Sample data selection method and system for fine tuning training of dialogue abstract generation model
By screening high-quality training samples, using the ROUGE scores and semantic vector similarity of candidate digests, combined with diversity sampling, the high cost and inefficiency problem of large-scale language models in the dialogue summary generation task is solved, and efficient fine-tuning training is achieved.
Patent Information
- Application Number
- CN202510578682.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-07
- Publication Date
- 2025-08-15
AI Technical Summary
Existing large-scale language models have high and inefficient fine-tuning training in conversation summary generation tasks, requiring high-quality samples from training data to reduce computational costs and improve training efficiency.
By using the first model to generate candidate digests and calculate ROUGE scores, candidate digests with low scores are selected; using the second model to generate session data semantic vectors, calculate similarity and obtain training samples through diversity sampling to reduce the data size.
On the premise of ensuring model performance, the training cost is significantly reduced, the training efficiency is improved, and the effect of the model in the dialogue summary generation task is improved.
Smart Images
Figure CN120492925A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of natural language processing technology, and in particular to a method and system for selecting sample data for fine-tuning training of a conversation summary generation model. Background Art
[0002] Dialogue summarization aims to condense coherent conversation text into a concise summary that contains key information. Compared to traditional summarization, conversation text has a specific format and is applicable to a variety of scenarios. Although LLM can currently serve as a solver for general language tasks, supervised fine-tuning is still required for DS tasks to adapt the model to these different conversation scenarios and specific summary formats.
[0003] Parameter Efficient Fine Tuning (PEFT) for open-source LLMs has recently garnered significant attention. This method uses only a small number of parameters in addition to the model parameters during training, significantly reducing training time while also avoiding issues such as poor model performance and catastrophic forgetting. However, performing PEFT on large-scale LLM data remains computationally expensive. Therefore, extracting more efficient data from training data for model training, thereby reducing computational costs and improving fine-tuning training efficiency, is crucial to addressing this issue. Summary of the Invention
[0004] To this end, the present invention provides a sample data selection method and system for fine-tuning training of a dialogue summary generation model, which solves the problem of unsatisfactory training cost and efficiency caused by insufficient training data in existing large-scale model fine-tuning training. Through a data selection strategy, high-instruction training samples are screened from the training data to ensure model performance during model fine-tuning training while reducing computing costs and improving efficiency.
[0005] According to the design scheme provided by the present invention, on the one hand, a method for selecting sample data for fine-tuning training of a conversation summary generation model is provided, comprising:
[0006] Obtaining a conversation summary generation model training sample dataset, where the training sample dataset includes conversation data and reference summaries corresponding to the conversation data;
[0007] Generate candidate summaries for each conversation data in the training sample dataset using the first model, calculate ROUGE scores for the candidate summaries and corresponding reference summaries and sort the scores, select candidate summaries and reference summaries with scores less than a first threshold, and obtain a candidate sample dataset based on the selected candidate summaries and reference summaries;
[0008] The second model is used to generate semantic vectors of conversation data in the candidate sample dataset, the similarity of the semantic vectors of conversation data in the candidate sample dataset is calculated, and based on the similarity and through diversity sampling, sample data for fine-tuning training of the conversation summary generation model is obtained.
[0009] As a method for selecting sample data for fine-tuning training of a conversation summary generation model of the present invention, further generating candidate summaries of each conversation data in a training sample dataset using the first model includes:
[0010] Construct the first prompt word based on the summary generation task;
[0011] The first prompt word is spliced with the conversation data in the training sample data set, the spliced data is input into the first model, and the first model is used to generate a candidate summary corresponding to the conversation data.
[0012] As a sample data selection method for fine-tuning the conversation summary generation model of the present invention, further calculating the ROUGE scores of the candidate summaries and the corresponding reference summaries and ranking the scores includes:
[0013] The ROUGE method is used to calculate the word overlap between the candidate summary corresponding to the conversation data and the reference summary, and the calculated overlap is used as the ROUGE score between the candidate summary corresponding to the conversation data and the reference summary;
[0014] Sort the conversation data in the training sample dataset from low to high according to the ROUGE score.
[0015] As a sample data selection method for fine-tuning and training a conversation summary generation model of the present invention, further generating a semantic vector of conversation data in a candidate sample dataset using a second model includes:
[0016] Building a second model using the open source model, the second model including an embedding layer, an average pooling layer, and an L2 regularization layer;
[0017] The conversation data in the candidate sample dataset is input into the second model to generate word embedding vectors for the conversation data using the embedding layer in the second model. The word embedding vectors are then reduced in dimension through the average pooling layer, and the L2 regularization layer is used to convert the reduced dimensionality word embedding vectors into unit semantic vectors for the conversation data.
[0018] As a method for selecting sample data for fine-tuning training of a conversation summary generation model according to the present invention, further, based on similarity and through diversity sampling, sample data for fine-tuning training of a conversation summary generation model is obtained, including:
[0019] Divide the candidate sample dataset into p subsets, and calculate the similarity between the candidate sample dataset and the conversation data in each subset;
[0020] Based on similarity, a domain diversity sampling method is used to sample and obtain q training samples from the subset. The domain diversity sampling method is based on the similarity of the session data in the subset and the similarity of the session data in the candidate sample dataset and uses hyperparameters to balance the similarity of the session data in the subset and the difference of the session data in the candidate sample dataset;
[0021] The p×q training samples sampled from the p subsets are used as sample data pairs for fine-tuning training of the dialogue summary generation model, where p and q are both integers greater than 1.
[0022] As a sample data selection method for fine-tuning the conversation summary generation model of the present invention, the sampling process of the intra-domain diversity sampling method is further expressed as follows:
[0023] Among them, S represents the candidate sample dataset, U represents the unselected sample dataset, sim(x1,x2) represents the similarity between the original conversation data x1 and x2, and λ is a hyperparameter.
[0024] In another aspect, the present invention further provides a sample data selection system for fine-tuning training of a dialogue summary generation model, comprising: a data acquisition module, a data screening module, and a data selection module, wherein:
[0025] A data acquisition module is used to acquire a training sample dataset for a conversation summary generation model, wherein the training sample dataset includes conversation data and a reference summary corresponding to the conversation data;
[0026] a data screening module, configured to generate candidate summaries for each conversation data in the training sample dataset using the first model, calculate and sort the ROUGE scores of the candidate summaries and corresponding reference summaries, select candidate summaries and reference summaries with scores less than a first threshold, and obtain a candidate sample dataset based on the selected candidate summaries and reference summaries;
[0027] The data selection module is used to use the second model to generate semantic vectors of conversation data in the candidate sample data set, calculate the similarity of the semantic vectors of conversation data in the candidate sample data set, and obtain sample data pairs for fine-tuning training of the conversation summary generation model based on the similarity and through diversity sampling.
[0028] In another aspect, the present invention further provides a method for fine-tuning a conversation summary generation model, comprising the following:
[0029] Obtaining conversation sample data pairs based on the above sample data selection method;
[0030] The obtained conversation sample data is used to fine-tune the conversation summary generation model.
[0031] Beneficial effects of the present invention:
[0032] The present invention selects a small number of high-quality training samples from the training dataset to perform fine-tuning training on the important language model for the conversation summarization task, reducing the scale of training data, improving training efficiency, and making the performance of the trained LLM comparable to that of training on the complete dataset. It has good application prospects in natural language processing fields such as conversation summary generation. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 This is a diagram illustrating the sample data selection process for fine-tuning the conversation summary generation model in the embodiment;
[0034] Figure 2 This is a schematic diagram of the sample data selection algorithm architecture in the embodiment;
[0035] Figure 3 This is a schematic diagram of closed-source LLM-assisted prediction in the embodiment;
[0036] Figure 4 Schematic diagram of the diversity sampling process in the embodiment. DETAILED DESCRIPTION
[0037] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention is further described in detail below with reference to the accompanying drawings and technical solutions.
[0038] In the conversation summarization task, PEFT training of the open-source LLM requires a large amount of text data, which is computationally expensive. The challenge is to select high-quality training samples from this data to reduce the data size, improve training efficiency, and ensure that the trained LLM performs comparable to that trained on the full dataset.
[0039] First of all, from the perspective of the summary task, it is about how to effectively evaluate data samples in the dataset. It is necessary to define evaluation indicators of sample quality in the dialogue summary task, such as the accuracy, completeness, and consistency of the summary.
[0040] Secondly, regarding the data selection problem, what methods should be used to screen the data? Traditional methods include: using clustering algorithms to group data and identify groups of samples with similar characteristics, thereby selecting highly representative samples; and outlier detection to identify and exclude abnormal samples that differ significantly from the majority of samples and may have a negative impact on model training.
[0041] Finally, what methods should be used to evaluate the performance of the trained model? We can evaluate the performance of the open-source LLM after PEFT on the conversation summarization test dataset. In the field of text summarization, the ROUGE method is commonly used for evaluation.
[0042] Since different data samples will affect the model performance to different degrees during the PEFT process, in order to more comprehensively and effectively analyze and evaluate the quality of each data sample in the data set, the embodiment of the present invention, such as Figure 1 As shown in the figure, a method for selecting sample data for fine-tuning training of a conversation summary generation model is provided, which specifically includes the following contents:
[0043] S101: Obtain a training sample dataset for a conversation summary generation model, where the training sample dataset includes conversation data and reference summaries corresponding to the conversation data;
[0044] S102: Generate candidate summaries for each conversation data in the training sample dataset using the first model, calculate ROUGE scores for the candidate summaries and corresponding reference summaries, sort the scores, select candidate summaries and reference summaries with scores less than a first threshold, and obtain a candidate sample dataset based on the selected candidate summaries and reference summaries;
[0045] S103. Generate semantic vectors of conversation data in the candidate sample dataset using the second model, calculate the similarity of the semantic vectors of conversation data in the candidate sample dataset, and obtain sample data pairs for fine-tuning training of the conversation summary generation model based on the similarity and through diversity sampling.
[0046] like Figure 2 In the algorithmic framework shown, the closed-source LLM is first used as the first model to quickly generate candidate summaries for the entire training dataset. These summaries are then ranked according to their ROUGE scores relative to the reference summaries, and the portion of data with a lower prediction score (LPS) is selected. Next, the open-source LLM is used as the second model to generate semantic vectors for the LPS data. The similarities between these vectors are calculated, and those with high similarities are removed, selecting a small number of diverse samples. During the model fine-tuning training (PEFT) phase, the open-source LLM is fine-tuned using the final 500 data samples to improve efficiency.
[0047] The candidate summary of each conversation data in the training sample dataset is generated by using the first model, which can be designed to include:
[0048] Construct the first prompt word based on the summary generation task;
[0049] The first prompt word is spliced with the conversation data in the training sample data set, the spliced data is input into the first model, and the first model is used to generate a candidate summary corresponding to the conversation data.
[0050] Currently, existing LLMs already have the ability to summarize conversation texts into short summaries. Therefore, in this embodiment, candidate summaries will be generated with the assistance of LLM to obtain an overall picture of the data set. However, the deployment cost of open source LLMs is high and requires a lot of computing costs. In order to improve the efficiency of the selection method, the API of more powerful closed-source LLMs (such as ChatGPT, ChatGLM4, etc.) is selected to quickly generate candidate summaries, and it can improve the accuracy of the summary results. The complete data set can be evaluated and the data set can be sorted according to the evaluation results.
[0051] like Figure 3 As shown, for the complete labeled conversation summary training dataset X, the training samples are s=(x,y label ) indicates that the original dialogue is x and the reference summary is y label To generate a short candidate summary using the closed-source LLM, set the prompt to “Summarize the following dialogue in one sentence:”.
[0052] First, the prompt word is concatenated with the original text to generate the corresponding JSON file; then, the corresponding candidate summary is generated by calling the closed-source LLM API tool and represented as y predict (As shown in Formula 1).
[0053] y predict =Predict LLM (prompt,x) (1)
[0054] The ROUGE method can be used to calculate the word overlap between the candidate summary corresponding to the conversation data and the reference summary, and the calculated overlap is used as the ROUGE score between the candidate summary corresponding to the conversation data and the reference summary; the conversation data in the training sample dataset is sorted from low to high according to the ROUGE score.
[0055] Among them, using y predict with y label The similarity between them is evaluated by the overlap of the vocabulary between them. π(s) represents the ROUGE-1 score of the corresponding sample s, as shown in Formula 2:
[0056] π(s)=Score ROUGE-1 (y predict ,y label ) (2)
[0057] Sort the samples s from low to high according to the π(s) score. After sorting, the sample sequence X={s1,s2,…,s |X|}. Correspondingly, π(si )≤π(s i+1 )(0 <i<n)。
[0058] For some data sets with high ROUGE-1 scores, it indicates that the generated summaries have significant word overlap with the reference summaries. This is due, firstly, to the word overlap between the reference summaries and the original text, which makes it relatively easy for the LLM to generate a highly similar summary by simply extracting key information and organizing it. Secondly, the language style learned by the LLM during pre-training is similar to that of the reference summaries, resulting in word matching in the candidate summaries. Using these training sets in PEFT would not significantly improve the model's performance. Therefore, during data selection, these data sets should be removed to reduce the size of the fine-tuning training dataset.
[0059] The opposite is true for the data with lower scores. This is often due to the high degree of condensed human-written reference summaries, or the significant deviation in the language style of the LLM pre-training data from the reference summaries. This may be "new knowledge" for the LLM. Supervised fine-tuning training can help the LLM fit this data, thereby improving its generalization ability to new tasks and new data.
[0060] Therefore, in this embodiment, with the help of the prediction results of LLM, the candidate dataset X with a lower prediction score (Lower PredictScore, LPS) is selected. c (accounting for p% of the original data set) for the next sampling process.
[0061] A dataset with a lower prediction score was obtained. To further reduce the size of the dataset, the LPS data can be screened based on the diversity sampling method to remove data samples with high similarity.
[0062] Specifically, the second model is used to generate the semantic vector of the conversation data in the candidate sample dataset, which can be designed to include:
[0063] Building a second model using the open source model, the second model including an embedding layer, an average pooling layer, and an L2 regularization layer;
[0064] The conversation data in the candidate sample dataset is input into the second model to generate word embedding vectors for the conversation data using the embedding layer in the second model. The word embedding vectors are then reduced in dimension through the average pooling layer, and the L2 regularization layer is used to convert the reduced dimensionality word embedding vectors into unit semantic vectors for the conversation data.
[0065] like Figure 4As shown in the figure, after obtaining the candidate data, we use the open-source LLM word embedding to obtain the word embedding of the original text x. Based on this, we use mean pooling to reduce the dimensionality. Finally, we use L2 regularization to convert the embedding into a unit vector e(x) (as shown in Formula 4), which is used for similarity calculation in the subsequent sampling method.
[0066] e(x)=L2Norm(MeanPooling(Embedding LLM (x))) (4)
[0067] In conversation summarization tasks, key information in the conversation is relatively scattered, and even multiple topics may appear. Traditional clustering methods based on topic extraction are not applicable. Therefore, sentence vectors generated using the open-source LLM contain the most complete conversation information. Furthermore, similarity calculation methods can be used to identify the distribution characteristics of the data.
[0068] To obtain text vectors more accurately, LLM has advantages over the traditional BERT model because it uses more text data and a larger number of parameters during pre-training, and the calculated text vectors contain richer semantic information. Furthermore, LLM is more advantageous when processing long texts. Therefore, in this embodiment, the LLM used in fine-tuning training is used to calculate text vectors, ensuring consistency and effectiveness from data selection to fine-tuning training.
[0069] The candidate sample dataset is divided into p subsets, and the similarity of the candidate sample dataset and the conversation data in each subset is calculated respectively; based on the similarity and using the intra-domain diversity sampling method, q training samples are sampled and obtained from the subsets. The intra-domain diversity sampling method is based on the similarity of the conversation data in the subset and the similarity of the conversation data in the candidate sample dataset and uses hyperparameters to balance the similarity of the conversation data in the subset and the difference of the conversation data in the candidate sample dataset; the p×q training samples sampled from the p subsets are used as sample data pairs for fine-tuning training of the conversation summary generation model, where p and q are both integers greater than 1.
[0070] In-domain diversity sampling (IDDS) is a sampling method used in active learning. Its basic idea is to maintain the difference between the selected samples to improve the diversity of the sampling. Its formula is as follows:
[0071]
[0072] sim(x1,x2)=<e(x1),e(x2)> (6)
[0073] Here, U represents the set of all unselected samples, and S represents the set of selected samples. sim(x1,x2) represents the similarity between the original samples x1 and x2. λ is a hyperparameter used to balance the similarity of the former with the difference of the latter.
[0074] Since this method needs to continuously update the samples of the unselected and selected sets, and the number of unselected samples is relatively large, the computational efficiency during the sampling period is reduced. To address this problem, in this embodiment, the candidate data set X is first c (accounting for p% of the original data set) is divided into p data subsets X={X1,X2,…,X p}, for subset X i Sample text x in the calculation of X i The average similarity within the selected samples is used to replace the average similarity between unselected samples.
[0075] Finally, q training samples are selected from each subset through the diversity sampling method, and finally p×q training samples are obtained for model fine-tuning training.
[0076] Furthermore, based on the above method, an embodiment of the present invention also provides a sample data selection system for fine-tuning training of a conversation summary generation model, comprising: a data acquisition module, a data screening module, and a data selection module, wherein:
[0077] A data acquisition module is used to acquire a training sample dataset for a conversation summary generation model, wherein the training sample dataset includes conversation data and a reference summary corresponding to the conversation data;
[0078] a data screening module, configured to generate candidate summaries for each conversation data in the training sample dataset using the first model, calculate and sort the ROUGE scores of the candidate summaries and corresponding reference summaries, select candidate summaries and reference summaries with scores less than a first threshold, and obtain a candidate sample dataset based on the selected candidate summaries and reference summaries;
[0079] The data selection module is used to use the second model to generate semantic vectors of conversation data in the candidate sample data set, calculate the similarity of the semantic vectors of conversation data in the candidate sample data set, and obtain sample data pairs for fine-tuning training of the conversation summary generation model based on the similarity and through diversity sampling.
[0080] Furthermore, an embodiment of the present invention also provides a method for fine-tuning a conversation summary generation model, comprising the following:
[0081] Obtaining conversation sample data pairs based on the above sample data selection method;
[0082] The obtained conversation sample data is used to fine-tune the conversation summary generation model.
[0083] Unless otherwise specifically stated, the relative steps, numerical expressions and values of the components and steps set forth in these embodiments do not limit the scope of the present invention.
[0084] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.
[0085] The units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person of ordinary skill in the art may use different methods to implement the described functions for each specific application, but such implementation is not considered to be beyond the scope of the present invention.
[0086] Those skilled in the art will appreciate that all or part of the steps in the above method can be performed by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a magnetic disk, or an optical disk. Alternatively, all or part of the steps in the above embodiment can be implemented using one or more integrated circuits. Accordingly, each module / unit in the above embodiment can be implemented in the form of hardware or software functional modules. The present invention is not limited to any specific combination of hardware and software.
[0087] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the above-described embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above-described embodiments within the technical scope disclosed by the present invention, or replace some of the technical features therein with equivalents. Such modifications, changes, or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.
Claims
1. A method for selecting sample data for fine-tuning training of a conversation summary generation model, characterized in that: Include: Obtaining a conversation summary generation model training sample dataset, where the training sample dataset includes conversation data and reference summaries corresponding to the conversation data; Generate candidate summaries for each conversation data in the training sample dataset using the first model, calculate ROUGE scores for the candidate summaries and corresponding reference summaries and sort the scores, select candidate summaries and reference summaries with scores less than a first threshold, and obtain a candidate sample dataset based on the selected candidate summaries and reference summaries; The second model is used to generate semantic vectors of conversation data in the candidate sample dataset, the similarity of the semantic vectors of conversation data in the candidate sample dataset is calculated, and based on the similarity and through diversity sampling, sample data pairs for fine-tuning training of the conversation summary generation model are obtained.
2. The method for selecting sample data for fine-tuning training of a conversation summary generation model according to claim 1, characterized in that: Generate candidate summaries for each conversation data in the training sample dataset using the first model, including: Construct the first prompt word based on the summary generation task; The first prompt word is spliced with the conversation data in the training sample data set, the spliced data is input into the first model, and the first model is used to generate a candidate summary corresponding to the conversation data.
3. The method for selecting sample data for fine-tuning training of a conversation summary generation model according to claim 1, characterized in that: Calculate the ROUGE scores of the candidate abstracts and the corresponding reference abstracts and sort the scores, including: The ROUGE method is used to calculate the word overlap between the candidate summary corresponding to the conversation data and the reference summary, and the calculated overlap is used as the ROUGE score between the candidate summary corresponding to the conversation data and the reference summary; Sort the conversation data in the training sample dataset from low to high according to the ROUGE score.
4. The method for selecting sample data for fine-tuning training of a conversation summary generation model according to claim 1, characterized in that: The second model is used to generate semantic vectors of conversation data in the candidate sample dataset, including: Building a second model using the open source model, the second model including an embedding layer, an average pooling layer, and an L2 regularization layer; The conversation data in the candidate sample dataset is input into the second model to generate word embedding vectors for the conversation data using the embedding layer in the second model. The word embedding vectors are then reduced in dimension through the average pooling layer, and the L2 regularization layer is used to convert the reduced dimensionality word embedding vectors into unit semantic vectors for the conversation data.
5. The method for selecting sample data for fine-tuning training of a conversation summary generation model according to claim 1, characterized in that: Based on similarity and through diversity sampling, we obtain sample data pairs for fine-tuning the conversation summary generation model, including: Divide the candidate sample dataset into p subsets, and calculate the similarity between the candidate sample dataset and the conversation data in each subset; Based on similarity, a domain diversity sampling method is used to sample and obtain q training samples from the subset. The domain diversity sampling method is based on the similarity of the session data in the subset and the similarity of the session data in the candidate sample dataset and uses hyperparameters to balance the similarity of the session data in the subset and the difference of the session data in the candidate sample dataset; The p×q training samples sampled from the p subsets are used as sample data pairs for fine-tuning training of the dialogue summary generation model, where p and q are both integers greater than 1.
6. The method for selecting sample data for fine-tuning training of a conversation summary generation model according to claim 5, characterized in that: The sampling process of the intra-domain diversity sampling method is expressed as: Among them, S represents the candidate sample dataset, U represents the unselected sample dataset, sim(x1,x2) represents the similarity between the original conversation data x1 and x2, and λ is a hyperparameter.
7. A sample data selection system for fine-tuning training of a conversation summary generation model, characterized by: Contains: data acquisition module, data screening module and data selection module, among which, A data acquisition module is used to acquire a training sample dataset for a conversation summary generation model, wherein the training sample dataset includes conversation data and a reference summary corresponding to the conversation data; a data screening module, configured to generate candidate summaries for each conversation data in the training sample dataset using the first model, calculate and sort the ROUGE scores of the candidate summaries and corresponding reference summaries, select candidate summaries and reference summaries with scores less than a first threshold, and obtain a candidate sample dataset based on the selected candidate summaries and reference summaries; The data selection module is used to use the second model to generate semantic vectors of conversation data in the candidate sample data set, calculate the similarity of the semantic vectors of conversation data in the candidate sample data set, and obtain sample data pairs for fine-tuning training of the conversation summary generation model based on the similarity and through diversity sampling.
8. A fine-tuning training method for a dialogue summary generation model, characterized in that: Contains the following content: Obtaining a session sample data pair based on the method of claim 1; The obtained conversation sample data is used to fine-tune the conversation summary generation model.
9. An electronic device, characterized in that: include: at least one processor, and a memory coupled to the at least one processor; The memory stores a computer program, and the computer program can be executed by the at least one processor to implement the method according to any one of claims 1 to 6.
10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed, the method according to any one of claims 1 to 6 can be implemented.