A pre-trained model-based domain frontiers dynamic progressive summarization system
By designing a dynamic progressive summarization system for cutting-edge fields based on a pre-trained model, the problems of insufficient personalization and logical progression in existing technologies are solved, realizing personalized and dynamic summarization generation and meeting users' needs for quickly obtaining cutting-edge information in the field.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TONGJI UNIV
- Filing Date
- 2025-05-13
- Publication Date
- 2026-07-07
AI Technical Summary
Existing technologies struggle to provide personalized, dynamic summary generation for rapidly changing cutting-edge information in various fields, and lack adaptability to user needs and logical progression.
Design a frontier dynamic progressive summarization system based on a pre-trained model, including an information acquisition module, a text summarization generation module, and a personalized agent interaction module. By crawling frontier dynamic data, designing feature vectors and network models, and combining user feedback to optimize summary generation, a closed-loop optimization system is formed.
It enables personalized and dynamic summary generation, meeting the information needs of different users at different times, improving the efficiency and accuracy of information acquisition, and generating detailed summaries that meet user needs.
Smart Images
Figure CN120523938B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of natural language processing, specifically relating to a frontier dynamic progressive summarization system based on a pre-trained model. Background Technology
[0002] Traditional extractive summarization techniques typically transform sentences in a corpus into sequences of semantic units, representing words and sentences by extracting abstract semantics and sequence information. This algorithm suffers from limitations in generalization, fluency, and sentence redundancy. With the rapid development of artificial intelligence, particularly in Natural Language Processing (NLP), pre-trained models have demonstrated powerful versatility and transfer learning capabilities, significantly driving innovation and progress in text processing tasks. The rise of pre-trained models allows for accurate and coherent summarization of the original text, effectively addressing the challenges of maintaining semantic integrity and generating diversity in traditional methods, and enabling applications in more flexible scenarios.
[0003] Existing large-scale text summarization models often overlook the dynamic interaction and human feedback during the summarization process. When dealing with long, multi-layered, and complex texts such as academic papers, the lack of guidance from human thought processes often results in summaries that fail to clearly demonstrate the progressive structure and logical relationships.
[0004] The sheer volume, complexity, and rapid pace of cutting-edge information across various disciplines today make it difficult for existing information platforms to meet users' needs for quick and accurate access to specific fields of cutting-edge information. Besides insufficient immediacy and timeliness, existing methods also struggle to adapt to the changing needs of different users or the same user at different times, hindering personalized information access. Summary of the Invention
[0005] To address the problems of existing technologies, this invention utilizes a pre-trained large model to perform dynamic, progressive summarization of cutting-edge information in a field; it also enables personalized summary generation for different users, helping them quickly obtain cutting-edge information and dynamics in fields of interest, thereby improving efficiency.
[0006] Technical solution of the present invention:
[0007] A frontier dynamic progressive summarization system based on a pre-trained model includes: an information acquisition module, a text summarization generation module, and a personalized agent interaction module;
[0008] The information acquisition module is used to crawl cutting-edge developments in the field, annotate and process the data, construct a dataset, and provide it to the text summarization generation module.
[0009] The text summarization generation module obtains a prediction model by designing input feature vectors, designing a network model, and designing a summary evaluation algorithm. The prediction model processes the dataset obtained by the information acquisition module to obtain candidate summaries, and adjusts the intermediate model with optimal parameters according to the summary evaluation algorithm. The intermediate model then generates a preliminary summary.
[0010] The personalized agent interaction module provides the user with the preliminary summary output by the text summarization generation module; the user reads the summary, selects a backtracking summary, and submits feedback and evaluation to the personalized interaction module after use; this module collects the user's summary quality feedback, classifies and saves the feedback, and feeds the summary quality feedback to the summary evaluation algorithm of the text summarization generation module, so that the text summarization generation module can continuously adjust parameters and optimize intermediate models.
[0011] Beneficial effects
[0012] 1) This invention addresses the problems of high cost, scattered information, and lack of professionalism in acquiring cutting-edge information in a specific field. By crawling top conferences and authoritative journals in a specific field, it obtains cutting-edge information and ultimately generates personalized summaries for different user profiles, providing a new solution to the problems of high information acquisition cost, scattered information, and lack of professionalism.
[0013] 2) This invention addresses the dynamic and progressive summarization task at the forefront of the field. It proposes to obtain a preliminary summary by adjusting the loss function through candidate summaries, then form a user profile through user feedback, adjust the summary evaluation criteria, and finally obtain a refined summary that conforms to personalized summarization.
[0014] 3) The designed personalized summary generation agent can collect user feedback after reading the summary and automatically understand their needs for various aspects of the summary based on the feedback, gradually forming a user profile. The agent will combine the domain characteristics reflected in the user profile to adjust the output importance of papers in different domains in subsequent summaries, perform personalized training and optimization of the model, and generate a personalized summary model with user style, thereby meeting their corresponding needs and realizing personalized and refined summary generation. Attached Figure Description
[0015] Figure 1 This is a schematic diagram of the overall system framework of the present invention;
[0016] Figure 2 This is a schematic diagram of the information acquisition module structure in an embodiment of the present invention;
[0017] Figure 3 This is a schematic diagram of the text summarization generation module structure according to an embodiment of the present invention;
[0018] Figure 4This is a schematic diagram of the personalized Agent interaction module processing procedure in an embodiment of the present invention. Detailed Implementation
[0019] The following will describe in detail the implementation of the present invention with reference to the accompanying drawings and embodiments, so as to fully understand how the present invention uses technical means to solve technical problems and achieve technical effects and to implement it accordingly.
[0020] A cutting-edge dynamic progressive summarization system based on a pre-trained model includes: an information acquisition module, a text summarization generation module, and a personalized agent interaction module, such as... Figure 1 :
[0021] The information acquisition module is used to crawl cutting-edge developments in the field, annotate and process the data, and construct a dataset to provide to the text summarization generation module;
[0022] The text summarization generation module obtains a prediction model by designing input feature vectors, designing a network model, and designing a summary evaluation algorithm. The prediction model processes the dataset obtained by the information acquisition module to obtain candidate summaries, and adjusts the intermediate model with optimal parameters according to the summary evaluation algorithm. The intermediate model then generates a preliminary summary.
[0023] The personalized agent interaction module provides the user with the preliminary summary output by the text summarization generation module; the user reads the summary, selects a backtracking summary, and submits feedback and evaluation to the personalized interaction module after use; this module collects the user's summary quality feedback, classifies and saves the feedback, and feeds the summary quality feedback to the summary evaluation algorithm of the text summarization generation module, so that the text summarization generation module can continuously adjust parameters and optimize intermediate models;
[0024] Specifically as follows:
[0025] like Figure 2 As shown, the information acquisition module includes a text reading and parsing submodule and a text cleaning submodule, as detailed below.
[0026] Text reading and parsing submodule:
[0027] The system reads text data from the source website, identifies the character encoding of the text, and converts it to a unified encoding standard. For non-plain text formats (such as HTML, Markdown, and XML), it parses the content and extracts the usable plain text information.
[0028] Text cleaning submodule:
[0029] Complete the text cleaning operation, including:
[0030] Check whether the data conforms to predefined rules and constraints, such as data type and whether fields are empty;
[0031] Based on language characteristics (English), words are segmented using spaces;
[0032] Remove irrelevant characters and non-text content, such as hyperlinks, footers, copyright information, XML tags, and other non-body information;
[0033] Standardize the formatting, unifying the formatting of line breaks, indentation, spaces, etc.
[0034] Remove redundant characters, and delete irrelevant control symbols, duplicate spaces, etc.
[0035] Perform case conversion to convert all English characters to lowercase to prevent them from affecting the model.
[0036] like Figure 3 As shown, the text summarization generation module includes: a text vector construction submodule, a global attention decoder submodule, and an evaluation and optimization submodule, as detailed below:
[0037] Text vector construction submodule:
[0038] First, the input text needs to be segmented. The tokenizer corresponding to the model is used to segment the original text and the summary, breaking down continuous text into words or sub-words. Then, word embedding is performed, mapping the segmentation results to a predefined vocabulary, and converting the segmented words into integer indices based on the model's vocabulary.
[0039] The word embeddings are then encoded by the encoder. Through the stacking and computation of multiple encoder modules, corresponding text vectors are generated. By adding a self-attention mechanism, the model can capture the dependencies between any two positions in the input text. However, since it is essentially based on the content of the input vector and is independent of the position of each word in the sequence, relative position encoding is also added to the word embeddings in the module to preserve sequence information.
[0040] Using the attention scores as weights, the value vectors at all positions are summed in a weighted manner to obtain the attention output at the current position i:
[0041]
[0042] Where A ij V represents the attention score. j Let n be a value vector, where n is the length of the sequence, i.e., the number of words.
[0043] Based on the obtained positional encoding and attention weights, convergence is optimized through residual connections to improve the model's training stability. A feedforward layer is added to the module to perform non-linear feature extraction, enhancing the model's expressive power and adaptability to input data. Finally, the resulting text vector is provided to 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 meticulously encoded in order to achieve better summary generation.
[0046] For each time step, a self-attention layer is still used to understand the context of the generated parts, and a causal mask is added to ensure that each position only interacts with all positions preceding it. A cross-attention layer is set up to interface with the context information of the entire input sequence, and combined with the hidden state of the current time step, a cross-attention context vector is obtained.
[0047] Residual connections and feedforward layers are still used to stabilize the training state and accelerate convergence. Word prediction is then performed on the final hidden information to obtain the probability distribution of the corresponding word at the current time step.
[0048] A greedy decoder strategy is adopted, selecting 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 This represents the word output at time step t, where ω is a word in the vocabulary V, and Y... t X represents the word sequence generated before time step t, and X is the input data.
[0051] Finally, decoding is completed, and a text summary (i.e., candidate summary) is obtained.
[0052] Evaluation and Optimization Submodule:
[0053] The two modules described above constitute the basic architecture of the model. During model training, the training dataset consists of several text segments and their corresponding reference summaries X (i.e., standard summaries). During training, these text segments are processed by the two modules to generate a candidate summary set {y1, y2, y3, ..., y...}. tTo 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's accuracy. The model's learning process 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 summaries y to be evaluated. i The quality of computer-generated text summarization is measured by its overlap with the standard summary X in n-grams. Commonly used 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 abstract, and the numerator is the number of n-grams shared by the standard abstract and the candidate abstract.
[0057]
[0058] Where β is a parameter, and LCS(X,y) i ) are X and y i The length of the longest common subsequence, considering order. m and n represent the lengths of the reference summary and the generated summary, respectively (length refers to the number of words in the summary).
[0059] Based on ROUGE, an evaluation function B(y) for candidate abstracts is further set. 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), where θ is the model parameter,
[0063] Combining objective evaluation indicators B(y) i ) and subjective human evaluation index S(y) i ), thus obtaining the contrastive learning loss function L com :
[0064]
[0065] Where α is a hyperparameter
[0066] The contrastive learning loss function L com Combining this with the cross-entropy loss function L(ω), we obtain the final loss function L:
[0067] L=λL com +L(ω),λ are hyperparameters.
[0068] The cross-entropy loss function L(ω) is calculated using the following formula:
[0069] ω i It is the model's predicted probability for the i-th category.
[0070] Once the loss function is determined, the accuracy of the model can be improved by iteratively training the model. After multiple iterations, training is stopped to obtain the final model, which is then used to generate the final summary.
[0071] like Figure 4 As shown, the Personalized Summary Agent module collects multi-dimensional structured feedback data through multiple rounds of interaction with users, combines user behavior analysis and deep learning technology to dynamically generate user preference weights, and adjusts model parameters based on the weights to achieve personalized summary generation.
[0072] After fine-tuning the model, it has reached the level of generating summaries, but it does not yet have human-computer interaction capabilities. This module will further build a human-computer interaction and visualization platform based on the model, integrate human feedback learning mechanisms, and realize multi-round interaction and iterative improvement between the large summarization model and users.
[0073] This module, together with the information acquisition module and the text summarization module, forms a closed-loop optimization system. The specific technical implementation is as follows:
[0074] Multi-dimensional user feedback collection and processing:
[0075] First, various user interaction data are collected, including:
[0076] 1. Behavioral data: Record user click heatmaps of the summary (such as the duration of time spent on highlighted parts), frequency and location of backtracking to the original text, and summary paragraph expansion / collapse operations 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 abstract quality through methods such as sliding scores (1-5 points), multi-dimensional tag selection (such as "logical clarity", "professional depth", "language conciseness"), and text feedback (natural language evaluation).
[0078] 3. Implicit Preference Data: Analyze users' browsing history and abstract collection / sharing behavior to construct implicit interest vectors for users (such as domain keyword weights and document type preferences).
[0079] The specific data processing methods for user interaction data are as follows:
[0080] For text-based feedback data, a BERT-based sentiment analysis model is used to extract sentiment polarity (positive / negative) and fine-grained requirements (such as "add experimental data" and "shorten the summary length"), and these are mapped to structured tags.
[0081] Temporal modeling (LSTM) is used to capture user interaction patterns for behavioral data, and an attention mechanism is combined to calculate the importance scores of different summary segments, forming a 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 weights are generated using the following mechanism:
[0084] 1. Preference Feature Fusion: This involves fusing the user behavior feature matrix M... attention Explicit rating vector S score Implicit interest vector I interest Concatenate into a unified feature vector F user ∈R k The input multilayer perceptron (MLP) generates user preference weights W. preference ∈R m , where m corresponds to the adjustable parameter dimension (attention head weight, loss function coefficient) in the text summarization generation module.
[0085] 2. Adaptive Weight Allocation: Through a contrastive learning strategy, W is weighted... preference Align the model with the original parameter distribution of the pre-trained model to ensure that the personalized model still maintains domain generality.
[0086] The following model parameter adjustment method is adopted:
[0087] Introduce a user preference term into the loss function of the text summarization generation module:
[0088] in For model parameters, λ3 controls the intensity of personalized adjustments, serving as the baseline model parameter.
[0089] Adjusting the decoder's attention mechanism:
[0090] W preferenceAs a gating weight, the output of different attention heads is dynamically scaled:
[0091] Where h represents the number of attention heads, which strengthens the priority of user-focused content during the generation process.
[0092] Closed-loop optimization and progressive iteration:
[0093] Connections with preceding 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, and the personalized agent iteratively optimizes user profiles and weight generation strategies through feedback data, forming a closed loop of "data-generation-feedback-optimization".
[0094] Long-term personalized training: Building an independent fine-tuning dataset D for each user user It includes historical interaction data and preference weights, and regularly updates personalized model branches through incremental learning to avoid global model drift.
[0095] Innovation
[0096] Innovation Point 1: A Progressive Method for Generating Dynamic Abstracts of Frontier Fields
[0097] This invention sets up both objective and subjective summary evaluation metrics to assess the candidate summaries initially generated by the pre-trained model. The resulting evaluation data is then fed back to the model, allowing it to assign higher generation probabilities to candidate summaries with higher evaluation scores, thereby adjusting the model to achieve higher accuracy. Simultaneously, this invention can feed user feedback data back to the model, enabling continuous improvement through user interaction and the generation of multi-layered, progressive summaries that better meet user needs.
[0098] Innovation Point Two: Personalized Summary Agent Design Method
[0099] Designing a personalized summary generation agent can collect user feedback after reading summaries and automatically understand their needs for various aspects of the summaries based on this feedback, gradually forming a user profile. The agent will then combine the domain characteristics reflected in the user profile to adjust the output importance of papers from different domains in subsequent summaries, performing personalized training and optimization of the model to generate a personalized summary model with the user's style, thereby meeting their corresponding needs and achieving personalized, refined summary generation.
[0100] The above description is merely a description of preferred embodiments of this application and is not intended to limit the scope of this application in any way. Any changes or modifications made by those skilled in the art based on the above-disclosed technical content should be considered as equivalent and valid embodiments and fall within the scope of protection of the technical solution of this application.
Claims
1. A frontier dynamic progressive summarization system based on a pre-trained model, characterized in that, include: The module includes an information acquisition module, a text summary generation module, and a personalized agent interaction module. The information acquisition module is used to crawl cutting-edge developments in the field, annotate and process the data, construct a dataset, and provide it to the text summarization generation module. The text summarization generation module obtains a prediction model by designing input feature vectors, designing a network model, and designing a summary evaluation algorithm. The prediction model processes the dataset obtained by the information acquisition module to obtain candidate summaries, and adjusts the intermediate model with optimal parameters according to the summary evaluation algorithm. The intermediate model then generates a preliminary summary. The personalized agent interaction module provides the user with the preliminary summary output by the text summarization generation module; the user reads the summary, selects a backtracking summary, and submits feedback and evaluation to the personalized interaction module after use; this module collects the user's summary quality feedback, classifies and saves the feedback, and feeds the summary quality feedback to the summary evaluation algorithm of the text summarization generation module, so that the text summarization generation module can continuously adjust parameters and optimize intermediate models; The text summarization generation module includes: a text vector construction submodule, a global attention decoder submodule, and an evaluation and optimization submodule; The evaluation and optimization submodule processes the following: During model training, the training dataset consists of several text segments and their corresponding reference summaries X. During training, each text segment is used to generate a candidate summary set through two modules: a text vector construction submodule and a global attention decoder submodule. The loss function is calculated, and the parameters are then updated to improve the model accuracy. The loss function is defined as follows: For candidate summary set The ROUGE automatic summary evaluation algorithm is used, as follows: ， Wherein, the denominator is the number of n-grams in the standard abstract, and the numerator is the number of n-grams common to both the standard abstract and the candidate abstract; , Where β is a parameter, yes and The length of the longest common subsequence, considering the order; and These represent the lengths of the reference abstract and the generated abstract, respectively. The length refers to the number of words contained in the abstract. Based on ROUGE, an evaluation function for candidate abstracts is further set. for: , And set the scoring function : , For model parameters, Combined with objective evaluation indicators and subjective human evaluation indicators The contrastive learning loss function is obtained. : ， in, For hyperparameters; The contrastive learning loss function With cross-entropy loss function The final loss function L is obtained by combining these: , For hyperparameters; Among them, the cross-entropy loss function The calculation formula is as follows: , The model is for the first The predicted probabilities of each category; Once the loss function is determined, the accuracy of the model can be improved by iteratively training the model. After multiple iterations, training is stopped to obtain the final model, which is then used to generate the final summary.
2. The domain-frontier dynamic progressive summarization system based on a pre-trained model as described 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 into a unified encoding standard; for files that are not plain text, it parses their content and extracts usable plain text information. The text cleaning submodule completes the text cleaning operation, including: Check whether the data conforms to predefined rules and constraints; Based on language characteristics, word segmentation is performed using spaces; Remove irrelevant characters and non-text content; Standardize the formatting, and unify the formatting of line breaks, indentation, and spaces; Remove redundant characters, and delete irrelevant control symbols and duplicate spaces; Perform case conversion to convert all English characters to lowercase to prevent them from affecting the model.
3. The domain-frontier dynamic summarization system based on a pre-trained model as described in claim 1, characterized in that, The text vector construction submodule first needs to perform word segmentation on the input text. The tokenizer corresponding to the model is used to segment the original text and the summary into words or sub-word units. Then, word embedding is performed to map the word segmentation results to a predefined vocabulary. The word segmentation is converted into integer indices according to the model vocabulary. The word embeddings are then encoded by the encoder. Through the stacking and operation of multiple encoder modules, corresponding text vectors are generated. By adding a self-attention mechanism, the model can 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. The current position is obtained by weighting the value vectors of all positions using the attention score as the weight. Attention output: in For attention score, For value vectors, The length of the sequence is the number of words. Based on the obtained positional encoding and attention weights, convergence is optimized through residual connections to improve the training stability of the model. A Feedforward layer was added to the module to perform non-linear feature extraction, thereby enhancing the model's expressive power and adaptability to input data. The resulting text vector is then provided to the global attention decoder submodule for further processing. 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 meticulously encoded to achieve better summary generation. For each time step, the self-attention layer is still used to understand the context of the generated part, and a causal mask is added to ensure that each position only interacts with all positions before it; the cross-attention layer is set to connect to the context information of the entire input sequence, and combined with the hidden state of the current time step, the cross-attention context vector is obtained. Residual connections and feedforward layers are still used to stabilize the training state and accelerate convergence; vocabulary prediction is performed on the final hidden information to obtain the probability distribution of the corresponding words at the current time step. A greedy decoder strategy is adopted, selecting the word with the highest probability as the next word each time: in Indicates time step Output words, It is a vocabulary list One of the words, Indicates time step The previously generated word sequence, Input data; Finally, decoding is completed, and a text summary, or candidate summary, is obtained.
4. The domain-frontier dynamic summarization system based on a pre-trained model as described in claim 1, characterized in that, The personalized agent interaction module functions as follows: The personalized summary agent module collects multi-dimensional structured feedback data through multiple rounds of interaction with users, combines user behavior analysis and deep learning technology to dynamically generate user preference weights, and adjusts model parameters based on the weights to achieve personalized summary generation; This module, together with the information acquisition module and the text summarization module, forms a closed-loop optimization system. 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 employ the following mechanism to generate weights: 1) Preference Feature Fusion: Integrating user behavior feature matrices Explicit rating vector Implicit interest vectors Concatenate into a unified feature vector User preference weights are generated by inputting a multilayer perceptron (MLP). ,in The adjustable parameter dimension in the corresponding text summarization generation module; 2) Adaptive weight allocation: Through a contrastive learning strategy, weights are assigned adaptively... Align the model with the original parameter distribution of the pre-trained model to ensure that the personalized model still maintains domain generality; Closed-loop optimization and progressive iteration are as follows: 1) Relationship with preceding modules: The information acquisition module provides cutting-edge data in the field to support the generation of summaries, the text summarization generation module dynamically updates model parameters based on user feedback, and the personalized agent iteratively optimizes user profiles and weight generation strategies through feedback data, forming a closed loop of "data-generation-feedback-optimization". 2) Long-term personalized training: Build independent fine-tuning datasets for each user. It includes its historical interaction data and preference weights, and regularly updates the personalized model branches through incremental learning to avoid global model drift.
5. The domain-frontier dynamic summarization system based on a pre-trained model as described in claim 4, characterized in that, The multi-dimensional user feedback collection and processing includes: First, various user interaction data are collected, including: 1) Behavioral data: Record user click heatmaps of summaries, 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 the abstract quality through sliding rating, multi-dimensional tag selection, and text feedback; 3) Implicit Preference Data: Analyze users' browsing history and summary collection / sharing behavior to construct users' implicit interest vectors; The specific data processing methods for user interaction data are as follows: For text-based feedback data, a BERT-based sentiment analysis model is used to extract sentiment polarity and fine-grained requirements, and these are then mapped to structured labels. For behavioral data, time-series modeling using LSTM captures user interaction patterns, and an attention mechanism is used to calculate the importance scores of different summary segments, forming a user attention matrix. , The number of sentences in the summary. For feature dimensions.
6. The domain-frontier dynamic summarization system based on a pre-trained model as described in claim 4, characterized in that, The dynamic weight generation and model optimization are specifically as follows: Introduce a user preference term into the loss function of the text summarization generation module: ， in For model parameters, As the baseline model parameters, Control the intensity of personalized adjustments; Adjusting the decoder's attention mechanism: Will As a gating weight, the output of different attention heads is dynamically scaled: ， in To prioritize user-focused content during the generation process, the number of attention points is used.