Text summarization generation method based on large language model

Abstract

The application relates to the technical field of natural language processing, and discloses a text abstract generation method based on a large language model. The method receives text data to be processed, an abstract length range and a type, extracts a text key theme, a sentiment tendency and semantic complexity by using a preset text feature extraction model, dynamically generates an initial abstract generation strategy, selects a large language model and sets input parameters. The abstract quality score is predicted by using a preset text quality prediction model, the weight parameters of the initial strategy are adjusted, the abstract generation priority is updated by using an iterative optimization algorithm, and a final text abstract is output. The method can improve the accuracy and quality of abstract generation, balance the generation efficiency and resource consumption, and effectively meet the needs of users for text abstracts.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a text summarization method based on a large language model. Background Technology

[0002] In today's information-saturated world, text data is growing exponentially. Whether it's news, academic literature, corporate reports, or social media content, people face a massive amount of textual information to process every day. In this context, how to quickly and accurately extract key information from large amounts of text has become a crucial issue for improving information processing efficiency and decision-making quality. Text summarization technology has emerged to address this need, aiming to extract long text content into concise, summary texts, helping users quickly understand the core points of the text.

[0003] Early text summarization methods were primarily rule-based and statistical. Rule-based methods rely on manually defined grammatical and semantic rules to extract key information from text. However, this approach is time-consuming and labor-intensive, and the rules are difficult to encompass all linguistic phenomena, resulting in poor adaptability. Statistical methods, on the other hand, determine important sentences for summarization by calculating statistical features of words and sentences in the text, such as word frequency and sentence length. However, these methods only address surface features and fail to delve into the semantics and logical relationships of the text, often resulting in low-quality summaries that do not meet user needs.

[0004] With the development of machine learning technology, machine learning-based text summarization methods have gradually emerged. These methods train models by learning from large amounts of text data to predict the importance of text and thus generate summaries. Common machine learning algorithms such as support vector machines and decision trees are applied to the field of text summarization. Although these methods have made some progress compared to earlier rule-based and statistical methods, they still have some limitations. For example, their ability to extract text features is limited, and they cannot fully capture the semantic information of the text, resulting in a lack of coherence and logic in the generated summaries.

[0005] The emergence of deep learning technology has brought new breakthroughs to the field of text summarization. Neural network-based text summarization models, such as recurrent neural networks (RNNs) and their variants Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRUs), can better handle the sequential features of text and learn its semantic representation. However, these models still face challenges such as high computational resource consumption and long training times when processing large-scale text data.

[0006] In recent years, Large Language Models (LLMs) have achieved remarkable results in various fields of natural language processing, including text summarization, thanks to their powerful language understanding and generation capabilities. By pre-training on large-scale text corpora, LLMs learn rich linguistic knowledge and semantic information, enabling them to generate high-quality text summaries. However, current LLM-based text summarization methods still have some shortcomings. On the one hand, different LLMs perform differently in different text scenarios; how to select the appropriate LLM based on the characteristics of the text and reasonably set its input parameters remains an unsolved problem. On the other hand, the quality of generated summaries varies, lacking effective quality assessment and optimization mechanisms, making it difficult to ensure that the generated summaries accurately and comprehensively reflect the core content of the original text. Furthermore, in practical applications, the generation efficiency and resource consumption of the model need to be considered to meet the needs of different users and scenarios. In summary, existing text summarization methods have many shortcomings in terms of accuracy, efficiency, and resource utilization, failing to adequately meet the needs of practical applications. Therefore, researching a more efficient, intelligent, and accurate LLM-based text summarization method is of significant practical importance. Summary of the Invention

[0007] The purpose of this invention is to provide a text summarization method based on a large language model to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a text summarization generation method based on a large language model, the method comprising:

[0009] Receive the text data to be processed, as well as the user-specified digest length range and digest type;

[0010] Based on a preset text feature extraction model, features are extracted from the text data to be processed to determine the key themes, sentiment tendencies, and semantic complexity of the text.

[0011] Based on the key themes, sentiment tendencies, and semantic complexity, an initial summary generation strategy is dynamically generated, which includes the selection of a large language model and the setting of input parameters.

[0012] Based on a preset text quality prediction model, the quality score of the generated summary under different summary generation strategies is predicted, and the dynamic weight parameters in the initial summary generation strategy are adjusted.

[0013] Under the adjusted summary generation strategy, the priority of summary generation is updated through an iterative optimization algorithm, and the final text summary is output to the user.

[0014] Preferably, the construction steps of the text feature extraction model include:

[0015] We collected a large amount of text data from different fields and constructed a training sample set that includes text topics, sentiment polarity, lexical complexity, and sentence structure.

[0016] The training sample set is used to learn features through a convolutional neural network in deep learning to generate initial feature labels.

[0017] The initial feature labels are corrected by combining expert annotations to obtain multi-level feature rules;

[0018] The multi-level feature rules are embedded into a pre-trained language model to generate a dynamically updatable text feature extraction model.

[0019] Preferably, the dynamic initial summary generation strategy includes:

[0020] Based on the key themes, sentiment tendencies, and semantic complexity, initial weight coefficients are assigned to different large language models;

[0021] Based on the performance metrics and real-time resource consumption of the large language model, the applicability score of each model is calculated.

[0022] The applicability score is weighted and fused with the initial weight coefficient to generate a selection strategy for the large language model;

[0023] Based on the model selection results, and considering the abstract length range and abstract type, set the input parameters for the large language model.

[0024] Preferably, the steps for constructing the text quality prediction model include:

[0025] Collect historical text summarization data and corresponding human-assessed quality scores to construct a text quality dataset;

[0026] The temporal correlation of summary quality is determined by extracting time-series features from the text quality dataset using a recurrent neural network.

[0027] By combining the semantic similarity and grammatical correctness of the text, a quality correlation matrix is ​​established to quantify the relationship between different summary generation strategies and summary quality;

[0028] The time correlation and the quality correlation matrix are input into a long short-term memory network to generate the text quality prediction model.

[0029] Preferably, the method further includes:

[0030] Based on the prediction results of the text quality prediction model, identify summary generation strategies with quality scores below a preset threshold;

[0031] In the initial summary generation strategy, low-priority tags are assigned to the identified strategies;

[0032] Based on the low-priority tags, resource allocation to the large language model under the corresponding strategy is reduced, and other high-priority strategies are tried first.

[0033] Preferably, the calculation of the applicability score includes:

[0034] Obtain the real-time response time, memory usage, and error rate of a large language model, and construct a dynamic performance evaluation matrix;

[0035] The relative importance of each model under different performance indicators is calculated using the analytic hierarchy process (AHP) to generate basic applicability values.

[0036] The applicability score is obtained by weighting the dynamic performance evaluation matrix with the basic applicability value.

[0037] The formula for calculating the applicability score is as follows:

[0038] ;

[0039] In the formula, Indicates the applicability score. Indicates the first Real-time values ​​of each performance metric Indicates the first The weight of each performance metric, Indicates the basic applicability value. This indicates the total number of performance metrics.

[0040] Preferably, the correction of the multi-level feature rules includes:

[0041] The initial feature labels are evaluated for confidence, and low-confidence samples are filtered out.

[0042] The low-confidence samples are labeled using an active learning algorithm;

[0043] Based on the completed labels, the hierarchical logic of the multi-level feature rules is reconstructed.

[0044] Preferably, the execution of the iterative optimization algorithm includes:

[0045] Define an optimization objective function for the summary generation strategy, wherein the function includes constraints on summary quality, generation efficiency, and resource consumption;

[0046] The dynamic weight parameters are adjusted iteratively through multiple rounds using a genetic algorithm;

[0047] In each iteration, the population is dynamically filtered based on the degree of constraint violation, and the optimal solution set is retained;

[0048] Output the summary generation strategy that satisfies the preset convergence condition;

[0049] The optimization objective function is:

[0050] ;

[0051] In the formula, Denotes the objective optimization function. This indicates the generation efficiency score. Indicates resource consumption score, This indicates the quality score of the abstract. , and These are the normalized weighting coefficients for efficiency, resource consumption, and quality, respectively.

[0052] Preferably, the method further includes:

[0053] After assigning the low-priority tags, the progress of summary generation under the corresponding strategy is monitored in real time;

[0054] If the generation progress does not reach the expected rate, an adaptive adjustment mechanism is triggered to dynamically adjust the input parameters of the large language model or replace the model.

[0055] Preferably, the setting of the resource consumption constraint includes:

[0056] Collect rated power and historical resource consumption data of each processing device, and construct a resource consumption baseline model;

[0057] Calculate the dynamic resource consumption limit based on the real-time device load rate;

[0058] The upper limit of dynamic resource consumption is used as a constraint condition for the optimization objective function;

[0059] The formula for calculating the upper limit of dynamic resource consumption is as follows:

[0060] ;

[0061] In the formula, Indicates the upper limit of dynamic resource consumption. Indicates the rated power of the equipment. Indicates real-time load rate. This is the equipment performance correction factor.

[0062] Compared with the prior art, the beneficial effects of the present invention are:

[0063] In terms of accuracy, this invention, through a pre-defined text feature extraction model, can comprehensively and deeply analyze the text data to be processed. It not only identifies the key themes of the text and accurately grasps its core points, but also understands the sentiment, ensuring that the generated summary maintains consistency with the original text in terms of emotional tone, thus avoiding deviations in information delivery. Simultaneously, semantic complexity analysis helps to organize the summary content more logically, making the summary more consistent with the text's internal logic, greatly improving the accuracy of the summary, and enabling users to accurately understand the core content of the original text through the summary.

[0064] This invention demonstrates high intelligence and specificity in the selection and parameter setting of large language models. It dynamically generates an initial summary generation strategy based on the key themes, sentiment tendencies, and semantic complexity of the text, fully considering the characteristics and applicable scenarios of different models. By calculating the applicability scores of each model and integrating them with the initial weight coefficients, the most suitable large language model is selected. Furthermore, by combining the summary length range and type settings with the input parameters, it ensures that the summaries generated by the model meet user needs in both content and form, thus improving the quality and practicality of the summaries.

[0065] The text quality prediction model collects historical text summarization data and human-assessed quality scores. It extracts time-series features using a recurrent neural network and establishes a quality correlation matrix by combining semantic similarity and grammatical correctness. This allows for accurate prediction of summary quality scores under different summarization strategies. This enables strategy evaluation and adjustment before summarization, allowing for early optimization of summary quality. Furthermore, strategies with low quality scores are assigned low-priority labels, reducing resource allocation and prioritizing higher-priority strategies. This avoids wasting resources on low-quality strategies and improves overall summarization efficiency.

[0066] In the process of generating summaries, the application of iterative optimization algorithms further improves the quality of summaries and the efficiency of resource utilization. An optimization objective function is set, incorporating constraints on summary quality, generation efficiency, and resource consumption. Dynamic weight parameters are adjusted through multiple rounds of iteration using a genetic algorithm. During the iteration process, the population is dynamically selected based on the degree of constraint violation, retaining the optimal solution set, and finally outputting a summary generation strategy that satisfies the preset convergence conditions. This not only ensures that the generated summaries achieve a high level of quality but also controls generation efficiency and resource consumption, achieving a balance among the three.

[0067] Furthermore, this invention also features an adaptive adjustment mechanism. After assigning low-priority tags, the progress of summary generation is monitored in real time. If the progress does not meet the expected rate, the input parameters of the large language model can be dynamically adjusted or the model can be replaced. This mechanism can flexibly respond to various unforeseen circumstances, ensuring the smooth progress of the summary generation process and improving the stability and reliability of the system. In summary, this invention optimizes the text summarization process from multiple dimensions, improving the accuracy, quality, and efficiency of summarization while rationally controlling resource consumption. It provides strong support for the development and practical application of text summarization technology, and has broad application prospects and significant economic value. Attached Figure Description

[0068] Figure 1 This is a schematic diagram illustrating the working principle of the text summarization generation method based on a large language model as described in this invention.

[0069] Figure 2 Flowchart for adjusting the summary generation strategy based on text quality prediction results;

[0070] Figure 3 Schematic diagram of the multi-level feature rule correction process;

[0071] Figure 4 The flowchart shows the optimization process for the summary generation strategy based on the genetic algorithm. Detailed Implementation

[0072] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] Please see Figures 1-4 This invention provides a text summarization generation method based on a large language model, the specific implementation steps of which are as follows:

[0074] Receiving Input Data: The system has a receiving function, capable of receiving text data to be processed, and simultaneously obtaining the user-specified summary length range and summary type. For example, a user might provide a news report as the text data to be processed, specifying a summary length of 100-200 words and a summary type of "key information extraction".

[0075] Text Feature Extraction: A pre-defined text feature extraction model is used to extract features from the received text data. This model identifies the key themes of the text, i.e., the core content around which the text revolves; analyzes the sentiment tendency of the text, determining whether it is positive, negative, or neutral; and assesses semantic complexity to understand the difficulty level of the text content. Assuming the input text is a scientific paper, the model will identify the core research topic as the key theme, determine the sentiment tendency based on word choice and tone, and assess semantic complexity based on vocabulary and sentence structure.

[0076] Dynamic Initial Summary Generation Strategy: Based on extracted key themes, sentiment trends, and semantic complexity, an initial summary generation strategy is dynamically developed. This strategy primarily involves the selection of a large language model and the setting of input parameters. Different large language models have their own advantages in processing different types of text. Combining text features to select an appropriate model and setting matching input parameters can improve the quality and efficiency of summary generation.

[0077] Predicting summary quality and adjusting strategies: Using a pre-defined text quality prediction model, the quality scores of generated summaries under different summarization strategies are predicted. Based on the prediction results, the dynamic weight parameters in the initial summary generation strategy are adjusted. If the predicted quality score of a particular strategy is low, the weight parameters related to that strategy are adjusted to reduce its influence on the summary generation process.

[0078] Iterative optimization and summary output: Under the adjusted summary generation strategy, an iterative optimization algorithm is used to update the priority of summary generation. Through continuous iteration, a balance is achieved between the generated summaries in terms of quality, generation efficiency, and resource consumption. Finally, the final text summary that meets the requirements is output to the user's end, satisfying the user's needs.

[0079] The implementation of the present invention will be further described below with reference to Examples 1 to 6.

[0080] Example 1:

[0081] Building a text feature extraction model involves multiple steps. First, a large amount of text data from various fields, such as news articles, academic papers, and literary works, is collected to construct a training sample set. This sample set contains rich information such as text topic, sentiment polarity, lexical complexity, and sentence structure. Taking a news report as an example, its topic is labeled as "Release of Scientific and Technological Achievements," its sentiment polarity is positive, its lexical complexity is determined based on the ratio of specialized vocabulary to common vocabulary, and its sentence structure is derived through grammatical analysis.

[0082] Next, convolutional neural networks (CNNs) from deep learning are used to learn features from the training sample set, generating initial feature labels. CNNs can automatically extract various features from text, such as identifying key phrases and sentiment characteristics of sentences. However, these initial feature labels may be inaccurate, so expert annotation is needed for correction. Experts review and adjust the initial feature labels based on their professional knowledge and experience, resulting in multi-level feature rules. During this process, the confidence level of the initial feature labels is evaluated, and low-confidence samples are filtered out. For these samples, labels are completed using an active learning algorithm. Active learning algorithms can select the most valuable samples from a large amount of unlabeled data for annotation, improving annotation efficiency and accuracy. Finally, based on the completed labels, the hierarchical logic of the multi-level feature rules is reconstructed to make it more reasonable and accurate. The corrected multi-level feature rules are embedded into a pre-trained language model to generate a dynamically updatable text feature extraction model. As new text data is continuously added, the model can continuously learn and update, improving its ability to extract text features.

[0083] This process involves revising multi-level feature rules. Suppose that when processing text in a certain domain, the initial feature labels are inaccurate in judging the sentiment of some texts. These low-confidence samples are filtered out through confidence assessment. The active learning algorithm selects some representative samples, which are then re-labeled by experts to complete the label information. Then, based on these completed labels, the hierarchical logic of the multi-level feature rules is reconstructed. For example, the priority of sentiment judgment rules may be adjusted or new judgment conditions may be added, enabling the model to extract text features more accurately in subsequent processing.

[0084] Example 2:

[0085] The dynamic initial summary generation strategy first assigns initial weight coefficients to different large language models based on the text's key themes, sentiment tendencies, and semantic complexity. For example, for a literary review with a clear sentiment tendency, if a certain large language model performs well in processing sentiment analysis-related texts, it is assigned a relatively high initial weight coefficient.

[0086] Next, the real-time response time, memory usage, and error rate of the large language model are obtained to construct a dynamic performance evaluation matrix. These three metrics directly reflect the model's performance during operation. The relative importance of each model under different performance metrics is calculated using the Analytic Hierarchy Process (AHP) to generate basic applicability values. The AHP determines the weights of different performance metrics through steps such as establishing a hierarchical structure model, constructing a judgment matrix, and calculating weight vectors, thereby obtaining basic applicability values.

[0087] Then, the dynamic performance evaluation matrix is ​​weighted and calculated with the basic applicability value to obtain the applicability score. The calculation formula is as follows:

[0088] ;

[0089] In this formula, The applicability score measures the suitability of a large language model in the current context. Indicates the first The real-time values ​​of several performance metrics, namely the real-time response time, memory usage, and error rate mentioned earlier. Indicates the first The weights of each performance indicator are determined by the analytic hierarchy process (AHP) to reflect the relative importance of each performance indicator. This represents the basic applicability value, which comprehensively considers the model's performance under different performance metrics; This represents the total number of performance metrics, as described in this embodiment. .

[0090] Finally, the applicability score is weighted and fused with the initial weight coefficients to generate a selection strategy for the large language model. Based on the model selection results, and considering the summary length range and summary type, the input parameters of the large language model are set. For example, after determining which model to use, the maximum length parameter for generating summaries is set according to the summary length range, and relevant prompt information parameters are set according to the summary type to guide the model to generate summaries that meet the requirements.

[0091] Example 3:

[0092] This embodiment details the construction process of the text quality prediction model. Specific methods include:

[0093] A text quality dataset was constructed by collecting historical text summary data and corresponding human-assessed quality scores. The data sources are extensive, covering various text domains such as news, academic papers, and fiction. For news, reports on different topics (such as politics, economics, and science and technology) and their corresponding summaries were collected. Professional editors or researchers were invited to conduct human scoring based on multiple dimensions, including content completeness, information accuracy, language fluency, and logical coherence, with a maximum score of 10. For example, a news summary about a scientific achievement received a score of 8 for content completeness, 9 for information accuracy, 7 for language fluency, and 8 for logical coherence, for a total score of 8. For academic paper summaries, in addition to the above dimensions, scoring was also based on aspects such as the extraction of academic value and the accuracy of the use of professional terminology. Through the collection of a large amount of data, a dataset containing rich text summaries and their corresponding quality scores was constructed, which will serve as the basis for model training.

[0094] Recurrent Neural Networks (RNNs) are used to extract time-series features from a text quality dataset to determine the temporal correlation of summary quality. RNNs possess memory capabilities, enabling them to process data with time-series characteristics. Text quality data is arranged chronologically, such as according to the release date of news articles or the publication date of papers, and input into the RNN. Through iterative learning, the RNN uncovers the potential relationships between the quality of summaries generated at different points in time. For example, analysis reveals that the urgency of news events during a certain period leads to a decline in the quality of some news summaries, while the quality improves over time as the reports become more refined and standardized. This discovery of temporal correlation helps the model better understand the changing trends in summary quality, providing more comprehensive information for subsequent predictions.

[0095] By combining semantic similarity and grammatical correctness, a quality correlation matrix is ​​constructed to quantify the relationship between different summarization strategies and summarization quality. Semantic similarity is measured by calculating the degree of semantic matching between the summary and the original text. A common method is cosine similarity calculation based on word vectors. For example, the summary and the original text are converted into word vector representations, and the semantic similarity is obtained by calculating the cosine value of the two vectors. Grammatical correctness is assessed using grammar checking tools, such as Stanford CoreNLP, to check for grammatical errors in the summary, count the number of errors, and quantify them. Based on the quantification results of semantic similarity and grammatical correctness, a quality correlation matrix is ​​constructed for different summarization strategies. The rows of the matrix represent different summarization strategies, the columns represent different quality assessment indicators (such as semantic similarity, number of grammatical errors, etc.), and the elements in the matrix represent the specific values ​​of each indicator under the corresponding strategy. In this way, the quality correlation matrix clearly shows the quantitative relationship between different summarization strategies and summarization quality.

[0096] A text quality prediction model is generated by inputting the temporal correlation and quality correlation matrix into a Long Short-Term Memory (LSTM) network. LSTM is a special type of recurrent neural network that effectively solves long-term dependency problems and has advantages in processing complex time-series data. Using the temporal correlation features extracted by the RNN and the quality correlation matrix as input to the LSTM, the LSTM automatically captures the complex nonlinear relationship between summarization strategies and summary quality through deep learning of this information. During training, the parameters of the LSTM are continuously adjusted to make the model's prediction results as close as possible to the actual human-assessed quality scores. After training and optimization on a large amount of data, a text quality prediction model is finally generated. This model can accurately predict the potential summary quality of different summarization strategies before summary generation, providing strong support for selecting the optimal summary generation strategy and thus improving the overall quality of text summarization.

[0097] Example 4:

[0098] Based on the prediction results of the text quality prediction model, summary generation strategies with quality scores below a preset threshold are identified. This preset threshold is not arbitrary but determined after extensive experimentation and in-depth analysis of real-world application scenarios. Taking news text summarization as an example, statistical analysis of a large amount of existing manually evaluated news summary data revealed that a summary quality score of 85 points (out of 100) or higher effectively meets users' needs for obtaining key information. Therefore, 85 points was set as the preset threshold. In actual operation, the model performs a quality prediction score for each summary generation strategy. Once a strategy's prediction score falls below 85 points, that strategy is accurately identified. For example, when processing a news article about a scientific achievement, if a strategy generates a summary with a predicted quality score of 80 points, this triggers subsequent processing steps.

[0099] After identifying low-quality strategies, these strategies are assigned a low-priority label in the initial summary generation strategy. This labeling is like attaching a special tag to the strategy, indicating its lower priority in the subsequent summary generation process. This labeling method builds a filtering mechanism within the system, enabling more rational subsequent resource allocation and strategy selection. Based on the low-priority label, the system reduces resource allocation to the large language model under the corresponding strategy and prioritizes trying other high-priority strategies. For example, in a multi-model parallel processing architecture, the large language model corresponding to a low-quality strategy was originally allocated 4 computing cores and 8GB of memory resources. After being labeled as low-priority, the number of computing cores is reduced to 2, and the memory resources are reduced to 4GB. These freed-up resources are then prioritized for models corresponding to high-priority strategies, ensuring that strategies more likely to generate high-quality summaries receive sufficient resource support.

[0100] After assigning tags to low-priority strategies, the system monitors the summary generation progress under the corresponding strategies in real time. This monitoring process is continuous and dynamic, achieved through precise timestamp recording and task progress tracking algorithms. Taking the generation of a 1000-word article summary as an example, based on past experience and system-set standards, it is expected that 80% of the summary generation work will be completed within 10 seconds. If, during actual monitoring, it is found that only 50% of the summary generation task has been completed after 10 seconds, meaning the generation progress has not reached the expected rate, this will immediately trigger an adaptive adjustment mechanism.

[0101] Once the adaptive adjustment mechanism is activated, it will first attempt to dynamically adjust the input parameters of the large language model. For example, the limit parameter for the length of the generated summary, originally set at 150 characters, might be adjusted to 200 characters in this case, hoping to obtain more complete information and thus improve the summary quality by increasing the summary length. Simultaneously, the model's temperature parameter (a parameter controlling the randomness of generated text) will be adjusted, appropriately lowering its value to make the generated text more focused and stable, reducing the generation of irrelevant information. If adjusting the input parameters still cannot improve the generation progress, the system will consider changing the model. For example, if the large language model originally used focuses on general text processing, it might be switched to a model optimized for a specific domain (such as the technology field to which the current news belongs), leveraging its advantages in understanding and expressing professional knowledge to improve the efficiency and quality of summary generation. Through this series of closely coordinated operations, the summary generation process is continuously optimized to provide users with high-quality text summaries.

[0102] Example 5:

[0103] The iterative optimization algorithm first defines the objective function for the summary generation strategy, which is:

[0104] ;

[0105] In this formula, The objective function is a comprehensive indicator for evaluating the quality of a summary generation strategy. This represents the generation efficiency score, which is used to evaluate the speed of summary generation. For example, the shorter the time required to generate a summary, the higher the score. The resource consumption score reflects the consumption of computing resources during the summary generation process; the less resources consumed, the higher the score. The abstract quality score is based on factors such as the completeness, accuracy, and fluency of the language. Higher quality abstracts receive higher scores. , and These are the normalized weighting coefficients for efficiency, resource consumption, and quality, used to adjust the relative importance of these three factors in the optimization objective function. Their values ​​range from 0 to 1. .

[0106] Next, a genetic algorithm is used to iteratively adjust the dynamic weight parameters over multiple rounds. The genetic algorithm simulates the natural evolutionary process, continuously optimizing the weight parameters through selection, crossover, and mutation operations, gradually bringing the summary generation strategy closer to the optimal solution. In each iteration, the population is dynamically filtered based on the degree of constraint violation, retaining the optimal solution set. For example, if a strategy exceeds the resource consumption limit when generating a summary, it is penalized, reducing its survival probability in the population, and only strategies that satisfy the resource consumption constraints and have a high overall score are retained.

[0107] Finally, the output summarizes the summary generation strategy that meets the preset convergence criteria. These criteria can be that the change in the objective function over multiple iterations is less than a certain threshold, or that a preset maximum number of iterations has been reached. When these criteria are met, a relatively ideal summary generation strategy is considered to have been found and can be used to generate the final text summary.

[0108] Example 6:

[0109] This embodiment details how to collect data to construct a resource consumption baseline model, calculate the dynamic resource consumption upper limit, and apply it to the constraints of the optimization objective function. Specific methods include:

[0110] Collect rated power and historical resource consumption data for each processing device to construct a resource consumption baseline model. Processing devices may include various hardware facilities such as servers and high-performance computers. Taking a commonly used server as an example, its rated power can be obtained by referring to the device's technical specifications, let's assume it's 500 watts. Collecting historical resource consumption data requires specialized system monitoring tools, such as Zabbix and Prometheus. These tools can monitor various resource consumption indicators of the server in real time during text summarization tasks, such as CPU utilization, memory usage, and disk I / O read / write operations. Over a period of time, collect this data when the server processes different types of text summarization tasks. For example, when processing news text summarization, record the server's CPU utilization at different time points (1 minute, 5 minutes, 10 minutes) after the task starts, along with the corresponding memory usage and disk I / O read / write operations. By organizing and analyzing a large amount of such data, and using statistical methods and data modeling techniques, construct a resource consumption baseline model. This model can reflect the resource consumption patterns of the server when processing text summarization tasks of different scales and difficulties under normal conditions.

[0111] Calculate the dynamic resource consumption limit based on the real-time device load rate. The real-time device load rate can be obtained in real time through system monitoring tools, reflecting the current workload of the devices. Assuming the monitoring tool shows that the server's current real-time load rate is 60%, it means the server is currently operating at 60% of its full capacity. The device performance correction factor δ is a parameter that takes into account the impact of factors such as device aging and changes in the operating environment on device performance. It is usually set based on historical operating data and experience; here, we assume it is set to 0.9. According to the formula...

[0112] ;

[0113] in Indicates the upper limit of dynamic resource consumption. The server's rated power is 500 watts, as previously obtained. The real-time load rate is 60% (converted to 0.6 decimals). Substituting these values ​​into the formula yields: The watts refer to the maximum dynamic resource consumption of the server under this load condition, which is 270 watts. This means that under the current working conditions, when the server is processing text summarization tasks, the resource consumption should not exceed 270 watts to ensure stable operation of the device and avoid system failure or significant performance degradation due to overload.

[0114] The dynamic resource consumption cap is used as a constraint on the optimization objective function. During the summary generation process, the optimization objective function comprehensively considers factors such as summary quality, generation efficiency, and resource consumption. By incorporating the dynamic resource consumption cap as a constraint, adjustments to the summary generation strategy using optimization algorithms such as genetic algorithms ensure that the resource consumption of each adjusted strategy does not exceed this cap. For example, in a certain round of genetic algorithm iteration, a potential summary generation strategy may perform well in terms of summary quality and generation efficiency, but calculations show that its resource consumption may reach 300 watts, exceeding the current dynamic resource consumption cap of 270 watts. In this case, the strategy will be adjusted or discarded based on the degree of constraint violation, and other strategies that meet the resource consumption constraints and have a higher overall score will be selected. In this way, while ensuring summary quality and generation efficiency, resource consumption is reasonably controlled, achieving effective utilization of system resources and ensuring that the entire text summarization generation process can operate stably and efficiently under resource constraints.

[0115] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0116] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A text summarization generation method based on a large language model, characterized in that, include: Receive the text data to be processed, as well as the user-specified digest length range and digest type; Based on a preset text feature extraction model, features are extracted from the text data to be processed to determine the key themes, sentiment tendencies, and semantic complexity of the text. Based on the key themes, sentiment tendencies, and semantic complexity, an initial summary generation strategy is dynamically generated, which includes the selection of a large language model and the setting of input parameters. Based on a preset text quality prediction model, the quality score of the generated summary under different summary generation strategies is predicted, and the dynamic weight parameters in the initial summary generation strategy are adjusted. Under the adjusted summary generation strategy, the priority of summary generation is updated through an iterative optimization algorithm, and the final text summary is output to the user.

2. The text summarization method based on a large language model according to claim 1, characterized in that, The construction steps of the text feature extraction model include: We collected a large amount of text data from different fields and constructed a training sample set that includes text topics, sentiment polarity, lexical complexity, and sentence structure. The training sample set is used to learn features through a convolutional neural network in deep learning to generate initial feature labels. The initial feature labels are corrected by combining expert annotations to obtain multi-level feature rules; The multi-level feature rules are embedded into a pre-trained language model to generate a dynamically updatable text feature extraction model.

3. The text summarization method based on a large language model according to claim 1, characterized in that, The dynamic initial summary generation strategy includes: Based on the key themes, sentiment tendencies, and semantic complexity, initial weight coefficients are assigned to different large language models; Based on the performance metrics and real-time resource consumption of the large language model, the applicability score of each model is calculated. The applicability score is weighted and fused with the initial weight coefficient to generate a selection strategy for the large language model; Based on the model selection results, and considering the abstract length range and abstract type, set the input parameters for the large language model.

4. The text summarization method based on a large language model according to claim 1, characterized in that, The steps for constructing the text quality prediction model include: Collect historical text summarization data and corresponding human-assessed quality scores to construct a text quality dataset; The temporal correlation of summary quality is determined by extracting time-series features from the text quality dataset using a recurrent neural network. By combining the semantic similarity and grammatical correctness of the text, a quality correlation matrix is ​​established to quantify the relationship between different summary generation strategies and summary quality; The time correlation and the quality correlation matrix are input into a long short-term memory network to generate the text quality prediction model.

5. The text summarization method based on a large language model according to claim 4, characterized in that, Also includes: Based on the prediction results of the text quality prediction model, identify summary generation strategies with quality scores below a preset threshold; In the initial summary generation strategy, low-priority tags are assigned to the identified strategies; Based on the low-priority tags, resource allocation to the large language model under the corresponding strategy is reduced, and other high-priority strategies are tried first.

6. The text summarization method based on a large language model according to claim 3, characterized in that, The calculation of the applicability score includes: Obtain the real-time response time, memory usage, and error rate of a large language model, and construct a dynamic performance evaluation matrix; The relative importance of each model under different performance indicators is calculated using the analytic hierarchy process (AHP) to generate basic applicability values. The applicability score is obtained by weighting the dynamic performance evaluation matrix with the basic applicability value. The formula for calculating the applicability score is as follows: ； In the formula, Indicates the applicability score. Indicates the first Real-time values ​​of each performance metric Indicates the first The weight of each performance metric, Indicates the basic applicability value. This indicates the total number of performance metrics.

7. The text summarization method based on a large language model according to claim 2, characterized in that, The correction of the multi-level feature rules includes: The initial feature labels are evaluated for confidence, and low-confidence samples are filtered out. The low-confidence samples are labeled using an active learning algorithm; Based on the completed labels, the hierarchical logic of the multi-level feature rules is reconstructed.

8. The text summarization method based on a large language model according to claim 1, characterized in that, The execution of the iterative optimization algorithm includes: Define an optimization objective function for the summary generation strategy, wherein the function includes constraints on summary quality, generation efficiency, and resource consumption; The dynamic weight parameters are adjusted iteratively through multiple rounds using a genetic algorithm; In each iteration, the population is dynamically filtered based on the degree of constraint violation, and the optimal solution set is retained; Output the summary generation strategy that satisfies the preset convergence condition; The optimization objective function is: ； In the formula, Denotes the objective optimization function. This indicates the generation efficiency score. Indicates resource consumption score, This indicates the quality score of the abstract. , and These are the normalized weighting coefficients for efficiency, resource consumption, and quality, respectively.

9. The text summarization method based on a large language model according to claim 5, characterized in that, Also includes: After assigning the low-priority tags, the progress of summary generation under the corresponding strategy is monitored in real time; If the generation progress does not reach the expected rate, an adaptive adjustment mechanism is triggered to dynamically adjust the input parameters of the large language model or replace the model.

10. The text summarization method based on a large language model according to claim 8, characterized in that, The setting of the resource consumption constraints includes: Collect rated power and historical resource consumption data of each processing device, and construct a resource consumption baseline model; Calculate the dynamic resource consumption limit based on the real-time device load rate; The upper limit of dynamic resource consumption is used as a constraint condition for the optimization objective function; The formula for calculating the upper limit of dynamic resource consumption is as follows: ； In the formula, Indicates the upper limit of dynamic resource consumption. Indicates the rated power of the equipment. Indicates real-time load rate. This is the equipment performance correction factor.

Images