Text summarization generation method based on multi-teacher feedback learning and dynamic elimination

Abstract

The embodiment of the application provides a text abstract generation method based on multi-teacher feedback learning and dynamic elimination, and belongs to the technical field of data processing, and specifically comprises the following steps: step 1, obtaining a sample data set and dividing it into a training set and a validation set; step 2, constructing a Prompt template, using the Prompt template to perform supervised instruction fine-tuning with a reference abstract; step 3, inputting the training set into a main student model to obtain a difficult sample set, training and evaluating the current main student model according to the multi-teacher model set and the difficult sample set, and obtaining the main student model for the next round of training; step 4, when each round of training is completed, eliminating a teacher model according to a preset teacher elimination rule; step 5, repeating steps 3 and 4 until an iteration stopping condition is reached; and step 6, inputting a target text into an abstract generation model to obtain a target abstract. Through the scheme of the application, the generation efficiency, accuracy and adaptability are improved.

Description

TECHNICAL FIELD

[0001] The embodiment of the application relates to the technical field of data processing, and particularly relates to a text abstract generation method based on multi-teacher feedback learning and dynamic elimination. BACKGROUND

[0002] At present, a large language model (LLM) exhibits strong capability in a text generation task, and a method based on reinforcement learning exhibits great potential in improving the quality of a generation model, but the method mainly depends on a numerical reward signal, feedback information is compressed into a reward value, and only limited information can be provided, the model cannot know how to specifically improve the behavior from the reward value, and can only explore a better strategy through continuous trial and error, which leads to a slow and inefficient learning process. OpenAI and other institutions further propose a reinforcement learning based on human feedback (RLHF) method, a reward model is trained by letting humans compare preferences for model output, and the abstract model is optimized by the reward model. However, RLHF usually depends on high-quality artificial feedback, and the cost is high and it is difficult to obtain on a large scale, and human preference comparison only provides limited binary feedback information, and the amount of effective information conveyed by each evaluation is low.

[0003] Therefore, a text abstract generation method based on multi-teacher feedback learning and dynamic elimination with high generation efficiency, precision and adaptability is urgently needed. SUMMARY

[0004] Therefore, the embodiment of the application provides a text abstract generation method based on multi-teacher feedback learning and dynamic elimination, which at least partially solves the problem of poor generation efficiency, precision and adaptability in the prior art.

[0005] The embodiment of the application provides a text abstract generation method based on multi-teacher feedback learning and dynamic elimination, which comprises the following steps:

[0006] Step 1, obtaining a sample data set and dividing it into a training set and a validation set, wherein the sample data set comprises sample texts and corresponding reference abstracts;

[0007] Step 2, generating a preliminary abstract for the sample texts in the training set and constructing a Prompt template according to the preliminary abstract, using the Prompt template and the reference abstract to perform supervised instruction fine-tuning to obtain a main student model;

[0008] Step 3, initializing a multi-teacher model set, inputting the training set into the main student model to obtain a difficult sample set, and training and evaluating the main student model based on the multi-teacher model set and the difficult sample set to obtain a main student model for the next round of training;

[0009] Step 4: At the end of each training round, teacher models are eliminated according to the preset teacher elimination rules, and a new set of multi-teacher models is obtained to enter the next training round.

[0010] Step 5: Repeat steps 3 and 4 until the iteration stopping condition is met to obtain the summary generation model;

[0011] Step 6: Input the target text into the summary generation model to obtain the target summary.

[0012] According to a specific implementation of an embodiment of the present invention, step 2 specifically includes:

[0013] Step 2.1: Use a pre-trained model to generate a draft summary for the sample text in the training set;

[0014] Step 2.2: Construct a general feedback template for the first draft abstract to generate initial feedback suggestions;

[0015] Step 2.3: Construct a Prompt template based on the sample text, draft abstract, initial feedback suggestions, and instructions;

[0016] Step 2.4: Use the Prompt and reference summary to fine-tune the supervised instructions, calculate the first loss function, and initialize the student model accordingly. The initialized student model This serves as the master student model for the first round of multi-teacher feedback learning training.

[0017] According to a specific implementation of an embodiment of the present invention, the first loss function is:

[0018] ;

[0019] in, For student model parameters, For reference summary The sequence length, For time step index, For target summary The target token at time step t. For the target summary One token, Student model The conditional probability distribution is given by , where i is the index of the training sample and N is the total number of samples. This indicates the Prompt template.

[0020] According to a specific implementation of an embodiment of the present invention, step 3 specifically includes:

[0021] Step 3.1: Construct a set of multi-teacher models with complementary styles;

[0022] Step 3.2: Generate predicted summaries for the sample texts in the training set using the current round of the master student model, calculate the difference between the predicted summaries and the reference summaries, and use this difference to filter the set of difficult samples. ;

[0023] Step 3.3: Each teacher model in the multi-teacher model set generates target feedback suggestions for the difficult samples in the current round;

[0024] Step 3.4: Construct a feedback learning Prompt based on the sample text, predicted summary, target feedback suggestions, and instructions, and use the current master student model. Based on this, a set of student model branches, guided by different teacher models, are trained using feedback learning Prompt and a second loss function. Wherein, the second loss function is

[0025] ;

[0026] in, This represents the loss during the training of the student model branch in the r-th round. Represents the teacher model in the r-th round. For the sample Constructed feedback learning prompt;

[0027] Step 3.5: Evaluate each student branch model in the validation set, calculate the comprehensive score, and select the student branch model with the highest score as the master student model for the next round of training.

[0028] According to a specific implementation of an embodiment of the present invention, step 3.2 specifically includes:

[0029] Step 3.2.1: When the current multi-teacher feedback learning training round r=1, the difficult sample set is the sample text set, i.e. ;

[0030] Step 3.2.2: When the current multi-teacher feedback learning training round r>1, the current round master-student model... Generated prediction summary Reference Abstract By comparison, the difference values ​​are obtained:

[0031] ;

[0032] in, , For hyperparameters, satisfying , This represents the difference between the current predicted abstract and the reference abstract. This is a function for calculating the longest common subsequence overlap between the reference digest and the generated digest. This is a function for calculating the semantic similarity between the reference summary and the generated summary.

[0033] Step 3.2.3: Construct a set of difficult samples based on the difference values. :

[0034] ;

[0035] Among them, if Then this sample is considered a difficult sample in the current round. Let r be the set of difficult samples in the r-th round. is the threshold for screening difficult samples in the r-th round.

[0036] According to a specific implementation of an embodiment of the present invention, step 3.5 specifically includes:

[0037] Step 3.5.1: Randomly select a performance evaluation dataset according to a preset ratio to evaluate the performance of each student model branch;

[0038] Step 3.5.2, calculate each student model branch. In the validation set Overall score :

[0039] ;

[0040] ;

[0041] ;

[0042] ;

[0043] Teacher model The average ROUGE-L score of the guided student model branch on the r-th round validation set is used to measure the average word-level similarity between the student model's generated summary and the reference summary. Teacher model The average value of the guided student model branch on the r-th round validation set A score is used to measure the semantic similarity between the student-generated summary and the reference summary. Teacher model The average summary length penalty term of the guided student model branch on the r-th round validation set is used to measure the deviation in length between the generated summary and the reference summary. For hyperparameters, This indicates that the i-th validation sample is generated by the teacher model. Summary generated by the guided student model branch This is a reference summary for the validation sample, where i is the local index of the current validation set sample. It is a function for calculating the length of the abstract. The total number of samples in the validation set;

[0044] Step 3.5.3: Select the teacher index with the highest score on the validation set. :

[0045] ;

[0046] in, This indicates selecting the score from all teacher models. The largest teacher model, and numbered it as The teacher model that wins in this round of training;

[0047] Step 3.5.4: Select the student model branch with the highest score as the master student model for the next round of training.

[0048] ;

[0049] in, For the next round of feedback learning, the master student model This is the student model branch corresponding to the teacher model that wins in this round of training.

[0050] According to a specific implementation of an embodiment of the present invention, step 4 specifically includes:

[0051] Step 4.1: At the end of each training round, maintain a teacher failure counter. For the teacher model corresponding to the winner of the current round, clear its failure count to zero. For the remaining teacher models, increment the failure count by one. The winner is the teacher model corresponding to the student branch model with the highest score in this round.

[0052] Step 4.2: Eliminate one teacher in each round according to the teacher elimination rules and update the multi-teacher model set.

[0053] According to a specific implementation of an embodiment of the present invention, step 4.2 specifically includes:

[0054] Step 4.2.1: When the number of teacher models in the multi-teacher model set is equal to 1, continue to retain the unique teacher until the iteration stops.

[0055] Step 4.2.2: When the number of teacher models in the multi-teacher model set is greater than 1, then construct the set of teachers who did not win. :

[0056] ;

[0057] in, The group of teachers who did not win in round r. For the r-th round of teacher set;

[0058] Select the subset of teachers with the most failures :

[0059] ;

[0060] in, The set of teachers who failed the most times in round r. For each teacher model A failure counter that accumulates the number of rounds without a win;

[0061] In the teacher subset The player with the lowest score in the current round will be eliminated.

[0062] ;

[0063] in, The index of the teacher with the lowest score in round r;

[0064] Final update of the next round of teacher selection:

[0065] ;

[0066] in, For the (r+1)th round of teacher models, This is the model for eliminating teachers in the current round.

[0067] According to a specific implementation of an embodiment of the present invention, the iteration stopping condition is any one of the following: the improvement of the validation set index is below the threshold for two consecutive rounds, there is only one teacher left and the validation index has not improved significantly, there are no difficult samples in the current round, and the maximum number of training rounds has been reached.

[0068] The text summarization generation scheme based on multi-teacher feedback learning and dynamic elimination in this embodiment of the invention includes: Step 1, acquiring a sample dataset and dividing it into a training set and a validation set, wherein the sample dataset includes sample texts and their corresponding reference summaries; Step 2, generating initial draft summaries for the sample texts in the training set and constructing a Prompt template accordingly, using the Prompt template and the reference summary for supervised fine-tuning to obtain a master student model; Step 3, initializing a multi-teacher model set, inputting the training set into the master student model to obtain a difficult sample set, and training and evaluating the current master student model based on the multi-teacher model set and the difficult sample set to obtain the master student model for the next round of training; Step 4, at the end of each round of training, eliminating teacher models according to a preset teacher elimination rule to obtain a new multi-teacher model set for the next round of training; Step 5, repeating steps 3 and 4 until the iteration stopping condition is met to obtain a summary generation model; Step 6, inputting the target text into the summary generation model to obtain the target summary.

[0069] The beneficial effects of this invention are as follows: The invention proposes a "feedback learning" mechanism that uses natural language feedback as a supervisory signal, bypassing the reliance on numerical reward modeling in traditional reinforcement learning. It uses structured prompts to guide student models in feedback-responsive optimization of initial draft summaries. It proposes using multiple large language models as "teachers" to replace expensive human experts providing feedback, constructing a student-multi-teacher interactive feedback optimization framework. This generates fine-grained, complementary natural language feedback, selecting and retaining the most effective feedback paths, ensuring that student models receive guidance from multiple perspectives, significantly enhancing the quality and accuracy of summary generation. It introduces a difficult sample mining mechanism based on the gap between the generated summary and the reference summary. In each round of feedback learning, the model is prioritized for training on samples where "feedback is difficult to improve." By focusing on these "weak" samples, the student model can specifically optimize its generation capabilities, significantly improving the overall summary quality. Finally, it introduces a dynamic teacher elimination mechanism based on validation set performance, employing a "failure-first, score-order-assisted" strategy to eliminate teachers with poor feedback performance round by round, achieving continuous optimization of the feedback source. Attached Figure Description

[0070] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0071] Figure 1 A flowchart illustrating a text summarization method based on multi-teacher feedback learning and dynamic elimination, provided in an embodiment of the present invention;

[0072] Figure 2 This is an overall framework diagram of a text summarization method based on multi-teacher feedback learning and dynamic elimination, provided for an embodiment of the present invention.

[0073] Figure 3 This is a system framework diagram for a multi-teacher feedback learning stage provided in an embodiment of the present invention. Detailed Implementation

[0074] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0075] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0076] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this invention, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0077] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. The illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0078] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0079] This invention provides a text summarization method based on multi-teacher feedback learning and dynamic elimination, which can be applied to the text summarization process in text processing scenarios.

[0080] See Figure 1 This is a flowchart illustrating a text summarization method based on multi-teacher feedback learning and dynamic elimination, provided by an embodiment of the present invention. Figure 1 and Figure 2 As shown, the method mainly includes the following steps:

[0081] Step 1: Obtain the sample dataset and divide it into a training set and a validation set, wherein the sample dataset includes sample text and its corresponding reference summary;

[0082] In this embodiment of the invention, the relevant parameters are defined as follows:

[0083] Abstract of the first draft during the cold start phase; : The r-th round of the main student model For training samples The generated summary; :instruction; : The Prompt template constructed from the i-th training sample during the cold start phase; Loss function for training during the cold start phase; : The loss function for training the student model branch in the r-th round; Initial set of teachers; The set of teacher models in the rth round; : The r-th round of the main student model; : The main student model in the (r+1)th round; Teachers in the rth round The corresponding student model branch; j: teacher index; The index of the teacher with the highest score on the validation set; Index of teachers to be phased out; The student model corresponding to the winning teacher model in the r-th round of training; : The set of difficult samples in the r-th round; : Threshold for screening difficult samples in the rth round; Student Branching Model The overall score on the validation set; Each teacher model A failure counter that accumulates the number of rounds without a win; Teacher Model in Round r For the sample Constructed prompt; Teacher Model in Round r right medium sample The generated feedback; Maximum number of training epochs; splicing operation; The set of input values ​​that maximizes the objective function; The set of input values ​​that minimizes the objective function; : A function to calculate the longest common subsequence overlap between the reference digest and the generated digest; : A function that calculates the semantic similarity between the reference summary and the generated summary.

[0084] In practice, a Chinese abstract dataset with an "original text-reference abstract" structure can be used for training, where each sample is represented as a supervised pair. .in The input text can be sourced from various text resources such as news, encyclopedia entries, government documents, legal documents, and financial reports. This indicates the corresponding manually written reference abstract. The data format is not limited to a specific platform or data source, but it must meet the basic quality standards of clear text-abstract structure, coherent content, and good semantic alignment.

[0085] To enable multi-round training optimization, the dataset was divided into two functional modules:

[0086] training set : accounting for 70% of the total data, used in the cold start phase and the multi-teacher feedback learning phase.

[0087] Validation set This comprises 30% of the total data and is used for performance evaluation of the student model, selection of the master model, and statistics of teacher win rates in each round of training. The system can randomly sample 35% of this data as a subset for performance evaluation in each round of training.

[0088] Step 2: Generate a draft summary for the sample text in the training set and construct a Prompt template based on it. Use the Prompt template and the reference summary to fine-tune the model using supervised instructions to obtain the master student model.

[0089] In practice, before formally entering multi-teacher feedback learning, to adapt the student model to the input pattern of "original text + initial draft summary + feedback," we designed a feedback-style instruction fine-tuning method as a cold-start training method to initialize the model. In the absence of real feedback and initial draft summaries, we adopted the following automatic construction strategy:

[0090] (1) Use the publicly available pre-trained summary generation model (BART) for each training sample Generate a rough summary This is an initial draft summary for the cold start phase.

[0091] (2) Construct some general-style feedback templates as natural language feedback for the cold start phase of each training sample. These feedback suggestions do not rely on actual reference summaries, such as: please remove redundant expressions and improve sentence variety; please add details of important events mentioned in the text; please adjust the word order to make the content more fluent; please make the summary more concise.

[0092] (3) Setting instructions The prompt is: "Based on the original text and the initial draft abstract, please optimize the expression and information structure according to the feedback suggestions to generate an improved abstract."

[0093] (4) Finally, the text First draft summary Feedback and suggestions ,instruction spliced ​​as As input to the prompt, the prompt structure remains consistent throughout the cold start phase and subsequent multi-teacher feedback learning phases.

[0094] ;

[0095] in, The Prompt input constructed for the i-th training sample during the cold start phase. This is for splicing operations.

[0096] (5) Set the loss function for the cold start phase Based on the Qwen2.5-1.5B model, the target reference summary To achieve the goal, train the initial student model. This training process helps the model learn how to understand and respond to natural language feedback cues, thereby improving the efficiency and response stability of subsequent multi-round feedback learning. It is expressed as follows:

[0097] ;

[0098] in, For student model parameters, Reference Summary for Target The sequence length, For time step index, For target summary The target token at time step t. For the target summary One token, Student model The conditional probability distribution is given by , where i is the index of the training sample and N is the total number of samples.

[0099] After completing the feedback-style instruction fine-tuning during the cold start phase, the student model... Having already acquired the basic ability to "understand feedback and rewrite summaries," they will serve as the primary student model for the first round of multi-teacher feedback learning.

[0100] Step 3: Initialize the multi-teacher model set, input the training set into the master student model to obtain the difficult sample set, and train and evaluate the current master student model based on the multi-teacher model set and the difficult sample set to obtain the master student model for the next round of training.

[0101] In practice, after obtaining the initial master-student model, the process immediately enters the second stage—multi-teacher feedback learning. The master-student model generated in the cold start stage... Building upon this foundation, we further enhance the quality of the abstracts. This stage is the core of our method, namely, introducing feedback from multiple teachers to optimize the student model branches. We employ a fixed set of teachers and continuously strengthen the capabilities of the student branch model through validation set evaluation, hard sample mining, dynamic elimination mechanisms, and master model update strategies. This allows the model to gradually learn to generate better abstracts on its own. The main framework and process are as follows: Figure 3 As shown.

[0102] The complete training process is as follows:

[0103] (1) Initialization of the teacher set:

[0104] The initial set of teachers is set as follows: , where K>=3.

[0105] The teacher models should be language models with summarization and generation capabilities, without being limited to specific model structures, sizes, or numbers. They can be publicly available large-scale Chinese language models (such as the Qwen series, ChatGLM series, GPT, etc.) or user-defined generative models. Multiple teacher models should be complementary in terms of feedback angles, styles, or strategies to achieve differentiated guidance effects during multiple rounds of fine-tuning. It is recommended that at least one general-purpose LLM teacher, one summarization task-specific LLM teacher, and one language style optimization LLM teacher be included.

[0106] (2) Iterative training:

[0107] Repeat the following steps for multiple rounds of training until the termination condition is met:

[0108] Summary regeneration:

[0109] In the r-th round of training, the current master student model is used. Re-training samples Generate summary .

[0110] Difficult sample mining:

[0111] Not all training samples need to undergo multiple rounds of feedback interaction. For simple samples where the model can already provide good summaries, applying feedback may be ineffective and a waste of computational resources. Instead, we aim to prioritize training on difficult samples where the model performs poorly, thus maximizing the benefits of feedback. Difficult samples in the training data typically refer to those that the model cannot currently predict correctly or have large errors. These samples often contain special cases or highly challenging patterns. If the model can focus on learning from these difficult samples, it can significantly improve the model's robustness and overall performance.

[0112] In each round of training, difficult samples For training set To simplify expression and operation, the samples in the selected subset are renumbered locally in each round. .

[0113] When r=1, That is, the set of difficult samples is all training samples.

[0114] The first round is the "initial exposure feedback" stage. At this time, the student model has just finished cold start training, and its generation quality is still weak. There are obvious differences in the summary output of most training samples. We directly carry out subsequent feedback learning on all training samples.

[0115] When r > 1, the main process is as follows:

[0116] 1. Change the current master-student model Generated summary Reference Abstract contrast:

[0117] ;

[0118] in, , For hyperparameters, satisfying , This represents the difference between the current abstract and the reference abstract. It is a function that calculates the longest common subsequence overlap between the reference digest and the generated digest. It is a function that calculates the semantic similarity between the reference summary and the generated summary.

[0119] 2. Constructing a set of difficult samples :

[0120] ;

[0121] like This sample is then considered a difficult sample in the current round. Let r be the set of difficult samples in the r-th round. is the threshold for screening difficult samples in the r-th round.

[0122] By screening for difficult samples, we can focus our limited teacher feedback resources on the areas that can best improve model performance, thus avoiding overfitting or ineffective updates caused by repeated training on known knowledge.

[0123] Multi-teacher feedback generation:

[0124] Each teacher For difficult samples Generate feedback (e.g., evaluation of abstract quality, score, or suggestions for improvement), this stage is equivalent to each teacher independently reviewing the student model's abstract and providing guidance. Each teacher's feedback prompt is designed as follows: "Please read the following article and abstract, point out the main problems in the abstract, and provide feedback suggestions on how to optimize it: Article: { };summary:{ }".

[0125] Constructing a multi-teacher feedback learning Prompt input:

[0126] Ultimately text First draft summary Feedback and suggestions ,instruction spliced ​​as As a multi-teacher feedback learning prompt input

[0127] ;

[0128] in, Teacher Model in Round r For the sample The constructed prompt.

[0129] Student model branch training:

[0130] Using the current round's master student model The model is trained on difficult samples to obtain a set of student model branches guided by different teacher models. .

[0131] ;

[0132] in, The loss function for training the student model branch in the r-th round.

[0133] Validation set optimization and master-student model update:

[0134] Although the training samples in each round are dynamically selected hard sample sets, keeping the validation set unchanged may cause the teacher model to dominate on certain types of samples in the long run, thus continuously winning in multiple rounds of training and affecting the fairness of teacher evaluation. To mitigate this problem, we dynamically sample the validation set for teacher evaluation in each round of training. This strategy ensures that the teacher's winning results are more representative and also alleviates the overfitting or selection bias problems caused by the single validation sample.

[0135] ;

[0136] Specifically, from the verification set A subset of samples is randomly selected at a ratio of 35% as the validation set for the current round. This is used to evaluate the actual optimization effect of each teacher's feedback and to guide the selection of the master student model and the teacher win rate statistics. We also renumber the samples in the validation set locally. This ensures that the evaluation in each round is independent of the location of the global dataset.

[0137] The main process is as follows:

[0138] 1. Each student model branch In this round of validation set Calculate the overall score :

[0139] ;

[0140] ;

[0141] ;

[0142] ;

[0143] , representing the teacher model The average ROUGE-L score of the guided student model on the r-th round validation set measures the average similarity between the student model's generated summary and the reference summary in terms of word-level overlap.

[0144] Teacher model The average value of the student model under supervision on the r-th round of validation set The score measures the semantic similarity between the student-generated summary and the reference summary;

[0145] Teacher model The average summary length penalty term of the guided student model on the r-th round of validation set is used to measure the deviation of the generated summary from the reference summary in length. The closer it is to 1, the more reasonable the length is, and the greater the deviation, the greater the penalty.

[0146] in, For hyperparameters, This indicates that the i-th validation sample is generated by the teacher model. Summary generated by the guided student model branch This is a reference summary for the validation sample, and i is the local index of the current validation set sample. It is a function that calculates the longest common subsequence overlap between the reference digest and the generated digest. It is a function that calculates the semantic similarity between the reference summary and the generated summary. It is a function for calculating the length of the abstract. This represents the total number of samples in the validation set.

[0147] 2. Select the teacher index with the highest score on the validation set. :

[0148] ;

[0149] in, Indicated in all teachers Select score The largest teacher, and numbered it as The teachers who win in this round of training.

[0150] 3. Select the student model with the highest score as the main student model for the next round (r+1 round):

[0151] ;

[0152] in, For the next round of feedback learning, the master student model This is the student model corresponding to the teacher model that wins in this round of training.

[0153] By selecting based on validation set metrics, we ensure that each student model advancing to the next stage has the highest possible summary quality.

[0154] Step 4: When each round of training ends, teacher models are eliminated according to the preset teacher elimination rules, and a new set of multi-teacher models is obtained to enter the next round of training.

[0155] In practice, within a multi-teacher feedback framework, different teacher models may possess different expression styles, task preferences, or model biases, resulting in varying degrees of guidance from their feedback to student models. Retaining poorly performing teachers for extended periods not only wastes training resources but may also mislead the optimization direction of student models.

[0156] Therefore, we adopt a strategy of "eliminating one teacher per round," using a dual criterion of a failure round counter and validation set performance to determine which teacher to eliminate. This mechanism aims to progressively remove the teacher model with the worst overall feedback performance, retain the most instructive teachers, improve training efficiency, prevent redundant signal perturbations, and provide a higher-quality feedback path for the student model.

[0157] 1) Teacher Failure Count Update Rules

[0158] Each teacher Maintenance failure counter The update rules are as follows:

[0159] For the current winner (i.e., the teacher corresponding to the best model in this round) Reset its failure count to zero:

[0160] .

[0161] For the remaining teachers Increment the failure count by one:

[0162] .

[0163] 2) The logic behind teacher elimination in each round

[0164] If multiple teachers fail to advance in a round, to avoid incorrect eliminations due to short-term fluctuations, we use a "failure-first, score-order-assisted" mechanism to eliminate the teacher with the worst overall learning feedback. First, among all the teachers who did not advance, those currently... The teacher with the highest value. If the current... If the situation is the same, the elimination strategy will revert to "worst score in this round". The main process is as follows:

[0165] When only one teacher remains in the teacher pool, that is... In this case, the sole teacher will remain until the conditions for termination of training are met.

[0166] when When the value is greater than 1, the main process is as follows:

[0167] 1. Construct the set of teachers who did not win. :

[0168] ;

[0169] in, This is the group of teachers who did not win in round r.

[0170] 2. Select the subset of teachers with the most failures. :

[0171] ;

[0172] in, This is the set of teachers who have failed the most times in round r.

[0173] 3. Eliminate the player with the lowest score in the current round:

[0174] ;

[0175] in, This is the index of the teacher with the lowest score in round r.

[0176] 4. Finally, update the next round of teacher selection:

[0177] ;

[0178] in, Let be the set of teacher models for the (r+1)th round.

[0179] Step 5: Repeat steps 3 and 4 until the iteration stopping condition is met to obtain the summary generation model;

[0180] In practice, to avoid overfitting and invalid iterations, we set the following termination conditions to end training. Iterative training will stop when any of the following conditions are met:

[0181] 1. The improvement in validation set performance metrics is below the preset threshold for two consecutive rounds:

[0182] ;

[0183] in, The validation set score corresponding to the winning teacher in the previous round.

[0184] 2. The teacher set has only 1 member remaining and the improvement in the validation set performance metrics is below the preset threshold for two consecutive rounds:

[0185] ;

[0186] 3. No difficult samples

[0187] =0;

[0188] 4. Maximum number of training rounds limit

[0189] ;

[0190] in, This is the maximum number of training rounds.

[0191] When one of the above conditions is met, we consider the student model to have converged sufficiently or the teacher set to be unable to provide new and diverse guidance. Therefore, we terminate the training process and finally output the student model obtained in the last round as the summary generation model trained by this method.

[0192] Step 6: Input the target text into the summary generation model to obtain the target summary.

[0193] In practice, when it is necessary to process a certain text to generate a corresponding summary, the target text can be input into the summary generation model to obtain the corresponding target summary.

[0194] This embodiment presents a text summarization method based on multi-teacher feedback learning and dynamic elimination. It proposes a "feedback learning" mechanism that uses natural language feedback as a supervisory signal, bypassing the reliance on numerical reward modeling in traditional reinforcement learning. Structured prompts guide student models to perform responsive optimization of initial draft summaries. The method utilizes multiple large language models as "teachers" to replace expensive human experts, constructing a student-multi-teacher interactive feedback optimization framework. This generates fine-grained, complementary natural language feedback, selecting and retaining the most effective feedback paths to ensure student models receive guidance from multiple perspectives, significantly enhancing the quality and accuracy of summarization. A difficult sample mining mechanism based on the gap between the generated summaries and reference summaries is introduced. In each round of feedback learning, the model is prioritized for training on samples where "feedback is difficult to improve." By focusing on these "weak" samples, student models can specifically optimize their generation capabilities, significantly improving overall summarization quality. Finally, a dynamic teacher elimination mechanism based on validation set performance is introduced, employing a "failure-first, score-order-assisted" strategy to eliminate teachers with poor feedback performance round by round, achieving continuous optimization of the feedback source.

[0195] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof.

[0196] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A text summarization method based on multi-teacher feedback learning and dynamic elimination, characterized in that, include: Step 1: Obtain the sample dataset and divide it into a training set and a validation set, wherein the sample dataset includes sample text and its corresponding reference summary; Step 2: Generate a draft summary for the sample text in the training set and construct a Prompt template based on it. Use the Prompt template and the reference summary to fine-tune the model using supervised instructions to obtain the master student model. Step 3: Initialize the multi-teacher model set, input the training set into the master student model to obtain the difficult sample set, and train and evaluate the current master student model based on the multi-teacher model set and the difficult sample set to obtain the master student model for the next round of training. Step 3 specifically includes: Step 3.1: Construct a set of multi-teacher models with complementary styles; Step 3.2: Generate predicted summaries for the sample texts in the training set using the current round of the master student model, calculate the difference between the predicted summaries and the reference summaries, and use this difference to filter the set of difficult samples. ; Step 3.3: Each teacher model in the multi-teacher model set generates target feedback suggestions for the difficult samples in the current round; Step 3.4: Construct a feedback learning Prompt based on the sample text, predicted summary, target feedback suggestions, and instructions, and use the current master student model. Based on this, a set of student model branches, guided by different teacher models, are trained using feedback learning Prompt and a second loss function. Wherein, the second loss function is ； in, This represents the loss during the training of the student model branch in the r-th round. For student model parameters, For reference summary The sequence length, For time step index, For target summary The target token at time step t. For the target summary One token, Student model The conditional probability distribution, where i is the index of the training sample. Represents the teacher model in the r-th round. For the sample Constructed feedback learning prompt; Step 3.5: Evaluate each student branch model in the validation set, calculate the comprehensive score, and select the student branch model with the highest score as the master student model for the next round of training. Step 4: At the end of each training round, teacher models are eliminated according to the preset teacher elimination rules, and a new set of multi-teacher models is obtained to enter the next training round. Step 5: Repeat steps 3 and 4 until the iteration stopping condition is met to obtain the summary generation model; Step 6: Input the target text into the summary generation model to obtain the target summary.

2. The method according to claim 1, characterized in that, Step 2 specifically includes: Step 2.1: Use a pre-trained model to generate a draft summary for the sample text in the training set; Step 2.2: Construct a general feedback template for the first draft abstract to generate initial feedback suggestions; Step 2.3: Construct a Prompt template based on the sample text, draft abstract, initial feedback suggestions, and instructions; Step 2.4: Use the Prompt and reference summary to fine-tune the supervised instructions, calculate the first loss function, and initialize the student model accordingly. The initialized student model This serves as the master student model for the first round of multi-teacher feedback learning training.

3. The method according to claim 2, characterized in that, The first loss function is ； Where N is the total number of samples. This indicates the Prompt template.

4. The method according to claim 3, characterized in that, Step 3.2 specifically includes: Step 3.2.1: When the current multi-teacher feedback learning training round r=1, the difficult sample set is the sample text set, i.e. ; Step 3.2.2: When the current multi-teacher feedback learning training round r>1, the current round master-student model... Generated prediction summary Reference Abstract By comparison, the difference values ​​are obtained: ； in, , For hyperparameters, satisfying , This represents the difference between the current predicted abstract and the reference abstract. This is a function for calculating the longest common subsequence overlap between the reference digest and the generated digest. A function for calculating the semantic similarity between the reference summary and the generated summary; Step 3.2.3: Construct a set of difficult samples based on the difference values. : ； Among them, if Then this sample is considered a difficult sample in the current round. Let r be the set of difficult samples in the r-th round. is the threshold for screening difficult samples in the r-th round.

5. The method according to claim 4, characterized in that, Step 3.5 specifically includes: Step 3.5.1: Randomly select a performance evaluation dataset according to a preset ratio to evaluate the performance of each student model branch; Step 3.5.2, calculate each student model branch. In the validation set Overall score : ； ； ； ； Teacher model The average ROUGE-L score of the guided student model branch on the r-th round validation set is used to measure the average word-level similarity between the student model's generated summary and the reference summary. Teacher model The average value of the guided student model branch on the r-th round validation set A score is used to measure the semantic similarity between the student-generated summary and the reference summary. Teacher model The average summary length penalty term of the guided student model branch on the r-th round validation set is used to measure the deviation in length between the generated summary and the reference summary. For hyperparameters, This indicates that the i-th validation sample is generated by the teacher model. Summary generated by the guided student model branch This is a reference summary for the validation sample, where i is the local index of the current validation set sample. It is a function for calculating the length of the abstract. The total number of samples in the validation set; Step 3.5.3: Select the teacher index with the highest score on the validation set. : ； in, This indicates selecting the score from all teacher models. The largest teacher model, and numbered it as The teacher model that wins in this round of training; Step 3.5.4: Select the student model branch with the highest score as the master student model for the next round of training. ； in, For the next round of feedback learning, the master student model This is the student model branch corresponding to the teacher model that wins in this round of training.

6. The method according to claim 5, characterized in that, Step 4 specifically includes: Step 4.1: At the end of each training round, maintain a teacher failure counter. For the teacher model corresponding to the winner of the current round, clear its failure count to zero. For the remaining teacher models, increment the failure count by one. The winner is the teacher model corresponding to the student branch model with the highest score in this round. Step 4.2: Eliminate one teacher in each round according to the teacher elimination rules and update the multi-teacher model set.

7. The method according to claim 6, characterized in that, Step 4.2 specifically includes: Step 4.2.1: When the number of teacher models in the multi-teacher model set is equal to 1, continue to retain the unique teacher until the iteration stops. Step 4.2.2: When the number of teacher models in the multi-teacher model set is greater than 1, then construct the set of teachers who did not win. : ； in, This is the group of teachers who did not win in round r. For the r-th round of teacher set; Select the subset of teachers with the most failures : ； in, The set of teachers who failed the most times in round r. For each teacher model A failure counter that accumulates the number of rounds without a win; In the teacher subset The player with the lowest score in the current round will be eliminated. ； in, The index of the teacher with the lowest score in round r; Final update of the next round of teacher selection: ； in, For the (r+1)th round of teacher models, This is the teacher elimination model for the current round.

8. The method according to claim 7, characterized in that, The iteration stopping condition is any one of the following: the improvement of the validation set index is below the threshold for two consecutive rounds, there is only one teacher left and the validation index has not improved significantly, there are no difficult samples in the current round, or the maximum number of training rounds has been reached.

Images