Domain frontier dynamic progressive abstract system based on pre-training model

Abstract

The invention belongs to the field of natural language processing, and particularly relates to a field frontier dynamic progressive abstract system based on a pre-training model. Comprising an information acquisition module, a text abstract generation module and a personalized Agent interaction module, wherein the information acquisition module is used for crawling a field leading edge dynamic state, labeling and processing data, and constructing a data set; the text abstract generation module is used for processing a data set by using a prediction model to obtain candidate abstracts, adjusting according to an abstract evaluation algorithm to obtain an intermediate model of optimal parameters, and generating a preliminary abstract by using the intermediate model; the personalized Agent interaction module is used for providing the initial abstract output by the text abstract generation module for a user; and the abstract quality feedback of the user is collected, the feedback is classified and stored, and the abstract quality is fed back to the abstract evaluation algorithm of the text abstract generation module, so that the text abstract generation module continuously adjusts parameters and optimizes the intermediate model.

Description

Technical Field

[0001] The present invention belongs to the field of natural language processing, and in particular relates to a domain-leading dynamic progressive summarization system based on a pre-training model. Background Art

[0002] Traditional extractive summarization techniques often convert sentences in a corpus into sequences of semantic units, representing words and sentences by extracting abstract semantics and sequence information. This algorithm suffers from shortcomings such as limited generalization, lack of fluency, and redundant sentences. With the rapid development of artificial intelligence technology, particularly in the field of Natural Language Processing (NLP), pre-trained models have demonstrated strong versatility and transfer learning capabilities, greatly promoting innovation and progress in text processing tasks. With the rise of pre-trained models, generative summarization based on pre-trained models can achieve accurate and coherent summaries of the original content, effectively addressing the challenges of traditional methods in maintaining semantic integrity and generating diversity, and can be applied to more flexible scenarios.

[0003] Existing large-scale text summarization models often overlook the dynamic interactions and human feedback during the summary generation process. When processing long, multi-layered, and complex texts like academic papers, the generated summaries often fail to clearly demonstrate their progressive structure and logical relationships due to the lack of human thought process guidance.

[0004] Today, cutting-edge information across various disciplines is vast, complex, and rapidly updated. Existing information platforms struggle to meet users' needs for fast and accurate access to cutting-edge information in specific fields. Besides lacking immediacy and timeliness, existing approaches also struggle to adapt to the personalized information needs of different users, or even the same user over time. Summary of the Invention

[0005] In response to the problems existing in the existing technology, the present invention uses a pre-trained large model to perform dynamic progressive summarization of the frontiers of the field; it realizes personalized summary generation for different users, helping users to quickly obtain the frontier information and dynamics of the field of interest, thereby improving efficiency.

[0006] Technical solution of the present invention:

[0007] A domain-leading dynamic progressive summarization system based on a pre-trained model, including: an information acquisition module, a text summary generation module, and a personalized agent interaction module;

[0008] The information acquisition module is used to crawl the latest developments in the field, annotate and process the data, and construct a data set to provide to the text summary generation module;

[0009] The text summary generation module obtains a prediction model by designing an input feature vector, a network model, and a summary evaluation algorithm. The prediction model processes the data set obtained by the information acquisition module to obtain candidate summaries, and adjusts the intermediate model with optimal parameters according to the summary evaluation algorithm to generate a preliminary summary.

[0010] The personalized agent interaction module provides the preliminary summary output by the text summary generation module to the user; the user reads the summary, selects a retrospective summary, and submits feedback and evaluation to the personalized interaction module after use; the module collects user feedback on summary quality, classifies and saves the feedback, and feeds back the summary quality to the summary evaluation algorithm of the text summary generation module, so that the text summary generation module can continuously adjust parameters and optimize the intermediate model.

[0011] Beneficial effects

[0012] 1) This invention addresses the problems of high cost of obtaining cutting-edge dynamics in a field, scattered information, and lack of professionalism. By crawling top conferences and authoritative journals in a specific field, we can obtain cutting-edge dynamics in the field and ultimately obtain personalized summaries for different user profiles, providing a new solution to the problems of high cost of information acquisition, scattered information, and lack of professionalism.

[0013] 2) Aiming at the dynamic progressive summarization task at the forefront of the field, the present invention proposes to obtain a preliminary summary through candidate summaries by adjusting the loss function, then form a user profile through user feedback, adjust the summary evaluation criteria, and finally obtain a refined summary that meets the personalized summary requirements.

[0014] 3) Design a personalized summary generation agent that collects user feedback after reading summaries and automatically understands their needs for various aspects of summaries based on this feedback, gradually forming a user profile. The agent then uses the domain characteristics reflected in the user profile to adjust the output importance of papers from different fields in subsequent summaries. This personalized training and optimization of the model generates a personalized summary model tailored to the user's style, thereby meeting their needs and achieving personalized and refined summary generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Schematic diagram of the overall framework of the system of the present invention;

[0016] Figure 2 This is a schematic diagram of the information acquisition module structure according to an embodiment of the present invention;

[0017] Figure 3 This is a schematic diagram of the structure of a text summary generation module according to an embodiment of the present invention;

[0018] Figure 4This is a schematic diagram of the processing process of the personalized Agent interaction module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The following will describe the embodiments of the present invention in detail with reference to the accompanying drawings and examples, so that the implementation process of how the present invention applies technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly.

[0020] A dynamic progressive summarization system based on a pre-trained model, including: information acquisition module, text summary generation module and personalized agent interaction module, such as Figure 1 :

[0021] The information acquisition module is used to crawl the latest developments in the field, annotate and process the data, and construct a data set to provide to the text summary generation module;

[0022] The text summary generation module obtains a prediction model by designing an input feature vector, a network model, and a summary evaluation algorithm. The prediction model processes the data set obtained by the information acquisition module to obtain candidate summaries, and adjusts the intermediate model with optimal parameters according to the summary evaluation algorithm to generate a preliminary summary.

[0023] The personalized agent interaction module provides the preliminary summary output by the text summary generation module to the user; the user reads the summary, selects a retrospective summary, and submits feedback and evaluation to the personalized interaction module after use; this module collects user feedback on summary quality, classifies and saves the feedback, and feeds summary quality feedback to the summary evaluation algorithm of the text summary generation module, so that the text summary generation module can continuously adjust parameters and optimize the intermediate model;

[0024] The details are as follows:

[0025] like Figure 2 As shown, the information acquisition module includes a text reading and parsing submodule and a text cleaning submodule, which are as follows

[0026] Text reading and parsing submodule:

[0027] Read text data from the source website, identify the character encoding of the text and convert it to a unified encoding standard. For files in non-plain text formats (such as HTML, Markdown, XML), parse their content and extract usable plain text information.

[0028] Text cleaning submodule:

[0029] Complete text cleaning operations, including:

[0030] Check whether the data complies with predefined rules and constraints, such as data type, whether the field is empty, etc.

[0031] According to the language characteristics (English), word segmentation is performed based on spaces;

[0032] Remove irrelevant characters and non-text content, such as hyperlinks, footers, copyright information, XML tags and other non-text information;

[0033] Standardize the format and unify the formats of line breaks, indentation, spaces, etc.

[0034] Remove redundant characters, delete irrelevant control symbols, repeated spaces, etc.;

[0035] Perform case conversion and convert all English characters to lowercase to prevent them from affecting the model.

[0036] like Figure 3 As shown in the figure, the text summary generation module includes: a text vector construction submodule, a global attention decoder submodule, and an evaluation and optimization submodule, which are as follows:

[0037] Text vector construction submodule:

[0038] First, the input text needs to be tokenized. The model's corresponding tokenizer is used to tokenize the original text and summary, breaking the continuous text into words or subword units. Word embedding is then performed, mapping the tokenized results into a predefined vocabulary. The tokens are then converted into integer indices based on the model's vocabulary.

[0039] The word embeddings are then encoded in the encoder, and the corresponding text vectors are generated through the stacking and operation of multiple layers of encoder modules. The addition of a self-attention mechanism enables the model to capture the dependencies between any two positions in the input text. However, because it is essentially based on the content of the input vector and is independent of the position of each word in the sequence, the module also adds relative position encoding to the word embeddings to preserve sequence information.

[0040] Using the attention score as the weight, we perform a weighted summation on the value vectors of all positions to obtain the attention output of the current position i:

[0041]

[0042] Among them A ij is the attention score, V j is the value vector, and n is the length of the sequence, that is, the number of words.

[0043] Based on the resulting positional encodings and attention weights, residual connections are used to optimize convergence and improve model training stability. A feedforward layer is added to the module to perform nonlinear feature extraction, enhancing the model's expressiveness and adaptability to input data. The resulting text vector is then fed into the global attention decoder submodule for further processing.

[0044] Global Attention Decoder Submodule:

[0045] The obtained text vector is input into the decoder, and the entire output sequence is selected as the context vector sequence to ensure that each part of the original text is carefully encoded to achieve better summary generation.

[0046] At each time step, the self-attention layer is still used to complete the contextual understanding of the generated part, and a causal mask is added to ensure that each position only interacts with all previous positions. The cross-attention layer is set to connect the contextual information of the entire input sequence and combine it with the hidden state of the current time step to obtain the cross-attention context vector.

[0047] Residual connections and feedforward layers are still used to stabilize the training state and accelerate convergence. The final hidden information is used to predict vocabulary and obtain the probability distribution of the corresponding vocabulary at the current time step.

[0048] Adopt the greedy decoder strategy and select the word with the highest probability as the next word each time:

[0049] y t =argmaxP(ω|Y t ,X),ω∈V

[0050] where y t represents the word output at time step t, ω is a word in the vocabulary V, and Y t represents the word sequence generated before time step t, and X is the input data.

[0051] Finally, the decoding is completed and the text summary (i.e., candidate summary) is obtained.

[0052] Evaluation and optimization submodule:

[0053] The above two modules constitute the basic architecture of the model. When training the model, the training data set consists of several paragraphs of text and their corresponding reference summaries X (i.e., standard summaries). During training, the text is generated into a candidate summary set {y1, y2, y3, ..., y t}, in order to make the summary generated by the model more accurate, it is necessary to calculate the loss function and then update the parameters to improve the model accuracy. The learning process of the model is the process of minimizing the loss function.

[0054] The loss function is defined as follows:

[0055] For the candidate summary set {y1,y2,y3,...,y t}, the automatic summary evaluation algorithm ROUGE is used, which mainly calculates the candidate summary y to be evaluated i The quality of computer-generated text summaries is measured by the degree of n-gram overlap with the standard summary X. Common evaluation metrics include ROUGE-N and ROUGE-L, which are calculated using the following formula:

[0056] The denominator is the number of n-grams in the standard summary, and the numerator is the number of n-grams shared by the standard summary and the candidate summary.

[0057]

[0058] Where β is a parameter, where LCS(X,y i ) are X and y i The length of the longest common subsequence of the two summaries, taking into account the order. m and n represent the lengths of the reference summary and the generated summary, respectively (the length refers to the number of words in the summary).

[0059] Based on ROUGE, we further set the evaluation function B(y i )for:

[0060] B(y i )=0.5*ROUGE-N(X,y i )+0.5*ROUGE-L(X,y i ),

[0061] And set the scoring function S(y i ):

[0062] S(y i )=log Model(θ) (y i |X,y 1:i-1 ,θ),θ is the model parameter,

[0063] Combined with the objective evaluation index B(y i ) and human subjective evaluation index S(y i ), and obtain the contrastive learning loss function L com :

[0064]

[0065] Among them, α is a hyperparameter

[0066] The contrastive learning loss function L com Combined with the cross entropy loss function L(ω) we get the final loss function L:

[0067] L=λL com +L(ω),λ are hyperparameters.

[0068] Among them, the calculation formula of the cross entropy loss function L(ω) is as follows:

[0069] ω i is the model's predicted probability for the i-th category.

[0070] After the loss function is determined, the accuracy of the model can be improved by continuously iterating the training model. After multiple iterations, the training is stopped to obtain the final model, which is used to generate the final summary.

[0071] like Figure 4 As shown in the figure, the personalized summary agent module collects multi-dimensional structured feedback data through multiple rounds of interaction with users, combines user behavior analysis with deep learning technology, dynamically generates user preference weights, and adjusts model parameters based on the weights to achieve personalized summary generation.

[0072] After completing model fine-tuning, the model is capable of generating summaries, but it does not yet have the ability to interact with humans. This module will further establish a human-computer interaction and visualization platform based on this model, integrate a human feedback learning mechanism, and realize multiple rounds of interaction and iterative improvement between the large summary model and users.

[0073] This module forms a closed-loop optimization system with the information acquisition module and the text summary generation module. The specific technical implementation is as follows:

[0074] Multi-dimensional user feedback collection and processing:

[0075] First, collect a variety of user interaction data, including:

[0076] 1. Behavioral data: This records the user's click heat map on the summary (such as the duration of the highlighted part), the frequency and location of backtracking to the original text, and the expansion / collapse operations of summary paragraphs, to extract the distribution characteristics of the core content that users are interested in.

[0077] 2. Explicit evaluation data: Obtain users' direct evaluation of the quality of the summary through sliding ratings (1-5 points), multi-dimensional label selection (such as "logical clarity", "professional depth", "linguistic conciseness"), and text feedback (natural language evaluation).

[0078] 3. Implicit preference data: Analyze users’ historical browsing history and abstract collection / sharing behaviors to construct users’ implicit interest vectors (such as domain keyword weights and document type preferences).

[0079] The specific data processing method for user interaction data is as follows:

[0080] A BERT-based sentiment analysis model is used to extract sentiment polarity (positive / negative) and fine-grained requirements (such as "increase experimental data" and "shorten summary length") from text feedback data, and then mapped them into structured labels.

[0081] Use time series modeling (LSTM) to capture user interaction patterns for behavioral data, and combine the attention mechanism to calculate the importance scores of different summary fragments to form the user attention matrix M attention ∈R n*d (n is the number of summary sentences, d is the feature dimension).

[0082] Dynamic weight generation and model optimization:

[0083] The following mechanism is used to generate weights:

[0084] 1. Preference feature fusion: user behavior feature matrix M attention , explicit rating vector S score , implicit interest vector I interest Spliced ​​into a unified feature vector F user ∈R k , input the multi-layer perceptron (MLP) to generate user preference weight W preference ∈R m , where m corresponds to the adjustable parameter dimension (attention head weight, loss function coefficient) in the text summary generation module.

[0085] 2. Weight adaptive allocation: Through contrast learning strategy, W preference Align with the original parameter distribution of the pre-trained model to ensure that the personalized model remains domain-universal.

[0086] Among them, the following model parameter adjustment method is adopted:

[0087] Introduce user preference terms into the loss function of the text summary generation module:

[0088] in are model parameters, is the baseline model parameter, and λ3 controls the intensity of personalized adjustment.

[0089] Adjust the decoder's attention mechanism:

[0090] W preferenceAs gating weights, dynamically scale the outputs of different attention heads:

[0091] Where h is the number of attention heads, which can enhance the priority of user-focused content during the generation process.

[0092] Closed-loop optimization and progressive iteration:

[0093] Relationship with previous modules: The information acquisition module provides cutting-edge data in the field to support summary generation. The text summary generation module dynamically updates model parameters based on user feedback. The personalized agent iteratively optimizes user portraits and weight generation strategies through feedback data, forming a "data-generation-feedback-optimization" closed loop.

[0094] Long-term personalized training: Build an independent fine-tuning dataset D for each user user , containing its historical interaction data and preference weights, and regularly updating the personalized model branches through incremental learning to avoid global model drift.

[0095] Innovation

[0096] Innovation 1: Progressive method for generating dynamic summaries of cutting-edge fields

[0097] The present invention uses both objective and subjective summary evaluation metrics to assess the candidate summaries initially generated by a pre-trained model. This candidate summary evaluation data is then fed back to the model, causing it to assign a higher generation probability to candidate summaries with higher evaluation metric scores, thereby adjusting the model for higher accuracy. Furthermore, the present invention can feed back user evaluation data to the model, allowing the model to continuously improve through user interaction, enabling it to generate multi-layered, progressive summaries that better meet user needs.

[0098] Innovation 2: Personalized Summarization Agent Design Method

[0099] A personalized summary generation agent is designed to collect user feedback after reading summaries. Based on this feedback, it automatically understands the user's needs for various aspects of summaries, gradually forming a user profile. The agent then uses the domain characteristics reflected in the user profile to adjust the output importance of papers in different fields in subsequent summaries. This personalized training and optimization of the model generates a personalized summary model tailored to the user's style, meeting their needs and enabling personalized, refined summary generation.

[0100] The above description is only a description of the preferred embodiments of the present application and does not limit the scope of the present application. Any changes or modifications made by any person skilled in the art based on the above disclosed technical content should be regarded as equivalent valid embodiments and fall within the scope of protection of the technical solution of the present application.

Claims

1. A domain-leading dynamic progressive summarization system based on a pre-trained model, characterized by: include: Information acquisition module, text summary generation module and personalized agent interaction module; The information acquisition module is used to crawl the latest developments in the field, annotate and process the data, and construct a data set to provide to the text summary generation module; The text summary generation module obtains a prediction model by designing an input feature vector, a network model, and a summary evaluation algorithm. The prediction model processes the data set obtained by the information acquisition module to obtain candidate summaries, and adjusts the intermediate model with optimal parameters according to the summary evaluation algorithm to generate a preliminary summary. The personalized agent interaction module provides the preliminary summary output by the text summary generation module to the user; the user reads the summary, selects a retrospective summary, and submits feedback and evaluation to the personalized interaction module after use; the module collects user feedback on summary quality, classifies and saves the feedback, and feeds back the summary quality to the summary evaluation algorithm of the text summary generation module, so that the text summary generation module can continuously adjust parameters and optimize the intermediate model.

2. A domain-frontier dynamic progressive summarization system based on a pre-trained model as claimed in claim 1, characterized in that: The information acquisition module includes a text reading and parsing submodule and a text cleaning submodule; Text reading and parsing submodule: reads text data from the source website, identifies the character encoding of the text and converts it to a unified encoding standard; for files in non-plain text format, parses their content and extracts usable plain text information; The text cleaning submodule completes the text cleaning operation, including: Checking data for compliance with predefined rules and constraints; According to the language characteristics, word segmentation is performed based on spaces; Remove irrelevant characters and non-text content; Standardize the format and unify the formats of line breaks, indentation, spaces, etc. Remove redundant characters, delete irrelevant control symbols, and repeated spaces; Perform case conversion and convert all English characters to lowercase to prevent them from affecting the model.

3. A domain frontier dynamic summarization system based on a pre-trained model as claimed in claim 1, characterized in that: The text summary generation module includes: a text vector construction submodule, a global attention decoder submodule and an evaluation and optimization submodule; The text vector construction submodule first needs to perform word segmentation on the input text, using the model's corresponding word segmenter Tokenizer to segment the original text and summary, and split the continuous text into words or subword units; then perform word embedding, map the word segmentation results to a predefined vocabulary, and convert the word segmentation into integer indexes according to the model vocabulary; The encoder then encodes the word embeddings. Through stacking and calculation of multiple layers of encoder modules, the corresponding text vectors are generated. A self-attention mechanism is added to enable the model to capture the dependency between any two positions in the input text. The module also adds relative position encoding to the word embeddings to preserve sequence information. Using the attention score as the weight, we perform a weighted summation on the value vectors of all positions to obtain the attention output of the current position i: Among them A ij is the attention score, V j is the value vector, n is the length of the sequence, that is, the number of words; Based on the obtained positional encoding and attention weights, residual connections are used to optimize convergence and improve the training stability of the model. A feedforward layer is added to the module to perform nonlinear feature extraction, enhancing the model's expressiveness and adaptability to input data. Finally, the text vector is obtained and provided to the global attention decoder submodule for the next step. The global attention decoder submodule inputs the obtained text vector into the decoder and selects to use the entire output sequence as the context vector sequence, thereby ensuring that each part of the original text is carefully encoded to achieve better summary generation; For each time step, the self-attention layer is still used to complete the contextual understanding of the generated part, and a causal mask is added to ensure that each position only interacts with all previous positions. A cross-attention layer is set to connect the contextual information of the entire input sequence and combine it with the hidden state of the current time step to obtain the cross-attention context vector. Residual connections and feedforward layers are still used to stabilize the training state and accelerate convergence. The final hidden information is predicted to obtain the probability distribution of the corresponding word at the current time step. Adopt the greedy decoder strategy and select the word with the highest probability as the next word each time: y t =argmaxP(ω|Y t ,X),ω∈V where y t represents the word output at time step t, ω is a word in the vocabulary V, and Y t represents the word sequence generated before time step t, and X is the input data; Finally, the decoding is completed and the text summary, that is, the candidate summary, is obtained.

4. A domain frontier dynamic summarization system based on a pre-trained model as claimed in claim 3, characterized in that: The evaluation and optimization submodule processes as follows: When training the model, the training dataset consists of several paragraphs of text and their corresponding reference summaries X. During training, the text is processed by the text vector construction submodule and the global attention decoder submodule to generate a candidate summary set {y1, y2, y3, ..., y t }, calculate the loss function and then update the parameters to improve the model accuracy; The loss function is defined as follows: For the candidate summary set {y1,y2,y3,...,y t }, the automatic summary evaluation algorithm ROUGE is used as follows: The denominator is the number of n-grams in the standard summary, and the numerator is the number of n-grams shared by the standard summary and the candidate summary. Where β is a parameter, where LCS(X,y i ) are X and y i The length of the longest common subsequence of the two summaries, taking the order into consideration; m and n represent the lengths of the reference summary and the generated summary respectively, and the length refers to the number of words contained in the summary; Based on ROUGE, we further set the evaluation function B(y i )for: B(y i )=0.5*RED-N(X,y i )+0.5*RED-L(X,y i ), And set the scoring function S(y i ): S(y i )=log Model(θ) (y i |X,y 1:i-1 ,θ),θ is the model parameter, Combined with the objective evaluation index B(y i ) and human subjective evaluation index S(y i ), and obtain the contrastive learning loss function L com : Among them, α is a hyperparameter The contrastive learning loss function L com Combined with the cross entropy loss function L(ω) we get the final loss function L: L=λL com +L(ω),λ are hyperparameters; Among them, the calculation formula of the cross entropy loss function L(ω) is as follows: ω i is the model’s predicted probability for the i-th category; After the loss function is determined, the accuracy of the model can be improved by continuously iterating the training model. After multiple iterations, the training is stopped to obtain the final model, which is used to generate the final summary.

5. A domain frontier dynamic summarization system based on a pre-trained model as claimed in claim 1, characterized in that: The personalized agent interaction module has the following functions: The personalized summary agent module collects multi-dimensional structured feedback data through multiple rounds of interaction with users, combines user behavior analysis with deep learning technology, dynamically generates user preference weights, and adjusts model parameters based on the weights to achieve personalized summary generation; This module forms a closed-loop optimization system with the information acquisition module and the text summary generation module. The specific technical implementation is as follows: Multi-dimensional user feedback collection and processing: First, collect and process various user interaction data; Dynamic weight generation and model optimization, using the following mechanism to generate weights: 1) Preference feature fusion: user behavior feature matrix M attention , explicit rating vector S score , implicit interest vector I interest Spliced ​​into a unified feature vector F user ∈R k , input the multi-layer perceptron MLP to generate user preference weight W preference ∈R m , where m corresponds to the adjustable parameter dimension in the text summary generation module; 2) Weight adaptive allocation: Through contrast learning strategy, W preference Align with the original parameter distribution of the pre-trained model to ensure that the personalized model remains domain-universal; Closed-loop optimization and progressive iteration are as follows: 1) Relationship with previous modules: The information acquisition module provides cutting-edge data to support summary generation. The text summary generation module dynamically updates model parameters based on user feedback. The personalized agent iteratively optimizes user profiles and weight generation strategies based on feedback data, forming a closed loop of "data-generation-feedback-optimization"; 2) Long-term personalized training: Build an independent fine-tuning dataset D for each user user , containing its historical interaction data and preference weights, and regularly updating the personalized model branches through incremental learning to avoid global model drift.

6. A domain frontier dynamic summarization system based on a pre-trained model as claimed in claim 5, characterized in that: The multi-dimensional user feedback collection and processing includes: First, collect a variety of user interaction data, including: 1) Behavioral data: Record user click heatmaps on summaries, the frequency and location of backtracking to the original text, and summary paragraph expansion / collapse operations to extract the distribution characteristics of core content that users are interested in; 2) Explicit evaluation data: Obtain users' direct evaluation of summary quality through sliding rating, multi-dimensional label selection, and text feedback; 3) Implicit preference data: Analyze user browsing history and summary collection / sharing behavior to construct user implicit interest vectors; The specific data processing method for user interaction data is as follows: A BERT-based sentiment analysis model is used to extract sentiment polarity and fine-grained requirements from text feedback data, and then mapped them into structured labels. Use time series modeling LSTM to capture user interaction patterns for behavioral data, and combine the attention mechanism to calculate the importance scores of different summary fragments to form the user attention matrix M attention ∈R n*d , n is the number of summary sentences, and d is the feature dimension.

7. A domain frontier dynamic summarization system based on a pre-trained model as claimed in claim 5, characterized in that: The dynamic weight generation and model optimization are specifically as follows: Introduce user preference terms into the loss function of the text summary generation module: L final =λ1L com +λ2L(ω)+λ3||W preference ·θ-θ base ||2, Where θ is the model parameter, θ base is the baseline model parameter, λ3 controls the intensity of personalized adjustment; Adjust the decoder's attention mechanism: W preference As gating weights, dynamically scale the outputs of different attention heads: Where h is the number of attention heads, which can enhance the priority of user-focused content during the generation process.

Images