Text abstract generation method based on large language model

Abstract

The invention relates to the technical field of natural language processing, and discloses a text abstract generation method based on a large language model. According to the method, to-be-processed text data and abstract length range and type are received, a preset text feature extraction model is utilized to extract a text key theme, emotional tendency and semantic complexity, an initial abstract generation strategy is dynamically generated according to the text key theme, emotional tendency and semantic complexity, a large language model is selected, and input parameters are set. Predicting the abstract quality score through a preset text quality prediction model, adjusting the weight parameter of the initial strategy, updating the abstract generation priority through an iterative optimization algorithm, and outputting a final text abstract. According to the method, the abstract generation accuracy and quality can be improved, the generation efficiency and resource consumption are balanced, and the requirement of a user for the text abstract is effectively met.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular to a text summary generation method based on a large language model. Background Art

[0002] In today's era of information explosion, text data is growing exponentially. Whether it's news, academic literature, corporate reports, or social media content, people are faced with massive amounts of textual information to process daily. In this context, how to quickly and accurately extract key information from this vast amount of text has become a critical issue for improving information processing efficiency and decision-making quality. Text summarization technology has emerged to address this issue. Its purpose is to distill long text into concise, summarized snippets, 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 formulated grammatical and semantic rules to extract key information from texts. However, this approach requires significant manpower and time, and the formulation of rules struggles to cover all linguistic phenomena, resulting in poor adaptability. Statistical methods, on the other hand, calculate statistical features of words and sentences in a text, such as word frequency and sentence length, to identify important sentences for summarization. However, these methods only consider surface features and are unable to deeply understand the semantics and logical relationships of the text. The resulting summaries are often of low quality and difficult to meet user needs.

[0004] With the development of machine learning technology, machine learning-based text summarization methods have gradually emerged. These methods train models based on large amounts of text data to predict the importance of text and generate summaries. Common machine learning algorithms such as support vector machines and decision trees have been applied to text summarization. Although these methods have made certain improvements over 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 the generated summaries lacking coherence and logic.

[0005] The emergence of deep learning technology has brought new breakthroughs in 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) networks, and gated recurrent units (GRUs), are able to better process the sequential nature of text and learn semantic representations. However, these models still face challenges when processing large amounts of text data, such as high computational resource consumption and long training times.

[0006] In recent years, large language models (LLMs), with their powerful language understanding and generation capabilities, have achieved remarkable results in various fields of natural language processing, including text summarization. By pre-training on large-scale text corpora, LLMs acquire rich linguistic knowledge and semantic information, enabling them to generate high-quality text summaries. However, current text summarization methods based on LLMs still have some shortcomings. First, different LLMs perform differently in different text scenarios, and selecting the appropriate LLM based on the characteristics of the text and properly configuring its input parameters remain unresolved. Second, the quality of generated summaries varies widely, and the lack of effective quality assessment and optimization mechanisms makes it difficult to ensure that the generated summaries accurately and comprehensively reflect the core content of the original text. Furthermore, in practical applications, model generation efficiency and resource consumption must be considered to meet the needs of different users and scenarios. In summary, existing text summarization methods suffer from numerous deficiencies in accuracy, efficiency, and resource utilization, failing to meet the demands of practical applications. Therefore, developing a more efficient, intelligent, and precise text summarization method based on LLMs is of great practical significance. Summary of the Invention

[0007] The purpose of the present invention is to provide a text summary generation method based on a large language model to solve the problems raised in the above background technology.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for generating a text summary based on a large language model, the method comprising: Receive text data to be processed and a user-specified summary length range and summary type; Based on a preset text feature extraction model, feature extraction is performed on the text data to be processed to determine the key themes, sentiment tendencies and semantic complexity of the text; Dynamically generate an initial summary generation strategy based on the key topics, sentiment tendency, and semantic complexity, the strategy including the selection of a large language model and the setting of input parameters; Based on a preset text quality prediction model, predict the quality scores of summaries generated under different summary generation strategies, and adjust the dynamic weight parameters in the initial summary generation strategy; Under the adjusted summary generation strategy, the priority of summary generation is updated through the iterative optimization algorithm, and the final text summary is output to the user end.

[0009] Preferably, the steps of constructing the text feature extraction model include: Collect a large amount of text data from different fields and build a training sample set that includes text topics, sentiment polarity, vocabulary complexity and sentence structure; Performing feature learning on the training sample set through a convolutional neural network in deep learning to generate initial feature labels; The initial feature labels are modified in combination with expert annotations to obtain multi-level feature rules; The multi-level feature rules are embedded in a pre-trained language model to generate the text feature extraction model that can be dynamically updated.

[0010] Preferably, the dynamic generation of the initial summary generation strategy includes: Assigning initial weight coefficients to different large language models based on the key topics, sentiment tendencies, and semantic complexity; Calculate the applicability score of each model based on the performance indicators and real-time resource usage of the large language model; Performing weighted fusion of the applicability score and the initial weight coefficient to generate a selection strategy for a large language model; Based on the model selection results, combined with the summary length range and summary type, set the input parameters of the large language model.

[0011] Preferably, the steps of constructing the text quality prediction model include: Collect historical text summary data and the corresponding manual evaluation quality scores to build a text quality dataset; Extracting time series features from the text quality dataset using a recurrent neural network to determine the temporal correlation of summary quality; 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 association matrix are input into a long short-term memory network to generate the text quality prediction model.

[0012] Preferably, the method further comprises: identifying summary generation strategies with quality scores below a preset threshold based on prediction results of the text quality prediction model; assigning a low priority tag to the identified policy in the initial summary generation policy; Based on the low-priority flag, resource allocation to the large language model under the corresponding strategy is reduced, and other high-priority strategies are preferentially tried.

[0013] Preferably, the calculation of the suitability score includes: Obtain the real-time response time, memory usage, and error rate of large language models and build a dynamic performance evaluation matrix; The relative importance of each model under different performance indicators is calculated through the hierarchical analysis method to generate the basic applicability value; Performing a weighted operation on the dynamic performance evaluation matrix and the basic applicability value to obtain the applicability score; The calculation formula for the applicability score is: ; Where, represents the suitability score, Indicates the Real-time values of performance indicators, Indicates the The weight of each performance indicator, represents the basic suitability value, Indicates the total number of performance indicators.

[0014] Preferably, the modification of the multi-level feature rules includes: Performing confidence evaluation on the initial feature labels and screening low-confidence samples; Performing label completion on the low-confidence samples through an active learning algorithm; According to the completed tags, the hierarchical logic of the multi-level feature rules is reconstructed.

[0015] Preferably, the execution of the iterative optimization algorithm includes: Setting an optimization objective function for the summary generation strategy, wherein the function includes constraints on summary quality, generation efficiency, and resource consumption; Performing multiple rounds of iterative adjustments on the dynamic weight parameters by using a genetic algorithm; In each round of iteration, the population is dynamically screened according to the degree of constraint violation, and the optimal solution set is retained; Outputting the summary generation strategy that meets the preset convergence condition; Wherein, the optimization objective function is: ; Where, represents the target optimization function, represents the generation efficiency score, Indicates the resource consumption score, represents the summary quality score, 、 and are the normalized weight coefficients of efficiency, resource consumption and quality respectively.

[0016] Preferably, the method further comprises: After assigning the low priority tag, monitoring the summary generation progress under the corresponding strategy in real time; If the generation progress does not reach the expected rate, the adaptive adjustment mechanism is triggered to dynamically adjust the input parameters of the large language model or replace the model.

[0017] Preferably, the setting of the resource consumption constraint includes: Collect the rated power and historical resource consumption data of each processing device to build a resource consumption baseline model; Calculate the upper limit of dynamic resource consumption based on the real-time device load rate; Using the upper limit of dynamic resource consumption as a constraint condition of the optimization objective function; The calculation formula for the upper limit of dynamic resource consumption is: ; Where, Indicates the upper limit of dynamic resource consumption, Indicates the rated power of the device. Indicates the real-time load rate. is the equipment performance correction factor.

[0018] Compared with the prior art, the present invention has the following beneficial effects: From an accuracy perspective, this invention, through its pre-defined text feature extraction model, enables a comprehensive and in-depth analysis of the text data being processed. It not only identifies the key themes of the text and accurately captures its core points, but also captures its emotional tendencies, ensuring that the generated summary maintains emotional consistency with the original text and avoiding discrepancies in information transmission. Furthermore, analysis of semantic complexity helps to more rationally organize the summary content, ensuring that it better aligns with the inherent logic of the text, significantly improving its accuracy and enabling users to accurately understand the core content of the original text through the summary.

[0019] The present invention is highly intelligent and targeted in its selection and parameter setting of large language models. It dynamically generates an initial summary generation strategy based on the key themes, sentiment, and semantic complexity of the text, fully considering the characteristics and applicable scenarios of different models. By calculating the applicability score of each model and integrating it with the initial weight coefficient, the most appropriate large language model is selected. Furthermore, input parameters are set based on the summary length range and type to ensure that the summary generated by the model meets user needs in both content and form, thereby improving the quality and practicality of the summary.

[0020] The text quality prediction model collects historical text summary data and manually assesses quality scores, uses a recurrent neural network to extract time series features, and builds a quality correlation matrix based on semantic similarity and grammatical correctness. This model accurately predicts summary quality scores for different summary generation strategies. This allows strategies to be evaluated and adjusted before generating a summary, optimizing summary quality in advance. Furthermore, strategies with low quality scores are assigned a low-priority tag, reducing resource allocation and prioritizing higher-priority strategies. This avoids wasting resources on low-quality strategies and improves overall summary generation efficiency.

[0021] During the summary generation process, the use of an iterative optimization algorithm further improves summary quality and resource efficiency. An optimization objective function is set that incorporates constraints on summary quality, generation efficiency, and resource consumption. Dynamic weight parameters are adjusted through multiple rounds of genetic algorithms. During the iterations, the population is dynamically screened based on the degree of constraint violation, retaining the optimal solution set. Ultimately, a summary generation strategy that meets the pre-set convergence criteria is output. This not only ensures a high level of summary quality, but also controls generation efficiency and resource consumption, achieving a balance between the three.

[0022] In addition, the present invention also has an adaptive adjustment mechanism. After assigning a low-priority tag, the summary generation progress is monitored in real time. If the progress does not reach 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 emergencies, ensure the smooth progress of the summary generation process, and improve the stability and reliability of the system. Overall, the present invention optimizes the text summary generation process from multiple dimensions, while improving the accuracy, quality and generation efficiency of the summary, and reasonably controls resource consumption, providing strong support for the development and practical application of text summary generation technology, and has broad application prospects and significant economic value. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a diagram showing the working principle of the text summary generation method based on a large language model according to the present invention; Figure 2 Adjust the flowchart for the summary generation strategy based on text quality prediction results; Figure 3 Modify the process schematic for multi-level feature rules; Figure 4 Optimizing a flow chart for a summary generation strategy based on a genetic algorithm. DETAILED DESCRIPTION

[0024] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0025] See also Figure 1-Figure 4 The present invention provides a method for generating a text summary based on a large language model, and the specific implementation steps are as follows: Receiving input data: The system has a receiving function that can receive text data to be processed and obtain the user-specified summary length range and summary type. For example, a user may provide a news article as the text data to be processed, specifying a summary length of 100-200 words and a summary type of "key information extraction".

[0026] Text Feature Extraction: This model uses a pre-set text feature extraction model to extract features from the incoming text data. This model identifies the key themes of the text—the core content surrounding it. It also analyzes the sentiment of the text to determine whether it is positive, negative, or neutral. It also assesses semantic complexity to understand the difficulty of the text. For example, if the input text is a scientific paper, the model will identify the core research topic as the key theme, determine sentiment based on word choice and tone, and assess semantic complexity based on vocabulary and sentence structure.

[0027] Dynamically generate an initial summary strategy: This strategy is dynamically generated based on the extracted key topics, sentiment, and semantic complexity. This strategy primarily involves selecting a large language model and setting its input parameters. Different large language models have their own strengths when processing different types of text. Choosing an appropriate model based on text features and setting appropriate input parameters can improve the quality and efficiency of summary generation.

[0028] Predicting summary quality and adjusting strategies: Using a pre-set text quality prediction model, we predict the quality scores of summaries generated under different summary generation strategies. Based on the prediction results, we adjust the dynamic weighting parameters in the initial summary generation strategy. If a particular strategy predicts a low summary quality score, we adjust the weighting parameters associated with that strategy to reduce its influence in the summary generation process.

[0029] Iterative optimization and summary output: Based on the adjusted summary generation strategy, an iterative optimization algorithm is used to update the summary generation priority. Through continuous iteration, the generated summary achieves a balance between quality, generation efficiency, and resource consumption. Ultimately, a final text summary that meets the requirements is output to the user end to meet their needs.

[0030] The implementation of the present invention will be further described below with reference to Examples 1 to 6.

[0031] Example 1: Building a text feature extraction model requires a multi-step process. First, a large amount of text data from various fields, such as news, academic papers, and literary works, is collected to construct a training sample set. This sample set contains rich information such as text topics, sentiment polarity, vocabulary complexity, and sentence structure. For example, a news report is labeled with the theme "Science and Technology Achievement Release," the sentiment polarity is positive, the vocabulary complexity is determined by the ratio of specialized vocabulary to common vocabulary, and the sentence structure is determined through grammatical analysis.

[0032] Next, a convolutional neural network (CNN) is used in deep learning to perform feature learning on the training sample set, generating initial feature labels. CNNs can automatically extract various features from text, such as identifying key phrases and sentence sentiment. However, these initial feature labels may be inaccurate, necessitating correction based on expert annotation. Experts review and adjust the initial feature labels based on their expertise and experience, generating multi-level feature rules. During this process, the initial feature labels are evaluated for confidence, filtering out low-confidence samples. For these samples, label completion is performed using an active learning algorithm. This algorithm selects 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 them more reasonable and accurate. The revised multi-level feature rules are embedded in a pre-trained language model to generate a dynamically updateable text feature extraction model. As new text data is added, the model continuously learns and updates, improving its ability to extract text features.

[0033] This process involves revising multi-level feature rules. For example, when processing text in a specific field, the initial feature labels may inaccurately judge the sentiment of some text. These low-confidence samples are screened out through confidence assessment. The active learning algorithm selects representative samples, which are then relabeled by experts to complete the label information. Then, based on these completed labels, the hierarchical logic of the multi-level feature rules is rebuilt. For example, the priority of the sentiment judgment rules may be adjusted or new judgment conditions may be added, enabling the model to more accurately extract text features in subsequent processing.

[0034] Example 2: The dynamic initial summary generation strategy first assigns initial weights to different large language models based on the key topics, sentiment, and semantic complexity of the text. For example, for a literary review with a strong sentiment profile, a large language model that performs well on sentiment analysis-related text will be assigned a relatively high initial weight.

[0035] 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 provide a direct reflection of the model's performance during operation. The analytic hierarchy process (AHP) is used to calculate the relative importance of each model under different performance indicators and generate a basic applicability value. The AHP method determines the weights of different performance indicators by establishing a hierarchical model, constructing a judgment matrix, and calculating a weight vector, thereby obtaining a basic applicability value.

[0036] Then, the dynamic performance evaluation matrix and the basic applicability value are weighted to obtain the applicability score. The calculation formula is: ; In this formula, Represents the applicability score, which is used to measure the applicability of the large language model in the current situation; Indicates the The real-time value of a performance indicator, where the performance indicators are the real-time response time, memory usage, and error rate mentioned above; Indicates the The weight of each performance indicator is determined by the hierarchical analysis method, reflecting the relative importance of each performance indicator; It represents the basic applicability value, which comprehensively considers the performance of the model under different performance indicators; Indicates the total number of performance indicators. In this embodiment .

[0037] Finally, the applicability score is weighted and fused with the initial weight coefficient to generate a large language model selection strategy. Based on the model selection results, combined with the summary length range and summary type, the input parameters of the large language model are set. For example, after selecting a model, the maximum length parameter for generated summaries is set based on the summary length range, and relevant prompt information parameters are set based on the summary type to guide the model to generate a summary that meets the requirements.

[0038] Example 3: This embodiment describes in detail the process of building a text quality prediction model. The specific method includes: We collected historical text summary data and its corresponding manually assessed quality scores to construct a text quality dataset. The data came from a wide range of sources, covering a wide range of text fields, such as news, academic papers, and fiction. For news, we collected reports and their corresponding summaries on various topics (such as politics, economics, and science). We then invited professional editors or researchers to manually score these summaries based on multiple criteria, including content completeness, information accuracy, language fluency, and logical coherence, with a maximum score of 10. For example, a news summary about a scientific and technological achievement received a manual evaluation score of 8 for content completeness, 9 for information accuracy, 7 for language fluency, and 8 for logical coherence, for an overall score of 8. For academic paper summaries, in addition to the aforementioned criteria, we also scored them based on the extraction of academic value and the accuracy of terminology. By collecting this large amount of data, we constructed a dataset containing a rich set of text summaries and their corresponding quality scores, which served as the foundation for model training.

[0039] A recurrent neural network (RNN) is used to extract time series features from a text quality dataset and determine the temporal correlation of summary quality. Recurrent neural networks (RNNs) have memory capabilities and are capable of processing data with time series characteristics. Text quality data is arranged chronologically, for example, by the time news or papers are published, and fed into the RNN. Through continuous iterative learning, the RNN uncovers potential correlations between the quality of summaries generated at different time points. For example, analysis revealed that the urgency of news event reporting caused the quality of some news summaries to decline within a certain period of time. However, as reporting progressed and became more standardized, summary quality improved. This discovery of temporal correlation helps the model better understand the changing trends in summary quality, providing more comprehensive information for subsequent predictions.

[0040] Combining the semantic similarity and grammatical correctness of the text, a quality association matrix is constructed to quantify the relationship between different summary generation strategies and summary quality. Semantic similarity is measured by calculating the degree of semantic match between the summary and the original text. A common method is cosine similarity based on word vectors. For example, the summary and original text are converted into word vector representations, and the cosine value of the two vectors is calculated to obtain semantic similarity. Syntactic correctness is assessed by using grammar checking tools such as Stanford CoreNLP to check for grammatical errors in the summary, count the errors, and quantify them. Based on the quantified results of semantic similarity and grammatical correctness, a quality association matrix is constructed for different summary generation strategies. The rows of the matrix represent different summary generation strategies, and the columns represent different quality assessment metrics (such as semantic similarity and number of grammatical errors). The elements in the matrix represent the specific values of each metric under the corresponding strategy. In this way, the quality association matrix clearly shows the quantitative relationship between different summary generation strategies and summary quality.

[0041] The temporal correlation and quality association matrices are input into a long short-term memory (LSTM) network to generate a text quality prediction model. LSTM is a special type of recurrent neural network that effectively addresses long-term dependencies and is advantageous for processing complex time series data. The temporal correlation features and quality association matrix extracted by the RNN are used as inputs to the LSTM. Through deep learning of this information, the LSTM automatically captures the complex nonlinear relationship between summary generation strategy and summary quality. During training, the LSTM parameters are continuously adjusted to ensure that the model's predictions are as close as possible to the actual quality scores assessed by human assessments. After training and optimization on a large amount of data, a text quality prediction model is ultimately generated. This model can accurately predict the quality of summaries produced by different summary generation strategies before they are generated, providing strong support for selecting the optimal summary generation strategy and thus improving the overall quality of text summaries.

[0042] Example 4: Based on the prediction results of the text quality prediction model, summary generation strategies with quality scores below a preset threshold are identified. This threshold is not set arbitrarily but is 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 manually evaluated news summary data revealed that summary quality scores above 85 (out of 100) effectively meet users' needs for key information. Therefore, 85 was set as the preset threshold. During actual operation, the model assigns a quality prediction score to each summary generation strategy. Once a strategy's prediction score falls below 85, it is accurately identified. For example, when processing a news release about a scientific and technological achievement, a summary generated by a certain strategy might have a predicted quality score of 80, triggering subsequent processing.

[0043] After low-quality strategies are identified, they are assigned a low-priority tag in the initial summary generation strategy. This tag is like attaching a special label to the strategy, indicating that it has a lower priority in the subsequent summary generation process. This tagging method builds a screening mechanism within the system, so that subsequent resource allocation and strategy selection can be carried out more reasonably. Based on the low-priority tag, the system will reduce the resource allocation to the large language model under the corresponding strategy, and give priority to trying other high-priority strategies. For example, in the multi-model parallel processing architecture, 4 computing cores and 8GB of memory resources were originally allocated to the large language model corresponding to the low-quality strategy. After being marked as low priority, the computing cores were reduced to 2 and the memory resources were reduced to 4GB. These released resources are preferentially allocated to the models corresponding to those high-priority strategies, thereby ensuring that strategies that are more likely to generate high-quality summaries can get sufficient resource support.

[0044] After assigning a tag to a low-priority policy, the system monitors the progress of summary generation for that policy in real time. This monitoring process is continuous and dynamic, enabled by precise timestamp recording and a task progress tracking algorithm. For example, for a 1,000-word article summary, based on past experience and system-defined standards, 80% of the summary generation process is expected to be completed within 10 seconds. If, during actual monitoring, only 50% of the summary generation task is complete after 10 seconds, indicating that the generation progress is not meeting the expected rate, an adaptive adjustment mechanism is immediately triggered.

[0045] 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 length limit for generated summaries, originally set at 150 words, might be adjusted to 200 words in this case. The hope is that increasing the length of the summary will capture more complete information, thereby improving summary quality. The model's temperature parameter (a parameter that controls the randomness of the generated text) is also 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 fails to improve generation progress, the system will consider replacing 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 field (such as the technology field of the current news article). This model leverages its strengths 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.

[0046] Example 5: The iterative optimization algorithm first sets the optimization objective function of the summary generation strategy, which is: ; In this formula, represents the target optimization function, which is a comprehensive indicator to measure the quality of summary generation strategy; Represents the generation efficiency score, which is used to evaluate the speed of abstract generation. For example, the shorter the time required to generate an abstract, the higher the score; Indicates the resource consumption score, which reflects the consumption of computing resources in the process of generating the summary. The lower the consumption, the higher the score. Indicates the quality score of the abstract, which is evaluated based on the content completeness, accuracy, language fluency and other aspects of the abstract. The higher the quality, the higher the score; 、 and are the normalized weight coefficients of efficiency, resource consumption and quality, respectively, which are used to adjust the relative importance of these three factors in the optimization objective function. Their values range from 0 to 1, and .

[0047] Next, the dynamic weight parameters are iteratively adjusted over multiple rounds using a genetic algorithm. This algorithm simulates the natural evolutionary process, continuously optimizing the weight parameters through operations such as selection, crossover, and mutation, gradually bringing the summary generation strategy closer to the optimal solution. In each iteration, the population is dynamically screened 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. Only those strategies that meet resource consumption constraints and have high overall scores are retained.

[0048] Finally, the algorithm outputs a summary generation strategy that meets the preset convergence criteria. These criteria can include the change in the optimization objective function being less than a certain threshold over multiple iterations, or reaching a preset maximum number of iterations. When these criteria are met, the algorithm considers the ideal summary generation strategy to have been found and is used to generate the final text summary.

[0049] Example 6: This example describes in detail how to collect data to build a resource consumption baseline model, calculate the dynamic resource consumption upper limit, and apply it to the constraints of the optimization objective function. The specific method includes: Collect the rated power and historical resource consumption data of each processing device to construct a resource consumption baseline model. Processing devices may include servers, high-performance computers, and other hardware. For example, the rated power of a common server can be obtained by consulting its technical specifications, assuming it is 500 watts. Collecting historical resource consumption data requires the use of specialized system monitoring tools, such as Zabbix and Prometheus. These tools can monitor various resource consumption metrics, such as CPU usage, memory usage, and disk I / O read / write volume, while the server processes text summarization tasks in real time. Over a period of time, collect this data as the server processes different types of text summarization tasks. For example, when processing news text summarization, record the server's CPU usage at different time points, such as 1 minute, 5 minutes, and 10 minutes after the task starts, along with the corresponding memory usage and disk I / O read / write volume. By organizing and analyzing this large amount of data, and applying statistical methods and data modeling techniques, a resource consumption baseline model is constructed. This model reflects the resource consumption patterns of servers under normal conditions when processing text summarization tasks of varying scale and difficulty.

[0050] Calculate the upper limit of dynamic resource consumption based on the real-time device load rate. The real-time device load rate can be obtained in real time through the system monitoring tool, which reflects the current busyness of the device. Assume that the monitoring tool knows that the current real-time load rate of the server is 60%, that is, the server's workload at this moment has reached 60% of its full load capacity. The device performance correction coefficient 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 the historical operating data and experience of the device. It is assumed that it is set to 0.9 here. According to the formula ; in Indicates the upper limit of dynamic resource consumption, The server rated power of 500 watts obtained earlier, The real-time load factor is 60% (converted to a decimal of 0.6). Substituting these values into the formula yields: Watts, indicating that the server's upper limit for dynamic resource consumption under this load is 270 watts. This means that under the current operating conditions, when processing text summarization tasks, the server's resource consumption should not exceed 270 watts to ensure stable operation and avoid system failures or significant performance degradation due to overload.

[0051] A dynamic resource consumption upper limit is used as a constraint for 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 upper limit into the constraint, when adjusting the summary generation strategy through optimization algorithms such as genetic algorithms, the resource consumption generated by each adjusted strategy will be ensured to not exceed this upper limit. For example, in a given genetic algorithm iteration, a potential summary generation strategy, while performing well in terms of summary quality and generation efficiency, may be calculated to consume 300 watts of resources, exceeding the current dynamic resource consumption upper limit of 270 watts. In this case, the strategy will be adjusted or discarded based on the degree of constraint violation, and a different strategy that meets the resource consumption constraint and has a higher overall score will be selected. In this way, while ensuring summary quality and generation efficiency, resource consumption is rationally controlled, achieving efficient utilization of system resources and ensuring the stable and efficient operation of the entire text summarization process under resource constraints.

[0052] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0053] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for generating text summarization based on a large language model, characterized in that: include: Receive text data to be processed and a user-specified summary length range and summary type; Based on a preset text feature extraction model, feature extraction is performed on the text data to be processed to determine the key themes, sentiment tendencies and semantic complexity of the text; Dynamically generate an initial summary generation strategy based on the key topics, sentiment tendency, and semantic complexity, the strategy including the selection of a large language model and the setting of input parameters; Based on a preset text quality prediction model, predict the quality scores of summaries generated under different summary generation strategies, and adjust the dynamic weight parameters in the initial summary generation strategy; Under the adjusted summary generation strategy, the priority of summary generation is updated through the iterative optimization algorithm, and the final text summary is output to the user end.

2. The text summarization method based on a large language model according to claim 1, characterized in that: The steps of constructing the text feature extraction model include: Collect a large amount of text data from different fields and build a training sample set that includes text topics, sentiment polarity, vocabulary complexity and sentence structure; Performing feature learning on the training sample set through a convolutional neural network in deep learning to generate initial feature labels; The initial feature labels are modified in combination with expert annotations to obtain multi-level feature rules; The multi-level feature rules are embedded in a pre-trained language model to generate the text feature extraction model that can be dynamically updated.

3. The text summarization generation method based on a large language model according to claim 1 is characterized in that The dynamic initial summary generation strategy includes: Assigning initial weight coefficients to different large language models based on the key topics, sentiment tendencies, and semantic complexity; Calculate the applicability score of each model based on the performance indicators and real-time resource usage of the large language model; Performing weighted fusion of the applicability score and the initial weight coefficient to generate a selection strategy for a large language model; Based on the model selection results, combined with the summary length range and summary type, set the input parameters of the large language model.

4. The text summarization generation method based on a large language model according to claim 1, characterized in that The steps of constructing the text quality prediction model include: Collect historical text summary data and the corresponding manual evaluation quality scores to build a text quality dataset; Extracting time series features from the text quality dataset using a recurrent neural network to determine the temporal correlation of summary quality; 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 association matrix are input into a long short-term memory network to generate the text quality prediction model.

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

6. The method for generating text summarization based on a large language model according to claim 3, characterized in that: The calculation of the suitability score includes: Obtain the real-time response time, memory usage, and error rate of large language models and build a dynamic performance evaluation matrix; The relative importance of each model under different performance indicators is calculated through the hierarchical analysis method to generate the basic applicability value; Performing a weighted operation on the dynamic performance evaluation matrix and the basic applicability value to obtain the applicability score; The calculation formula for the applicability score is: ； Where, represents the suitability score, Indicates the Real-time values of performance indicators, Indicates the The weight of each performance indicator, represents the basic suitability value, Indicates the total number of performance indicators.

7. The method for generating text summarization based on a large language model according to claim 2, characterized in that: The modification of the multi-level feature rule includes: Performing confidence evaluation on the initial feature labels and screening low-confidence samples; Performing label completion on the low-confidence samples through an active learning algorithm; According to the completed tags, 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: Setting an optimization objective function for the summary generation strategy, wherein the function includes constraints on summary quality, generation efficiency, and resource consumption; Performing multiple rounds of iterative adjustments on the dynamic weight parameters by using a genetic algorithm; In each round of iteration, the population is dynamically screened according to the degree of constraint violation, and the optimal solution set is retained; Outputting the summary generation strategy that meets the preset convergence condition; Wherein, the optimization objective function is: ； Where, represents the target optimization function, represents the generation efficiency score, Indicates the resource consumption score, represents the summary quality score, 、 and are the normalized weight coefficients of efficiency, resource consumption and quality respectively.

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

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

Images