Abstract generation method and device, medium and computer program product

Abstract

The invention discloses an abstract generation method and device, a medium and a computer program product, and relates to the technical field of data processing, and the method comprises the steps: determining each piece of fine tuning data in a training data set, the fine tuning data comprising an ASR text and a first abstract, and the first abstract comprising a plurality of concerns in the ASR text; determining a first thinking chain and a first evidence reason corresponding to the fine tuning data according to a preset large model; according to the first thinking chain and the first evidence reason corresponding to each piece of fine tuning data, updating each piece of fine tuning data in the training data set to obtain an updated training data set; and optimizing a to-be-trained target large language model according to the updated training data set until a preset training ending condition is met, obtaining a trained target large language model, and performing abstract generation according to the trained target large language model. According to the method and the device, after the recording content is converted into the ASR text, the abstract corresponding to the ASR text is generated according to the target large language model.

Description

Technical Field

[0001] The present application relates to a specific field of data processing technology, and in particular to a summary generation method, device, medium and computer program product. Background Art

[0002] At present, outbound calls can be used to obtain current real-time information and promptly identify users who are sensitive to electronic collection. Generally speaking, it takes multiple collection recordings (i.e., telephone recordings of outbound calls) for a user to go from the first contact to the final conversion. During this process, each collection recording requires the collection staff to record important information for customer group analysis, thereby improving the collection success rate. However, this method has certain defects, that is, when the collection staff records important information in the collection recording, information recording errors or omissions may occur. Therefore, how to convert the important information in the recording content into a summary for recording has become an urgent problem that needs to be solved.

[0003] The above contents are only used to assist in understanding the technical solution of the present application and do not constitute an admission that the above contents are prior art. Summary of the invention

[0004] The main purpose of this application is to provide a summary generation method, device, medium and computer program product, aiming to solve the technical problem of how to generate a summary corresponding to the recorded content.

[0005] To achieve the above objectives, the present application proposes a method for generating a summary, the method comprising:

[0006] Determine each fine-tuning data in a preset training data set, wherein the fine-tuning data includes an ASR text and a first summary, and the first summary includes a plurality of focus points in the ASR text;

[0007] Determine a first chain of thought and a first evidence reason corresponding to the fine-tuning data according to a preset large model, wherein the first chain of thought is used to represent a logical link for generating the first summary according to the ASR text, and the first evidence reason is used to support a focus point in the first summary;

[0008] According to the first thought chain and the first evidence reason corresponding to each fine-tuning data, each fine-tuning data in the training data set is updated to obtain an updated training data set, wherein any updated fine-tuning data in the updated training data set includes the ASR text, the first summary, the first thought chain and the first evidence reason;

[0009] The target large language model to be trained is optimized according to the updated training data set until a preset training end condition is met, thereby obtaining a trained target large language model, wherein a summary is generated according to the trained target large language model.

[0010] In one embodiment, the step of determining the first chain of thought and the first evidence reason corresponding to the fine-tuning data according to the preset macro model includes:

[0011] When the fine-tuning data is public text data, determining an external device having a preset macro model;

[0012] Sending the fine-tuning data to the external device, wherein the external device inputs the fine-tuning data into a preset large model to obtain a first chain of thought and a first evidence reason;

[0013] receiving a first chain of thought and a first reason of evidence fed back by the external device according to the fine-tuning data; or,

[0014] When the fine-tuning data is private text data, determining a preset large model deployed locally;

[0015] The fine-tuning data is input into the preset large model to obtain the first chain of thinking and the first evidence reason.

[0016] In one embodiment, the step of optimizing the target large language model to be trained according to the updated training data set includes:

[0017] For any updated fine-tuning data in the updated training data set, the ASR text and the first thought chain are input into the target large language model to be trained to obtain a prediction result, wherein the prediction result includes a second summary and a second evidence reason;

[0018] Calculate the knowledge distillation loss from the large model to the target large language model according to the prediction result, the first summary and the first evidence reason;

[0019] The target large language model is optimized based on the knowledge distillation loss.

[0020] In one embodiment, after the step of inputting the ASR text and the first thought chain into the target large language model to be trained and obtaining the prediction result, the following steps are included:

[0021] Inputting the prediction result, and the first summary and the first evidence reason corresponding to the prediction result into the loss function corresponding to the target large language model, and calculating the prediction loss;

[0022] The step of optimizing the target large language model based on the knowledge distillation loss includes:

[0023] The target large language model is optimized based on the prediction loss and the knowledge distillation loss.

[0024] In one embodiment, the target large language model includes a model ontology and an adapter.

[0025] The step of optimizing the target large language model based on the prediction loss and the knowledge distillation loss includes:

[0026] A model parameter of the adapter is updated based on the knowledge distillation loss and the prediction loss to optimize the target large language model.

[0027] In one embodiment, after the step of obtaining the trained target large language model, the following steps are included:

[0028] Selecting at least one updated fine-tuning data in the updated training data set as the first evaluation data, inputting the ASR text and the first thought chain in the first evaluation data into the trained target large language model, and outputting a third summary and a third evidence reason;

[0029] If feedback information input by the user based on the third summary and the third evidence reason is received, the first summary and the first evidence reason in the first evaluation data are adjusted according to the feedback information, and / or the first chain of thought in the first evaluation data is adjusted according to the feedback information to obtain second evaluation data;

[0030] The trained target large language model is optimized again based on the second evaluation data until a preset training end condition is met, thereby obtaining a new target large language model.

[0031] In one embodiment, the step of updating each fine-tuning data in the training data set according to the first chain of thought and the first evidence reason corresponding to each fine-tuning data includes:

[0032] For any fine-tuning data in the training data set, the first chain of thought and the first evidence reason corresponding to the fine-tuning data are aligned with the fine-tuning data in text component units to obtain updated fine-tuning data, wherein the text component unit alignment includes attributing the fine-tuning data, the first chain of thought and the first evidence reason to the text component unit belonging to the same fine-tuning data.

[0033] In addition, to achieve the above-mentioned purpose, the present application also proposes a summary generation device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the summary generation method described above.

[0034] In addition, to achieve the above-mentioned purpose, the present application also proposes a medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the summary generation method described above are implemented.

[0035] In addition, to achieve the above-mentioned purpose, the present application also provides a computer program product, wherein the computer program product includes a computer program, and when the computer program is executed by a processor, the steps of the summary generation method described above are implemented.

[0036] In an embodiment of the present application, each fine-tuning data in a preset training data set is determined, and the fine-tuning data includes an ASR text and a first summary, and the first summary includes multiple focus points in the ASR text. That is, in the training data set, each fine-tuning data may include an ASR text converted from audio data, and a first summary constructed from the focus points in the ASR text. The collection recording is essentially also a kind of audio data, and the ASR text is a kind of text data converted from audio data. Therefore, during the model training stage, the collection recording can be converted into an ASR text, and the focus points in the ASR text can be used as important information that needs to be recorded. Therefore, for the important information record of a collection recording, it can be embodied in the form of a first summary by converting the collection recording into an ASR text, and then extracting multiple focus points in the ASR text.

[0037] The first chain of thought and the first evidence reason corresponding to the fine-tuning data are determined through the preset large model, and the first chain of thought is used to characterize the logical link for generating the first summary based on the ASR text, and the first evidence reason is used to verify the focus points in the first summary. That is, in order to reflect that the focus points in the ASR text are important information that the user wants to record, the large model can be used to determine the logical link for generating the first summary based on multiple focus points in the ASR text through the first chain of thought generated by the large model, and the correctness of each focus point can be verified based on the first evidence reason.

[0038] By updating each fine-tuning data in the training data set according to the first chain of thought and the first reason of evidence corresponding to each fine-tuning data, the updated fine-tuning data in the obtained updated training data set includes ASR text, first summary, first chain of thought and first reason of evidence. Then, the training data set can be optimized and updated with the relatively accurate first chain of thought and first reason of evidence generated by the large model to improve the quality of the training data set. And the target large language model to be trained will be optimized according to the updated training data set until the preset training end condition is met, and the trained target large language model is obtained, and then the summary is generated according to the trained target large language model. Then, the target large language model can be optimized with the help of the model performance of the large model, so that the trained target large language model has the ability to generate summaries and reasons of evidence, that is, the trained target large language model can directly extract important information from the ASR text converted from the collection recording, such as the determination of the focus, and then generate a summary corresponding to the collection recording. Compared with the method of manually recording important information by the collection personnel, the intelligence is significantly improved, and the summary generated by the target large language model is more accurate, and there will be no information recording errors or missing phenomena. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0040] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0041] Figure 1 A flowchart diagram of the first embodiment of the method for generating an abstract of the present application is provided;

[0042] Figure 2 A schematic diagram of a scenario in which a large model in the summary generation method of this application outputs a summary;

[0043] Figure 3 A schematic diagram of a scenario in which a target large language model performs summary output in the summary generation method of this application;

[0044] Figure 4 This is a schematic diagram of the overall scenario framework flow in the summary generation method of the embodiment of the present application;

[0045] Figure 5 Schematic diagram of the device structure of the hardware operating environment involved in the abstract generation method in the embodiment of the present application.

[0046] The purpose, features and advantages of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0047] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and are not used to limit the present application.

[0048] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0049] It should be noted that the execution subject of this embodiment can be a computing service device with data processing, network communication and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc., or an electronic device, a summary generation device, etc. that can realize the above functions. The following takes the summary generation device as an example to illustrate this embodiment and the following embodiments. A target large language model to be trained is deployed in the summary generation device. The target large language model can be a language model. The language model is a model that can be used to perform natural language processing tasks, such as LLaMa2 13B, LLaMa27B, OPT 7B, etc. Optionally, the summary generation device can be a device whose computing power can support the training and operation of the target large language model, which can be deployed in banks, insurance companies and other enterprises.

[0050] Based on this, the present application embodiment provides a summary generation method, referring to Figure 1 , Figure 1 This is a flowchart of the first embodiment of the abstract generation method of this application.

[0051] In this embodiment, the summary generation method includes steps S10 to S40.

[0052] Step S10, determining each fine-tuning data in a preset training data set;

[0053] It should be noted that the fine-tuning data includes the ASR text and the first summary, and the first summary includes multiple focus points in the ASR text.

[0054] Optionally, the fine-tuning data can be a type of text data, which can be private text data or public text data. Optionally, ASR (Automatic Speech Recognition) text can convert human speech into text data. For example, converting a collection recording into text data. The ASR text includes all the content in the collection recording. Optionally, the collection recording can be input into the ASR model to obtain the corresponding ASR text. Optionally, role distinction is required in the ASR text. For example, determine the voice data spoken by the customer service, determine the voice data spoken by the input customer, and reflect it in the ASR text in the form of text.

[0055] Optionally, the focus point may be manually defined, and different focus points may be defined according to different scene requirements. Optionally, the focus point may be a text feature in the ASR text, including important information in the ASR text that the user wants to focus on.

[0056] Optionally, the focus points corresponding to each ASR text may be different or the same. For example, the focus points may include but are not limited to: whether the person answers the call, the promised repayment time, the user's willingness to repay, and the risk of complaints.

[0057] Optionally, a plurality of voice data stored in advance (such as voice data corresponding to the collection recording) can be obtained in the database, and each voice data is input into the ASR model to obtain the ASR text. The focus points in the ASR text are marked, and the first summary is generated according to the marked focus points.

[0058] Optionally, the ASR text can be input into the large model with the aid of a large model to obtain a first summary corresponding to the ASR text. Optionally, the first prompt and the ASR text can be input into the large model, and the large model outputs the corresponding first summary. The first prompt is an instruction that tells the large model that the summary extraction of the focus point needs to be completed based on the ASR text.

[0059] For example, Figure 2 As shown, the input of the large model can be determined based on the ASR text and input into the large model. The large model performs summary output to obtain a first summary. Optionally, the input can be "You are a professional debt collector. Based on the conversation between the agent and the customer, a summary is generated, where the summary content includes the following focus points: [whether the person answers the call], [willingness to repay], [promised repayment time]... The following is the conversation between the agent and the user: [agent]... [customer]... According to the conversation between the agent and the customer, the specific summary is as follows:". The summary output of the large model can be "customer [whether the person answers the call] is; [willingness to repay] high; [promised repayment time] tomorrow afternoon at 5 o'clock;..."

[0060] Step S20, determining the first chain of thought and the first evidence reason corresponding to the fine-tuning data according to the preset macro model;

[0061] It should be noted that the first chain of thought is used to represent the logical link for generating the first summary based on the ASR text, and the first evidence reason is used to support the focus of the first summary.

[0062] Optionally, the first chain of thought may be embodied in the form of text data. The first evidence reason may be embodied in the form of text data.

[0063] Optionally, a thought chain generation template can be constructed in advance. Optionally, the thought chain generation template can include specific text information of the ASR text, and require the large model to determine the first summary corresponding to the ASR text, and the reasoning process (i.e., logical link) of why it is this first summary, so as to generate the corresponding first thought chain. Optionally, after receiving the ASR text, the large model can generate a thought chain corresponding to the ASR text according to the thought chain generation template, and use it as the first thought chain.

[0064] Optionally, the large model can find specific evidence reasons in the ASR text to support the focus points to determine the validity of the first summary output by the large model.

[0065] For example, let's take the repayment ability as an example. The big model can use the manual annotation results (ASR text, low repayment ability) as input, generate a text segment in the ASR text that can reflect the low repayment ability, and determine the first evidence reason corresponding to the focus point based on the text segment. Optionally, if the text segment is "I really have no money now, I am not like this, but now I am unemployed and have no money to eat", then the big model further summarizes the first evidence reason as "the customer mentioned that he had no money to eat because he was unemployed, so the corresponding output of the repayment ability of this focus point is low."

[0066] Step S30, updating each fine-tuning data in the training data set according to the first thought chain and the first evidence reason corresponding to each fine-tuning data, to obtain an updated training data set;

[0067] It should be noted that any piece of fine-tuning data updated in the updated training dataset includes ASR text, first summary, first chain of thought and first reason of evidence;

[0068] Optionally, after determining the first chain of thought and the first evidence reason corresponding to each fine-tuning data in the training data set based on the large model, the training data set can be updated based on the first chain of thought and the first evidence reason, so that the updated training data set includes the ASR text, the first summary, the first chain of thought and the first evidence reason.

[0069] Optionally, when updating the training data set, each fine-tuning data in the training data set can be updated, and the corresponding first chain of thought and first evidence reason can be updated to the fine-tuning data. And when updating, the processing can be performed in parallel, which is not limited here. Optionally, for each fine-tuning data in the training data set, first determine the first chain of thought and first evidence reason corresponding to the fine-tuning data. Then input the first chain of thought and the first evidence reason into the training data set, and in the training data set, align the first chain of thought and the first evidence reason with the ASR text and the first summary in the fine-tuning data to update the first chain of thought and the first evidence reason to the fine-tuning data, and obtain updated fine-tuning data.

[0070] Further, in a feasible embodiment, in step S30, the step of updating each fine-tuning data in the training data set according to the first chain of thought and the first evidence reason corresponding to each fine-tuning data includes step S31.

[0071] Step S31, for any fine-tuning data in the training data set, align the first thought chain and the first evidence reason corresponding to the fine-tuning data with the text component units of the fine-tuning data to obtain updated fine-tuning data.

[0072] It should be noted that the alignment of text component units includes attributing the fine-tuning data, the first chain of thought and the first evidence reason to the text component unit belonging to the same fine-tuning data.

[0073] Optionally, the text component unit is a basic unit constituting the text, which can be obtained by artificial division, and each text component unit divided based on a set of division rules constitutes a vocabulary. For example, the text component unit is divided by each word. Optionally, the specific implementation method of text component unit alignment is not limited.

[0074] Optionally, if each fine-tuning data in the training data set is stored in the form of a table, the table storing each fine-tuning data may be used as the first table, and the first table contains the corresponding relationship between the ASR text and the first summary.

[0075] Optionally, after the first chain of thought and the first evidential reason corresponding to each fine-tuning data are determined according to the large model, they can be stored in the form of a table, and the table can be used as the second table.

[0076] Optionally, the text component units of the first table containing each fine-tuning data in the training data set and the second table containing the first chain of thought and the first evidence reason can be aligned. For example, if the text component unit included in the first table is a first text component unit containing a sequence number corresponding to the fine-tuning data, a second text component unit containing an ASR text, and a third text component unit containing a first summary. If the text component unit included in the second table is a fourth text component unit containing a sequence number corresponding to the fine-tuning data, a fifth text component unit containing the first chain of thought, and a sixth text component unit containing the first evidence reason. When aligning the text component units, the first text component unit in the first table that has not been aligned with the text component unit can be traversed, and according to the sequence number in the traversed first text component unit, the corresponding sequence number in the second table and the fourth text component unit where the sequence number is located are determined, and it is used as the fourth text component unit matching the first text component unit. Determine the fifth and sixth text components in the second table that are in the same row as the matched fourth text component, and edit them into the first table, and into a blank table in the same row as the traversed first text component, so that the traversed first text component, and the corresponding second text component, third text component, fifth text component and sixth text component are in the same row, thus completing the alignment of the text components. At this point, it can be considered that the first evidence reason and the first chain of thought are updated to the training data set, and the updated fine-tuning data in the updated training data set includes the ASR text, the first summary, the first chain of thought and the first evidence reason in the same row.

[0077] In this embodiment, by aligning the text component units of the first thought chain and the first evidence reason corresponding to each fine-tuning data with the fine-tuning data, updated fine-tuning data is obtained, thereby ensuring that the updated fine-tuning data includes ASR text, the first summary, the first thought chain and the first evidence, thereby facilitating the effective execution of subsequent training of the target large language model.

[0078] Step S40, optimizing the target large language model to be trained according to the updated training data set until a preset training end condition is met, thereby obtaining a trained target large language model.

[0079] It should be noted that the ASR text converted from the recorded data is summarized based on the trained target large language model.

[0080] Optionally, the preset training end condition can be a pre-set condition that needs to be met to end the training, which can be set as needed, and is not limited in this embodiment. For example, in a specific embodiment, it can be set to the loss function of the target large language model converges or the number of training rounds reaches a set number of rounds, or the training duration reaches a set duration, etc.

[0081] Optionally, after determining the updated training data set, the updated fine-tuning data in the updated training data set can be used to perform model training on the target large language model to be trained, and the model parameters of the target large language model can be optimized according to the prediction results of the model training and the updated fine-tuning data until the preset training end conditions are met. The iterative training process can then be terminated, and the final updated target large language model can be used as the trained target large language model.

[0082] Optionally, in the model application stage of the target large language model, the ASR text corresponding to the actual collection recording can be first obtained, and the ASR text corresponding to the actual collection recording can be input into the target large language model, and the target large language model outputs the corresponding summary and evidence reasons.

[0083] In this embodiment, each fine-tuning data in a preset training data set is determined, and the fine-tuning data includes an ASR text and a first summary, and the first summary includes multiple focus points in the ASR text. That is, in the training data set, each fine-tuning data may include an ASR text converted from audio data, and a first summary constructed from the focus points in the ASR text. The collection recording is essentially also a kind of audio data, and the ASR text is a kind of text data converted from audio data. Therefore, during the model training stage, the collection recording can be converted into an ASR text, and the focus points in the ASR text can be used as important information that needs to be recorded. Therefore, for the important information record of a collection recording, it can be embodied in the form of a first summary by converting the collection recording into an ASR text, and then extracting multiple focus points in the ASR text.

[0084] The first chain of thought and the first evidence reason corresponding to the fine-tuning data are determined through the preset large model, and the first chain of thought is used to characterize the logical link for generating the first summary based on the ASR text, and the first evidence reason is used to verify the focus points in the first summary. That is, in order to reflect that the focus points in the ASR text are important information that the user wants to record, the large model can be used to determine the logical link for generating the first summary based on multiple focus points in the ASR text through the first chain of thought generated by the large model, and the correctness of each focus point can be verified based on the first evidence reason.

[0085] By updating each fine-tuning data in the training data set according to the first chain of thought and the first reason of evidence corresponding to each fine-tuning data, the updated fine-tuning data in the obtained updated training data set includes ASR text, first summary, first chain of thought and first reason of evidence. Then, the training data set can be optimized and updated with the relatively accurate first chain of thought and first reason of evidence generated by the large model to improve the quality of the training data set. And the target large language model to be trained will be optimized according to the updated training data set until the preset training end condition is met, and the trained target large language model is obtained, and then the summary is generated according to the trained target large language model. Then, the target large language model can be optimized with the help of the model performance of the large model, so that the trained target large language model has the ability to generate summaries and reasons of evidence, that is, the trained target large language model can directly extract important information from the ASR text converted from the collection recording, such as the determination of the focus, and then generate a summary corresponding to the collection recording. Compared with the method of manually recording important information by the collection personnel, the intelligence is significantly improved, and the summary generated by the target large language model is more accurate, and there will be no information recording errors or missing phenomena.

[0086] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as those in the first embodiment can be referred to the above description, and will not be described in detail later. On this basis, in step S20, the steps of determining the first chain of thought and the first evidence reason corresponding to the fine-tuning data according to the preset large model include steps a10-a30.

[0087] Step a10, when the fine-tuning data is public text data, determining an external device having a preset large model;

[0088] Optionally, the public text data may be text data disclosed in the industry for training the target large language model, and the public text data is converted from public audio data, such as public text data converted from public debt collection recordings.

[0089] Optionally, type detection can be performed on the fine-tuning data in the training data set. If the fine-tuning data is public text data, the model performance of the local target large language model can be optimized with the help of the large model deployed in the external device, effectively reducing the pressure on local computing resources and improving the efficiency of model training.

[0090] Optionally, an external device with a preset large model may be determined first. The external device may be a computing service device with data processing, network communication and program running functions, such as a server, a tablet computer, a personal computer, a mobile phone, etc.

[0091] Optionally, the device executing this embodiment (i.e., the summary generation device) and the external device may be devices deployed in different enterprises. A preset large model is deployed in the external device, and a target large language model to be trained is deployed in the summary generation device. Optionally, the model parameter complexity of the target large language model is much smaller than the model parameter complexity of the large model. For example, the large model in the external device is a model of hundreds of billions, and the target large language model in the summary generation device is a model of tens of billions or billions.

[0092] Optionally, the large model may be a language model, and the target large language model may be a language model. A language model is a model that can be used to perform natural language processing tasks, such as LLaMa213B, LLaMa2 7B, OPT 7B, etc. The structure of the large model is more complex, and the number of model parameters is more, while the structure of the target large language model is relatively simple, and the number of parameters is less.

[0093] Step a20, sending the fine-tuning data to an external device, wherein the external device inputs the fine-tuning data into a preset large model to obtain a first chain of thought and a first evidence reason;

[0094] Optionally, the fine-tuning data belonging to the public text data can be sent to an external device. The external device inputs the fine-tuning data into the preset large model, that is, inputs the ASR text and the first summary into the large model, and outputs the corresponding first chain of thought and first evidence reason. Optionally, after receiving the fine-tuning data, the large model can generate two task links, namely a first task link for generating a chain of thought, and a second task link for generating the first evidence reason. The large model executes the first task link according to the fine-tuning data, and outputs the first chain of thought. The large model executes the second task link according to the fine-tuning data, and outputs the first evidence reason. The first task link and the second task link can be executed simultaneously. Optionally, the external device sends the first chain of thought and the first evidence reason output by the large model to the summary generation device.

[0095] Step a30, receiving the first chain of thought and the first evidence reason fed back by the external device based on the fine-tuning data.

[0096] Optionally, after receiving the first chain of thought and the first evidence reason fed back by the external device, the first chain of thought and the first evidence reason may be updated to the training data set, so as to optimize the target large language model according to the updated training data set.

[0097] In this embodiment, when the fine-tuning data is public text data, the fine-tuning data is sent to an external device, and the first chain of thought and the first evidence reason are received by the external device after processing the fine-tuning data according to the large model and feeding back, thereby reducing the pressure on local computing resources, and obtaining the first chain of thought and the first evidence reason with the help of the large model of the external device, thereby realizing the knowledge distillation from the large model to the small model when training the target large language model according to the first chain of thought and the first evidence reason, and utilizing the model performance of the large model to improve the model performance of the target large language model.

[0098] Further, in a feasible embodiment, in step S20, the step of determining the first chain of thought and the first evidential reason corresponding to the fine-tuning data according to the preset large model also includes steps a40-a50.

[0099] Step a40, when the fine-tuning data is private text data, determining a preset large model deployed locally;

[0100] Optionally, the private text data may be business data from various enterprises, such as ASR text data generated after conversion of conversations between customer service personnel of banks and insurance companies and customers when serving customers, and the first summary corresponding to the ASR text data.

[0101] Optionally, the fine-tuning data in the training data set can be type-checked, and if the fine-tuning data is private text data, the corresponding large model can be deployed locally, for example, using the open source Qwen72B to deploy the large model.

[0102] Step a50, input the fine-tuning data into the preset large model to obtain the first chain of thought and the first evidence reason.

[0103] Optionally, the fine-tuning data can be input into the large model, that is, the ASR text and the first summary are input into the large model, and the corresponding first thought chain and first evidence reason are output. Optionally, after receiving the fine-tuning data, the large model can generate two task links, namely a first task link for generating a thought chain, and a second task link for generating the first evidence reason. The large model executes the first task link according to the fine-tuning data, and outputs the first thought chain. The large model executes the second task link according to the fine-tuning data, and outputs the first evidence reason. The first task link and the second task link can be executed simultaneously.

[0104] In this embodiment, when the fine-tuning data is private text data, the fine-tuning data is directly input into the local big model to obtain the first chain of thought and the first evidence reason, thereby avoiding the leakage of the fine-tuning data and improving the data security of the fine-tuning data.

[0105] Based on the first embodiment or the second embodiment of the present application, in the third embodiment of the present application, the same or similar contents as those in the first embodiment or the second embodiment can be referred to the above introduction, and will not be described in detail later. On this basis, in step S40, the step of optimizing the target large language model to be trained according to the updated training data set includes steps b10-b30.

[0106] Step b10, for any updated fine-tuning data in the updated training data set, input the ASR text and the first thought chain into the target large language model to be trained to obtain a prediction result, wherein the prediction result includes a second summary and a second evidence reason;

[0107] Optionally, when performing iterative training of the target large language model, the ASR text and the first chain of thought in the updated fine-tuning data can be input into the target large language model to be trained. The target large language model preprocesses the input ASR text and the first chain of thought, and generates two tasks, namely, a summary task for generating a summary and an evidence reason task for generating evidence reasons. The target large language model executes the two tasks in parallel and outputs corresponding prediction results, namely, a second summary and a second evidence reason. Optionally, the ASR text and the first chain of thought input to the target large language model belong to the same ASR text and the first chain of thought in the updated fine-tuning data.

[0108] For example, the input of the target large language model can be determined based on the updated fine-tuning data and input into the target large language model. The target large language model will output a corresponding summary and also output a corresponding evidence reason. Figure 3 , that is, the input part of the target large language model can be: "You are a professional debt collector. Generate a summary based on the conversation between the agent and the customer. The summary content includes the following focus points: [whether the person answers the call], [willingness to repay], [promised repayment time]...

[0109] Thinking chain: [Did I answer the call]: When the recording clearly expresses that I don’t know the person, dialed the wrong number, or denied the person, then the focus point output is no,...

[0110] The following is the conversation between the agent and the user:

[0111] [Agent]...[Customer]...

[0112] According to the conversation between the agent and the customer, the specific summary is as follows:".

[0113] The summary output of the target large language model can be "[Did the customer answer the call in person] Yes, because he did not explicitly express that he did not know the customer during the conversation, and answered the agent's questions normally; [Willingness to repay] High; [Promised repayment time] Tomorrow afternoon at 5 o'clock; ..."

[0114] Step b20, calculating the knowledge distillation loss from the large model to the target large language model based on the prediction result, the first summary and the first evidence reason;

[0115] Step b30, optimize the target large language model based on knowledge distillation loss.

[0116] Alternatively, knowledge distillation refers to a model training method that transfers the knowledge of a teacher model to a student model.

[0117] Optionally, the large model can be used as a teacher model, and the target large language model can be used as a student model to achieve knowledge distillation from the large model to the target large language model. When performing knowledge distillation, if the first summary is also predicted based on the large model. The first summary and the first evidence reason can be used as the large model prediction results of the large model. And the knowledge distillation loss from the large model to the target large language model is calculated based on the prediction results and the large model prediction results. The specific calculation method of the knowledge distillation loss is not limited in this embodiment. For example, the knowledge distillation loss can be obtained by calculating the cross entropy between the prediction results and the large model prediction results.

[0118] Optionally, in the knowledge distillation stage, the first chain of thought and the first evidence reason generated by the large model can be used as supervisory signals to guide the target large language model to learn, and the target large language model needs to minimize the difference between the second summary and the first summary output, and also needs to generate a second evidence reason as close to the first evidence reason output by the large model as possible. That is, the knowledge distillation loss from the large model to the target large language model can be calculated based on the difference from the second summary to the first summary, and the difference from the second evidence reason to the first evidence reason. The specific calculation method can refer to the above.

[0119] Optionally, after determining the knowledge distillation loss, the summary generation device can optimize the target large language model according to the knowledge distillation loss based on the computing resources it has. Specifically, the model parameters in the target large language model can be updated with the goal of reducing the knowledge distillation loss. For example, the gradient descent algorithm can be used to calculate the gradient values ​​corresponding to the model parameters in the target large language model according to the knowledge distillation loss, and the model parameters of the target large language model can be optimized according to the gradient values, thereby achieving the purpose of optimizing the target large language model.

[0120] Furthermore, in a feasible embodiment, after the step of inputting the ASR text and the first thought chain into the target large language model to be trained and obtaining the prediction result in step b10, step b40 is included.

[0121] Step b40: input the prediction result, the first summary and the first evidence reason corresponding to the prediction result into the loss function corresponding to the target large language model to calculate the prediction loss.

[0122] Optionally, after obtaining the prediction result output by the target large language model, the loss function corresponding to the target large language model can also be determined. The first summary and the first evidence reason corresponding to the prediction result in the updated training data set are input into the loss function for calculation to obtain the prediction loss. The model parameters of the target large language model can be updated in reverse according to the prediction loss until the preset training end condition is met to obtain the trained target large language model.

[0123] In a feasible embodiment, in step b30, the step of optimizing the target large language model based on the knowledge distillation loss includes step b50.

[0124] Step b50, optimize the target large language model based on the prediction loss and the knowledge distillation loss.

[0125] Optionally, the prediction loss and the knowledge distillation loss can be added or weighted to obtain the total loss. The model parameters in the target large language model are updated with the goal of reducing the total loss. For example, the gradient descent algorithm can be used to calculate the gradient values ​​corresponding to the model parameters in the target large language model based on the total loss, and the model parameters in the target large language model are updated according to the gradient values, thereby achieving the purpose of optimizing the target large language model.

[0126] In this embodiment, by optimizing the target large language model according to the prediction loss and knowledge distillation algorithm, the positive "knowledge" in the large model can be transferred to the target large language model, so that the target large language model can generate a more accurate summary when performing summary generation tasks, thereby improving the model performance of the target large language model.

[0127] Further, in a feasible embodiment, the target large language model includes a model body and an adapter.

[0128] Optionally, in step b50, the step of optimizing the target large language model based on the prediction loss and the knowledge distillation loss includes step b60.

[0129] Step b60, updates the model parameters of the adapter based on the knowledge distillation loss and the prediction loss to optimize the target large language model.

[0130] Optionally, the target large language model may include a model body and an adapter. The adapter may include an adapter located at the head of the model body and an adapter located at the tail of the model body.

[0131] Optionally, when the target large language model is updated using knowledge distillation loss and prediction loss, the model parameters in the model body are kept unchanged, and the model parameters in the adapter are updated, such as updating the model parameters in the head adapter and / or the tail adapter.

[0132] In this embodiment, the model parameters of the adapter are updated according to the knowledge distillation loss and the prediction loss to optimize the target large language model, thereby reducing the computational complexity of the device when optimizing the target large language model, thereby reducing the computing power requirements of the summary generation device.

[0133] In addition, to assist in understanding the complete training process of the target large language model in this embodiment, so as to generate a summary based on the trained target large language model, an example is given below.

[0134] For example, Figure 4 As shown, the original fine-tuning data is obtained, including the ASR text and the first summary, and the first summary includes multiple focus points in the ASR text. And the large model (72B) can be used to generate the thought chain + evidence. That is, for each original fine-tuning data, it can be input into the large model, and the first thought chain and the first evidence reason are output. And it is updated to the corresponding original fine-tuning data to obtain the fine-tuning data with the thought chain + evidence added, that is, the new fine-tuning data, and the target large language model is fine-tuned according to the new fine-tuning data. It is also possible to obtain user feedback information based on the result generation of the target large language model, and reversely optimize the target large language model based on the feedback information, so as to achieve the purpose of optimizing the target large language model through reinforcement learning based on human feedback.

[0135] Based on the first embodiment, the second embodiment, or the third embodiment of the present application, in the fourth embodiment of the present application, the same or similar content as the first embodiment, the second embodiment, or the third embodiment can refer to the above introduction, and will not be repeated later. On this basis, in step S40, after the step of obtaining the trained target large language model, steps S50-S70 are also included.

[0136] Step S50, selecting at least one updated fine-tuning data in the updated training data set as the first evaluation data, inputting the ASR text and the first thought chain in the first evaluation data into the trained target large language model, and outputting a third summary and a third evidence reason;

[0137] Step S60: if feedback information inputted by the user based on the third summary and the third evidence reason is received, the first summary and the first evidence reason in the first evaluation data are adjusted according to the feedback information, and / or the first chain of thought in the first evaluation data is adjusted according to the feedback information, so as to obtain second evaluation data;

[0138] Step S70, optimizing the trained target large language model again based on the second evaluation data until a preset training end condition is met, thereby obtaining a new target large language model.

[0139] In this embodiment, after training is performed based on the updated training data set to obtain the trained target large language model, the target large language model can also be optimized based on user feedback information, such as manually regularly evaluating the results generated by the target large language model, optimizing the updated fine-tuning data in the updated training data set based on the evaluation results to obtain optimized fine-tuning data, and then performing reinforcement learning on the target large language model based on the optimized fine-tuning data to obtain a new target large language model, and using this new target large language model as the final trained target large language model.

[0140] Optionally, since any updated fine-tuning data of the updated training data set includes ASR text, the first summary, the first chain of thought, and the first evidence reason, when the user performs regular evaluation on the trained target large language model, a test data set can be constructed, and the test data set includes at least one evaluation data, such as the first evaluation data. Optionally, at least one updated fine-tuning data can be selected from the updated training data set as the first evaluation data, so as to construct a test data set based on the selected at least one first evaluation data, and then the trained target large language model is evaluated based on the test data set.

[0141] Optionally, when performing the evaluation process, for any first evaluation data in the test data set, the ASR text and the first chain of thought in the first evaluation data can be input into the trained target large language model, and the third summary and the third evidence reason are output. Among them, the third summary is the summary of the target large language model's prediction of the ASR text in the first evaluation data. The third evidence reason is the evidence reason for the target large language model's prediction of the ASR text in the first evaluation data, which is used to support the focus of the predicted third summary.

[0142] Optionally, the user can manually check the third summary and the third evidence reason to determine whether there is an error, and then input corresponding feedback information into the device. It is also possible to compare and analyze the third summary and the third evidence reason with the first summary and the first evidence reason in the first evaluation data to determine the first similarity between the third summary and the first summary, and the second similarity between the third evidence reason and the first evidence reason. If there is a first similarity greater than the first similarity threshold (the threshold set in advance by the user), and / or the second similarity greater than the second similarity threshold (the threshold set in advance by the user), it is determined that the evaluation result is abnormal, and the feedback information input by the user based on the abnormal evaluation result can be obtained. Optionally, the first similarity threshold and the second similarity threshold can be the same or different.

[0143] Optionally, the feedback information input by the user can be checked. If the feedback information indicates that the first evaluation data has an error and needs to be corrected, the first summary and the first evidence reason in the first evaluation data can be updated, and the updating method can be manually adjusted by the user. If it is determined based on the feedback information that the first thinking chain in the first evaluation data needs to be adjusted, the first thinking chain can also be modified and adjusted based on the thinking chain related information input by the user, and the modification and adjustment method can also be manually adjusted by the user.

[0144] Optionally, the first evaluation data including the updated first summary, the updated first evidence reason, the updated first chain of thought and the ASR text can be used as the second evaluation data. The first evaluation data including the updated first summary, the updated first evidence reason, the updated first chain of thought and the ASR text can also be used as the second evaluation data.

[0145] Optionally, the target large language model can be trained again based on all the second evaluation data and other fine-tuning data in the updated training data set except the updated fine-tuning data corresponding to the second evaluation data, so as to optimize the model parameters of the target large language model until the preset training end conditions are met to obtain a new target large language model.

[0146] Optionally, the new target large language model may continue to be evaluated. If the evaluation result is abnormal, the evaluation data may continue to be updated to train and optimize the target large language model again based on the updated evaluation data.

[0147] Optionally, a summary may be generated for the ASR text converted from the recording data according to the new target large language model.

[0148] In this embodiment, at least one updated fine-tuning data is selected as the first evaluation data in the updated training data set, and the ASR text and the first chain of thought in the first evaluation data are input into the trained target large language model, and the third summary and the third evidence reason are output. After receiving the feedback information input by the user, the first summary and the first evidence reason in the first evaluation data can be adjusted according to the feedback information, and the first chain of thought can also be adjusted to obtain the second evaluation data, and then the trained target large language model is optimized again according to the second evaluation data until the preset training end condition is met, and a new target large language model is obtained. That is, when training the target large language model, the target large language model can be adjusted according to the actual needs of the user, thereby improving the accuracy of the target large language model in summarizing.

[0149] In addition, to achieve the above purpose, the present application also proposes a summary generation device, including:

[0150] A determination module, configured to determine each fine-tuning data in a preset training data set, wherein the fine-tuning data includes an ASR text and a first summary, and the first summary includes a plurality of focus points in the ASR text;

[0151] A determination module, configured to determine a first chain of thought and a first evidence reason corresponding to the fine-tuning data according to a preset large model, wherein the first chain of thought is used to represent a logical link for generating a first summary according to the ASR text, and the first evidence reason is used to support a focus point in the first summary;

[0152] An updating module, used for updating each fine-tuning data in the training data set according to the first chain of thought and the first evidence reason corresponding to each fine-tuning data, to obtain an updated training data set, wherein each updated fine-tuning data in the updated training data set includes an ASR text, a first summary, a first chain of thought and a first evidence reason;

[0153] The training module is used to optimize the target large language model to be trained according to the updated training data set until the preset training end condition is met, and the trained target large language model is obtained, wherein the summary is generated according to the trained target large language model.

[0154] The summary generation device provided in the present application adopts the summary generation method in the above embodiment, and can generate a summary corresponding to the ASR text according to the target large language model after converting the recording content into ASR text. Compared with the prior art, the beneficial effects of the summary generation device provided in the present application are the same as the beneficial effects of the summary generation method provided in the above embodiment, and other technical features in the summary generation device are the same as the features disclosed in the above embodiment method, which will not be repeated here.

[0155] The present application provides a summary generation device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the summary generation method in the above-mentioned embodiment 1.

[0156] Reference below Figure 5 , which shows a schematic diagram of the structure of a summary generation device suitable for implementing the embodiment of the present application. The summary generation device in the embodiment of the present application may include but is not limited to mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 5 The summary generation device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0157] like Figure 5 As shown, the summary generation device may include a processing device 1001 (e.g., a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 to a random access memory (RAM) 1004. In the RAM 1004, various programs and data required for the operation of the summary generation device are also stored. The processing device 1001, the ROM 1002, and the RAM 1004 are connected to each other through a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Generally, the following systems can be connected to the I / O interface 1006: an input device 1007 including, for example, a touch screen, a touch pad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; an output device 1008 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 may allow the summary generation device to communicate with other devices wirelessly or by wire to exchange data. Although the figure shows a summary generation device with various systems, it should be understood that it is not required to implement or have each of the systems shown. More or fewer systems may be implemented or have instead.

[0158] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are executed.

[0159] The summary generation device provided by the present application adopts the summary generation method in the above embodiment, and can generate a summary corresponding to the ASR text according to the target large language model after converting the recording content into ASR text. Compared with the prior art, the beneficial effects of the summary generation device provided by the present application are the same as the beneficial effects of the summary generation method provided by the above embodiment, and other technical features in the summary generation device are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.

[0160] It should be understood that the various parts disclosed in this application can be implemented by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0161] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0162] The present application provides a computer-readable storage medium having computer-readable program instructions (ie, computer program) stored thereon, and the computer-readable program instructions are used to execute the summary generation method in the above embodiment.

[0163] The computer-readable storage medium provided in the present application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, systems or devices, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination of the above.

[0164] The computer-readable storage medium may be included in the summary generation device; or may exist independently without being assembled into the summary generation device.

[0165] The computer-readable storage medium carries one or more programs. When the one or more programs are executed by a summary generation device, the summary generation device can perform the steps of the summary generation method.

[0166] Computer program code for performing the operations of the present application may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0167] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0168] The modules involved in the embodiments described in this application may be implemented by software or hardware, wherein the name of the module does not constitute a limitation on the unit itself in some cases.

[0169] The readable storage medium provided in the present application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., a computer program) for executing the above-mentioned summary generation method, and can generate a summary corresponding to the ASR text according to the target large language model after converting the recording content into ASR text. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in the present application are the same as the beneficial effects of the summary generation method provided in the above-mentioned embodiment, which will not be repeated here.

[0170] The present application also provides a computer program product, including a computer program, which implements the steps of the above-mentioned summary generation method when executed by a processor.

[0171] The computer program product provided in this application can generate a summary corresponding to the ASR text according to the target large language model after converting the recording content into ASR text. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as the beneficial effects of the summary generation method provided in the above embodiment, which will not be repeated here.

[0172] The above descriptions are only some embodiments of the present application, and are not intended to limit the patent scope of the present application. All equivalent structural changes made using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect applications in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A method for generating a summary, characterized in that: The method comprises: Determine each fine-tuning data in a preset training data set, wherein the fine-tuning data includes an ASR text and a first summary, and the first summary includes a plurality of focus points in the ASR text; Determine a first chain of thought and a first evidence reason corresponding to the fine-tuning data according to a preset large model, wherein the first chain of thought is used to represent a logical link for generating the first summary according to the ASR text, and the first evidence reason is used to support a focus point in the first summary; According to the first thought chain and the first evidence reason corresponding to each fine-tuning data, each fine-tuning data in the training data set is updated to obtain an updated training data set, wherein any updated fine-tuning data in the updated training data set includes the ASR text, the first summary, the first thought chain and the first evidence reason; The target large language model to be trained is optimized according to the updated training data set until a preset training end condition is met, thereby obtaining a trained target large language model, wherein a summary is generated according to the trained target large language model.

2. The method for generating a summary according to claim 1, wherein: The step of determining the first chain of thought and the first evidence reason corresponding to the fine-tuning data according to the preset macro model includes: When the fine-tuning data is public text data, determining an external device having a preset macro model; Sending the fine-tuning data to the external device, wherein the external device inputs the fine-tuning data into a preset large model to obtain a first chain of thought and a first evidence reason; receiving a first chain of thought and a first reason of evidence fed back by the external device according to the fine-tuning data; or, When the fine-tuning data is private text data, determining a preset large model deployed locally; The fine-tuning data is input into the preset large model to obtain the first chain of thinking and the first evidence reason.

3. The method for generating a summary according to claim 1, wherein: The step of optimizing the target large language model to be trained based on the updated training data set includes: For any updated fine-tuning data in the updated training data set, the ASR text and the first thought chain are input into the target large language model to be trained to obtain a prediction result, wherein the prediction result includes a second summary and a second evidence reason; Calculate the knowledge distillation loss from the large model to the target large language model according to the prediction result, the first summary and the first evidence reason; The target large language model is optimized based on the knowledge distillation loss.

4. The method for generating a summary according to claim 3, wherein: After the step of inputting the ASR text and the first thought chain into the target large language model to be trained to obtain a prediction result, the method further comprises: Inputting the prediction result, and the first summary and the first evidence reason corresponding to the prediction result into the loss function corresponding to the target large language model, and calculating the prediction loss; The step of optimizing the target large language model based on the knowledge distillation loss includes: The target large language model is optimized based on the prediction loss and the knowledge distillation loss.

5. The method for generating a summary according to claim 3, wherein: The target large language model includes a model ontology and an adapter. The step of optimizing the target large language model based on the prediction loss and the knowledge distillation loss includes: A model parameter of the adapter is updated based on the knowledge distillation loss and the prediction loss to optimize the target large language model.

6. The method for generating a summary according to claim 1, wherein: After the step of obtaining the trained target large language model, the method further includes: Selecting at least one updated fine-tuning data in the updated training data set as the first evaluation data, inputting the ASR text and the first thought chain in the first evaluation data into the trained target large language model, and outputting a third summary and a third evidence reason; If feedback information input by the user based on the third summary and the third evidence reason is received, the first summary and the first evidence reason in the first evaluation data are adjusted according to the feedback information, and / or the first chain of thought in the first evaluation data is adjusted according to the feedback information to obtain second evaluation data; The trained target large language model is optimized again based on the second evaluation data until a preset training end condition is met, thereby obtaining a new target large language model.

7. The method for generating a summary according to any one of claims 1 to 6, characterized in that: The step of updating each fine-tuning data in the training data set according to the first chain of thought and the first evidence reason corresponding to each fine-tuning data includes: For any fine-tuning data in the training data set, the first chain of thought and the first evidence reason corresponding to the fine-tuning data are aligned with the fine-tuning data in text component units to obtain updated fine-tuning data, wherein the text component unit alignment includes attributing the fine-tuning data, the first chain of thought and the first evidence reason to the text component unit belonging to the same fine-tuning data.

8. A summary generation device, characterized in that: The device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the summary generation method according to any one of claims 1 to 7.

9. A medium, characterized in that The medium is a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the abstract generation method according to any one of claims 1 to 7 are implemented.

10. A computer program product, characterized in that The computer program product comprises a computer program, and when the computer program is executed by a processor, the steps of the abstract generation method according to any one of claims 1 to 7 are implemented.

Images