarity between the text retrieval and the output conversation text. The speech stream corresponding to the target speech is sent to the user as the result of the conversation interaction. The whole process uses the speech retrieval model to standardize and map the conversation text on the basis of the output conversation text. The standardized conversation text has the speech stream of the conversation output in the predetermined speech library. The speech stream can directly participate in the task stage of conversation data processing, avoiding frequent calls to speech translation during conversation data processing and improving the response time of conversation data.

[0040] In one possible implementation, the output conversation text of the user during the interaction process is usually a reply sentence obtained by question-answering reasoning based on the conversation model. Specifically, in the above embodiment, Figure 2 As shown, step 101 includes the following steps: 201. Obtain the user's input conversation text during the interaction process.

[0041] 202. Perform question-answering reasoning based on the input conversation text using a pre-trained conversation model to obtain an output conversation text.

[0042] In this embodiment, the conversation model is obtained by training using historical conversation texts, which are usually composed of input conversation samples and output conversation samples. The conversation model is trained by using the input conversation samples and the output conversation samples as the mapping relationship formed during the training process as question-answer labels, so that the trained conversation model can perform question-answer reasoning based on the input conversation text to obtain the output conversation text. Here, the conversation model can use an open source model.

[0043] In one possible implementation, the output conversation text of the user during the interaction process is usually a reply statement obtained by question-answering reasoning based on a conversation model. Specifically, in the process of obtaining the input conversation text of the user during the interaction process, the input conversation voice of the user during the interaction process can be first obtained, and then the input conversation voice can be converted into text using a speech recognition model to obtain the input conversation text. Here, the speech recognition model is trained using historical voice data, and the historical voice data consists of conversation text samples and corresponding conversation voice samples. The mapping relationship between the conversation voice samples and the conversation text samples formed during the training process is used to train the speech recognition model, so that the trained speech recognition model can perform text conversion based on the input conversation voice to obtain the input conversation text.

[0044] In the above embodiment, specifically, Figure 4 As shown, step 102 includes the following steps: 301. Use a speech retrieval model to compare the output conversation text with each standard speech in a predetermined speech library to obtain a similarity score.

[0045] 302. On the basis that the similarity score is greater than a preset threshold, the standard speech with the highest similarity score is determined as the target speech.

[0046] It is understandable that the output conversation text cannot be directly output to the user side, but needs to be converted into a voice stream before it can be received by the user. In order to avoid frequently calling the voice translation model, the embodiment of the present invention can retrieve each output conversation text through the speech retrieval model, and perform similarity matching between the output conversation text and each standard speech in the predetermined speech library. The standard speech here has a voice stream obtained through translation, so that the standard speech with the highest similarity is determined as the target speech, so that the voice stream corresponding to the target speech can be directly used as the output conversation voice in the future.

[0047] Furthermore, considering that the conversation model is directly applied to intelligent question-and-answer customer service, and the output conversation text obtained through the conversation model may have some abnormal problems, for example, the output conversation text generated is nonsense / factually incorrect / incomplete / logically disordered. This example can more conveniently obtain standard conversation words suitable for user replies by introducing a speech retrieval model to avoid the occurrence of the above abnormal problems.

[0048] It should be noted that the similarity score, as the similarity between the output conversation text and the standard speech, can reflect the degree of similarity in voice between the output conversation text and the standard speech to a certain extent. However, in the case of very low similarity, if the voice stream corresponding to the standard speech is still used as the output conversation voice, there may be certain errors. Here, a preset threshold is set for the similarity score as a prerequisite for using the voice stream corresponding to the standard speech. Only when the similarity score is greater than the preset threshold, the voice stream corresponding to the standard speech is used as the output conversation voice. Then correspondingly, on the basis that the similarity score is less than or equal to the preset threshold, further, in the above embodiment, if Figure 4 As shown, after step 201, the method further includes the following steps: 401. Based on the similarity score being less than or equal to a preset threshold, use a speech classification model to perform quality assessment on the output conversation text.

[0049] 402. If the quality assessment score indicates that the output conversation text belongs to the first category, the standard speech with the highest similarity score is determined as the target speech.

[0050] It is understandable that if the similarity score is less than or equal to the preset threshold, it means that there is no standard speech similar to the output conversation text in the predetermined speech library, and the speech classification model needs to be used to evaluate the quality of the output conversation text. For the output conversation text with poor quality, it means that even if the speech translation model is used, the output conversation text will not get the expected output conversation voice, and the speech stream corresponding to the standard speech with the highest similarity is still selected as the output conversation voice. The first category here mainly includes text types that cannot be converted by the speech transfer model, that is, text categories with poor quality, for example, text categories involving nonsense, inoperable text categories, and incomplete / grammatically disordered / unfriendly text categories. These text categories cannot allow the speech transfer model to output the expected output conversation voice.

[0051] Furthermore, if the quality assessment score indicates that the output conversation text belongs to the second category, the audio translation model is called to perform speech conversion on the output conversation text to obtain the output conversation speech. The second category here includes text types that can be converted into speech by the speech transfer model, that is, text types with higher quality. At this time, the audio translation model can be called to achieve speech conversion of the output conversation text to obtain the output conversation speech.

[0052] That is to say, when the similarity score is less than or equal to a preset threshold, the quality of the output conversation text can be evaluated by the speech classification model to determine whether to call the voice transfer model based on the quality evaluation result.

[0053] Furthermore, in the process of training the speech classification model, in the above embodiment, Figure 5 As shown, before step 401, the method further includes the following steps: 501. Get the conversation text of different reply types.

[0054] 502. Construct a reply sample using the reply quality of the conversation text as a classification label, and train a speech classification model by utilizing the mapping relationship between the reply sample and the classification labels of different reply qualities during the training process.

[0055] In this embodiment, the conversation text can be pre-set with the following four classification labels of different answer qualities: 0. High-quality text label, 1. Nonsense text label, 2. Inoperable text label, 3. Incomplete / grammatically incorrect / unfriendly text label. The classification label 0 here is equivalent to the output conversation text of the second category, and the classification labels 1-3 are equivalent to the output conversation text of the first category. During the training process, the above four classification labels are used to train the reply samples to obtain the mapping relationship between the reply samples and the classification labels of different reply qualities, and a speech classification model is constructed according to the mapping relationship.

[0056] Take an incomplete / grammatically incorrect text table as an example. The conversation text is "This insurance is worth a million yuan, with a deductible as low as 10,000 yuan and a wide range of coverage. I suggest you consider it." After being shuffled at random positions or segmented at any position, the following answer samples can be obtained: 1. This insurance is a million-dollar medical insurance with a minimum coverage range of up to 10,000 yuan. It provides great protection, so it is recommended that you consider it.

[0057] 2. This insurance provides medical compensation from one million to ten thousand, with no minimum insurance premium. The coverage is very wide, so it is recommended that you consider it.

[0058] 3. This insurance is a million-dollar medical insurance with a deductible as low as one.

[0059] 4. The medical insurance covers 40,000 yuan, with a deductible as low as 10,000 yuan and a wide range of coverage. It is recommended that you consider it.

[0060] When the output conversation text is evaluated by the speech classification model and the classification label is 0, it means that the output conversation text has high quality, and the audio translation model needs to be called to perform speech conversion on the output conversation text to obtain the output conversation speech. When the output conversation text is evaluated by the speech classification model and the classification label is 1-3, it means that the output conversation text has quality problems. For example, the output conversation text contains hallucinations, nonsense, unfriendly content, etc. At this time, the audio translation model will not be called to perform speech conversion on the output conversation text. Instead, the speech stream corresponding to the standard speech with the highest similarity is directly used as the output conversation speech to ensure the stability of the conversation system output.

[0061] This embodiment uses a speech classification model to evaluate the quality of the output conversation text, which greatly avoids the text quality problems that exist in question-answering reasoning by the conversation model. It also ensures that the output conversation text with higher quality in the task stage can directly call the speech transfer model for translation, thereby improving the intelligence of the conversation.

[0062] In actual application scenarios, the session data processing flow based on the large model is as follows: Figure 6 As shown, after receiving the user's conversation input voice, the user's conversation voice is first converted into conversation input text through the speech recognition model, and then input into the conversation model together with the historical conversation text for question-answering reasoning to obtain the conversation output text, and then use the speech retrieval model to search in the speech library to obtain the similarity score between the output conversation text and the standard speech. If the similarity score is greater than 0.9, the speech stream corresponding to the target speech with the highest similarity score in the predetermined speech library is selected as the output conversation voice. If the similarity is less than or equal to 0.9, the output conversation text is quality evaluated by the speech classification model. If the quality evaluation output conversation text belongs to category 1, 2, 3, that is, the output conversation text belongs to the first category with poor quality, then the speech stream corresponding to the target speech with the highest similarity score in the predetermined speech library is selected as the output conversation voice. If the quality evaluation output conversation text does not belong to category 1, 2, 3, that is, the output conversation text belongs to the second category with higher quality, the speech translation model is called to perform speech conversion on the output conversation text to obtain the output conversation voice.

[0063] Further, as Figure 1-5 The specific implementation of the method, the embodiment of the present application provides a session data processing device based on a large model, such as Figure 7 As shown, the device includes: an acquisition unit 61, a determination unit 62, and a sending unit 63.

[0064] An acquisition unit 61 is used to acquire the output conversation text of the user during the interaction process; A determination unit 62 is used to perform a search process on the output conversation text using a speech retrieval model to determine a target speech in a predetermined speech library, wherein the predetermined speech library records standard speech obtained by encoding and mapping speech streams in different fields, and the target speech is the standard speech in the predetermined speech library with the highest similarity to the output conversation text; The sending unit 63 is used to send the voice stream corresponding to the target speech as a conversation interaction result to the user.

[0065] The conversation data processing device based on a large model provided by the embodiment of the present invention is compared with the method of processing conversation data using a conversation model with low participation in the task stage in the current prior art. The present application obtains the output conversation text of the user during the interaction process, and uses the speech retrieval model to retrieve and process the output conversation text to determine the target speech in the predetermined speech library. The predetermined speech library records the standard speech obtained by encoding and mapping the voice streams in different fields. The target speech is the standard speech with the highest similarity between the text retrieval and the output conversation text. The voice stream corresponding to the target speech is sent to the user as the result of the conversation interaction. The whole process uses the speech retrieval model to standardize and map the conversation text on the basis of the output conversation text. The standardized conversation text has the voice stream of the conversation output in the predetermined speech library. The voice stream can directly participate in the task stage of conversation data processing, avoiding the frequent call of voice translation in the conversation data processing process, and improving the response time of conversation data.

[0066] In a specific application scenario, the acquisition unit includes: The first acquisition module is used to acquire the user's input conversation text during the interaction process; The reasoning module is used to perform question-answering reasoning based on the input conversation text using a pre-trained conversation model to obtain an output conversation text, wherein the conversation model is obtained by training using historical conversation texts.

[0067] In a specific application scenario, the first acquisition module is specifically used to: Obtain the user's input conversation voice during the interaction process; The input conversation speech is converted into text using a speech recognition model to obtain an input conversation text, wherein the speech recognition model is trained using historical speech data.

[0068] In a specific application scenario, the determining unit includes: A comparison module, for comparing the output conversation text with each standard speech in a predetermined speech library by using a speech retrieval model to obtain a similarity score; A determination module is used to determine the standard speech with the highest similarity score as the target speech based on the similarity score being greater than a preset threshold.

[0069] In a specific application scenario, the determining unit further includes: An evaluation module is used for comparing the output conversation text with each standard speech in a predetermined speech library by using the speech retrieval model to obtain a similarity score, and then, on the basis that the similarity score is less than or equal to a preset threshold, using a speech classification model to perform a quality evaluation on the output conversation text; The determination module is further configured to determine the standard speech with the highest similarity score as the target speech if the quality assessment score indicates that the output conversation text belongs to the first category.

[0070] In a specific application scenario, the determining unit further includes: The conversion module is used to call the audio translation model to perform speech conversion on the output conversation text to obtain the output conversation speech after the quality assessment is performed on the output conversation text using the speech classification model, if the quality assessment score indicates that the output conversation text belongs to the second category.

[0071] In a specific application scenario, the determining unit further includes: A second acquisition module is used to acquire conversation texts of different answer types before using the speech classification model to perform quality assessment on the output conversation text; The training module is used to construct reply samples using the reply quality of the conversation text as classification labels, and to train the speech classification model by utilizing the mapping relationship between the reply samples and classification labels of different reply qualities during the training process.

[0072] It should be noted that for other corresponding descriptions of the functional units involved in the session data processing device based on the large model provided in this embodiment, reference can be made to Figure 1-Figure 5 The corresponding description in will not be repeated here.

[0073] Based on the above Figure 1-Figure 5 The method shown in the embodiment of the present application accordingly provides a storage medium on which a computer program is stored, and when the program is executed by a processor, the above-mentioned Figure 1-Figure 5 The session data processing method based on the large model is shown.

[0074] Based on this understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.), and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each implementation scenario of the present application.

[0075] Based on the above Figure 1-Figure 5 The method shown, and Figure 7 In order to achieve the above-mentioned purpose, the embodiment of the present application also provides a physical device for processing session data based on a large model, which can be a computer, a smart phone, a tablet computer, a smart watch, a server, or a network device, etc. The physical device includes a storage medium and a processor; the storage medium is used to store a computer program; the processor is used to execute the computer program to achieve the above-mentioned Figure 1-Figure 5 The session data processing method based on the large model is shown.

[0076] Optionally, the physical device may also include a user interface, a network interface, a camera, a radio frequency (RF) circuit, a sensor, an audio circuit, a WI-FI module, etc. The user interface may include a display, an input unit such as a keyboard, etc., and the optional user interface may also include a USB interface, a card reader interface, etc. The network interface may optionally include a standard wired interface, a wireless interface (such as a WI-FI interface), etc.

[0077] In an exemplary embodiment, see Figure 8 The physical device includes a communication bus, a processor, a memory and a communication interface, and may also include an input / output interface and a display device, wherein each functional unit can communicate with each other through the bus. The memory stores a computer program, and the processor is used to execute the program stored in the memory and execute the session data processing method based on the large model in the above embodiment.

[0078] Those skilled in the art will appreciate that the entity device structure for session data processing based on a large model provided in this embodiment does not constitute a limitation on the entity device, and may include more or fewer components, or a combination of certain components, or different component arrangements.

[0079] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the entity device for the large model-based session data processing, and supports the operation of the information processing program and other software and / or programs. The network communication module is used to realize the communication between the components inside the storage medium, and the communication with other hardware and software in the information processing entity device.

[0080] Through the description of the above implementation methods, those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform, or by hardware. By applying the technical solution of the present application, compared with the current existing methods, the present application uses a speech retrieval model to standardize and map the conversation text on the basis of outputting the conversation text. The standardized conversation text has a voice stream of the conversation output in the predetermined speech library, and the voice stream can directly participate in the task stage of conversation data processing, avoiding frequent calls to speech translation during conversation data processing, and improving the response time of conversation data.

[0081] Those skilled in the art will appreciate that the accompanying drawings are only schematic diagrams of a preferred implementation scenario, and the modules or processes in the accompanying drawings are not necessarily necessary for implementing the present application. Those skilled in the art will appreciate that the modules in the devices in the implementation scenario can be distributed in the devices of the implementation scenario according to the description of the implementation scenario, or can be changed accordingly and located in one or more devices different from the present implementation scenario. The modules of the above-mentioned implementation scenario can be combined into one module, or can be further split into multiple submodules.

[0082] The above serial numbers of this application are only for description and do not represent the advantages and disadvantages of the implementation scenarios. The above disclosure is only a few specific implementation scenarios of this application, but this application is not limited to them, and any changes that can be thought of by technicians in this field should fall within the scope of protection of this application.

Claims

1. A session data processing method based on a large model, characterized in that: include: Get the user's output conversation text during the interaction process; The output conversation text is searched and processed using a speech retrieval model to determine a target speech in a predetermined speech library, wherein the predetermined speech library records standard speech obtained by encoding and mapping speech streams in different fields, and the target speech is the standard speech in the predetermined speech library with the highest similarity to the output conversation text; The voice stream corresponding to the target speech is sent to the user as a conversation interaction result.

2. The method according to claim 1, characterized in that The step of obtaining the output conversation text of the user during the interaction process includes: Get the user's input conversation text during the interaction process; A pre-trained conversation model is used to perform question-answering reasoning based on the input conversation text to obtain an output conversation text, wherein the conversation model is trained using historical conversation texts.

3. The method according to claim 2, characterized in that The step of obtaining the user's input conversation text during the interaction process includes: Obtain the user's input conversation voice during the interaction process; The input conversation speech is converted into text using a speech recognition model to obtain an input conversation text, wherein the speech recognition model is trained using historical speech data.

4. The method according to any one of claims 1 to 3, characterized in that The using of the speech retrieval model to perform retrieval processing on the output conversation text to determine the target speech in a predetermined speech library includes: Using a speech retrieval model, the output conversation text is compared with each standard speech in a predetermined speech library for similarity to obtain a similarity score; On the basis that the similarity score is greater than a preset threshold, the standard speech with the highest similarity score is determined as the target speech.

5. The method according to claim 4, characterized in that After using the speech retrieval model to compare the output conversation text with each standard speech in the predetermined speech library for similarity to obtain a similarity score, the method further includes: On the basis that the similarity score is less than or equal to a preset threshold, using a speech classification model to perform quality assessment on the output conversation text; If the quality assessment score indicates that the output conversation text belongs to the first category, the standard speech with the highest similarity score is determined as the target speech.

6. The method according to claim 5, characterized in that After the quality assessment of the output conversation text is performed using the speech classification model, the method further includes: If the quality assessment score indicates that the output conversation text belongs to the second category, the audio translation model is called to perform speech conversion on the output conversation text to obtain the output conversation speech.

7. The method according to claim 5, characterized in that Before using the speech classification model to perform quality assessment on the output conversation text, the method further includes: Get conversation texts with different reply types; The reply quality of the conversation text is used as a classification label to construct a reply sample, and the mapping relationship between the reply sample and the classification labels of different reply qualities during the training process is utilized to train the speech classification model.

8. A session data processing device based on a large model, characterized in that: include: An acquisition unit, used to acquire the output conversation text of the user during the interaction process; A determination unit, configured to perform a search process on the output conversation text using a speech retrieval model to determine a target speech in a predetermined speech library, wherein the predetermined speech library records standard speech obtained by encoding and mapping speech streams in different fields, and the target speech is the standard speech in the predetermined speech library with the highest similarity to the output conversation text; A sending unit is used to send the voice stream corresponding to the target speech as a conversation interaction result to the user.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the session data processing method based on the large model described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the session data processing method based on a large model as described in any one of claims 1 to 7 are implemented.

Citation Information

Patent Citations

  • Voice matching method of intelligent dialogue system, electronic device and computer device

    CN110534104A

  • Intelligent session generation method based on semantic matching and generation model

    CN116303930A

  • Business data processing method and device based on artificial intelligence, equipment and medium

    CN117131093A

  • Voice conversion method and device, computer equipment and storage medium

    CN117238295A

  • Text classification method and device, electronic equipment and storage medium

    CN117390181A