An efficient cross-document information extraction system and method based on large language model

Abstract

The present invention relates to the field of artificial intelligence technology, and in particular to an efficient cross-document information extraction system and method based on a large language model. The information compression module of the present invention not only effectively solves the information redundancy and fragmentation problems in cross-document event extraction through multi-format data conversion and efficient data cleaning, as well as feature analysis, metadata management, consistency verification and security measures, but also ensures the comprehensiveness and reliability of information processing. The event extraction module realizes accurate recognition and classification of events in complex texts through pre-trained large-scale language models and supervised sequence annotation models. The cross-document event collection module uses deep semantic analysis and intelligent similarity evaluation technology to effectively integrate relevant event information from different documents and different time points, providing users with a more complete and coherent event picture.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an efficient cross-document information extraction system and method based on a large language model. Background Art

[0002] In the era of information explosion, event extraction, as one of the core tasks in the field of information extraction, is crucial for understanding complex and ever-changing text data and mining potential events and their correlations. Especially when dealing with cross-document events, there are challenges such as large time spans between texts, scattered information distribution, and complex and ever-changing contexts.

[0003] Traditional methods often have difficulty effectively capturing and integrating the full picture of these events and their internal connections. Although current research has attempted to address the challenge of cross-document event extraction by fine-tuning small language models, in the face of increasingly complex and diverse text environments, these methods still face problems such as low event extraction accuracy and poor understanding of event relationships in complex contexts. Summary of the invention

[0004] In order to solve the technical problems existing in the above-mentioned background technology, the present invention provides an efficient cross-document information extraction method and system based on a large language model, which helps to improve the accuracy of event extraction and the ability to understand event relationships in complex contexts.

[0005] In order to implement the above technical solution, in a first aspect, the present invention provides an efficient cross-document information extraction system based on a large language model, comprising: an information compression module, which is used to perform in-depth analysis and summary of multiple documents in multiple formats through a large language model to obtain cross-document summary information;

[0006] The event extraction module is used to extract event information from the generated cross-document summary information through a large-scale language model and a supervised sequence labeling model;

[0007] The cross-document event collection module obtains the event picture through deep semantic analysis and vectorization representation based on the extracted event information;

[0008] Among them, the information compression module includes: a data preprocessing module, which is used to perform data conversion, data cleaning, feature analysis, metadata management, consistency verification and take security measures on multiple documents in multiple formats.

[0009] Furthermore, the data preprocessing module includes:

[0010] A metadata management unit for collecting and managing metadata for each document in a plurality of documents in a plurality of formats;

[0011] a document format conversion unit, configured to convert each of the plurality of documents in the plurality of formats into text data using a natural language processing tool;

[0012] A feature analysis unit, used to apply a machine learning algorithm to perform mean analysis and maximum and minimum length statistics on the converted text data to understand the text data;

[0013] A data cleaning unit, used for performing a thorough data cleaning on the text data based on the understood text data;

[0014] A data consistency verification unit performs text data consistency verification in the process of converting multiple documents of various formats into text data and in the process of cleaning the stored text data;

[0015] The Data Security Measures Implementation Unit is used to implement data security measures.

[0016] Furthermore, the information compression module also includes a text segmentation module, a summary extraction module and an information fusion module;

[0017] The text segmentation module is used to segment the text data of each document to obtain multiple text blocks of each document;

[0018] The summary extraction module is used to extract summary information of each document from multiple text blocks of each document;

[0019] The information fusion module is used to fuse the extracted summary information of each document to generate cross-document summary information.

[0020] Furthermore, the text segmentation module includes:

[0021] An efficient document loading and content extraction unit, used to identify the text data structure of each document to extract key content;

[0022] A recursive character text segmentation unit, for segmenting text data into a plurality of structured text blocks based on the extracted key content using a recursive algorithm;

[0023] Dynamic text block size adjustment unit, used to automatically adjust the size of the text block according to the length and complexity of the text data;

[0024] A sliding window technology application unit, used for applying a sliding window on a text block to obtain context information of the text block;

[0025] The adaptive text segmentation unit is used to dynamically adjust the segmentation strategy based on the context information of the obtained text block to obtain the semantic density and structural features.

[0026] Furthermore, the information compression module also includes: a multi-language support module and a summary quality assessment module;

[0027] The multi-language support module is used to evaluate the summary information obtained from the text data of each document in different language environments, and optimize the summary information based on the evaluation results;

[0028] The summary quality assessment module is used to perform quality assessment on the generated cross-document summary information and generate feedback information.

[0029] Furthermore, the event extraction module includes: a model training module, an event extraction module and an event evaluation module;

[0030] The model training module is used to train and fuse large-scale language models and supervised sequence labeling models;

[0031] The event extraction module is used to extract event information from cross-document summary information through a fused large-scale language model and a supervised sequence labeling model;

[0032] The event evaluation module is used to evaluate the performance of the deeply fused large-scale language model and the supervised sequence labeling model based on the extracted time information.

[0033] Further, the cross-document event collection module includes a deep semantic parsing and vectorized representation module, an intelligent similarity assessment and efficient grouping module, a cross-document event fusion and consistency maintenance module, and a cross-document event fusion and consistency maintenance module;

[0034] The deep semantic parsing and vectorization representation module is used to use text embedding technology to learn context information in the extracted event information through a deep neural network to convert the event information into a high-dimensional, dense semantic vector;

[0035] Intelligent similarity evaluation and efficient grouping module, based on the transformed semantic vector, uses the cosine similarity calculation method and clustering algorithm DBSCAN to intelligently identify and merge event information to generate merged event information;

[0036] The cross-document event fusion and consistency maintenance module is used to fuse and maintain the merged events to obtain the target event information;

[0037] The multi-dimensional event verification and information enhancement module is used to verify and enhance the target event information to obtain an event picture.

[0038] In a second aspect, the present invention provides an efficient cross-document information extraction method based on a large language model, comprising the following steps:

[0039] Step 1: Use a large language model to deeply analyze and summarize multiple documents in various formats to obtain summary information for each document;

[0040] Step 2: Extract event information from the generated cross-document summary information through a large-scale language model and a supervised sequence labeling model;

[0041] Step 3: Based on the extracted event information, deep semantic analysis and vectorization representation are performed to obtain the event picture;

[0042] Step 4: Continuously optimize the system through data feedback and algorithm iteration.

[0043] Furthermore, the step 1 comprises:

[0044] Preprocess multiple documents in multiple formats to obtain text data for each document;

[0045] Slice the text data of each document to obtain multiple text blocks of each document;

[0046] Extract summary information of each document from multiple text blocks of each document;

[0047] The summary information of each extracted document is fused to generate cross-document summary information.

[0048] In a third aspect, the present invention provides a computer-readable storage medium, which includes a stored program, wherein when the program is running, the device where the computer-readable storage medium is located is controlled to execute the above-mentioned efficient cross-document information extraction system based on a large language model.

[0049] The beneficial effects of the present invention are:

[0050] (1) The information compression module of the present invention not only effectively solves the information redundancy and fragmentation problems in cross-document event extraction through multi-format data conversion and efficient data cleaning, as well as feature analysis, metadata management, consistency verification and security measures, but also ensures the comprehensiveness and reliability of information processing. The event extraction module achieves accurate recognition and classification of events in complex texts through pre-trained large-scale language models and supervised sequence labeling models. The cross-document event collection module uses deep semantic analysis and intelligent similarity evaluation technology to effectively integrate relevant event information from different documents and different time points, providing users with a more complete and coherent event picture.

[0051] (2) In addition, the performance evaluation and optimization module continuously monitors the system operation status and ensures the continuous optimization and stability of system performance through data feedback and algorithm iteration. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0053] Figure 1 This is a principle block diagram of an efficient cross-document information extraction system based on a large language model of the present invention.

[0054] Figure 2 The present invention is a flowchart of an efficient cross-document information extraction method based on a large language model. DETAILED DESCRIPTION

[0055] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0056] It should be noted that the following detailed descriptions are all illustrative and are intended to provide further explanation of the present invention. Unless otherwise specified, each technical and scientific term used in this embodiment has the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0057] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0058] In the present invention, terms such as "upper", "lower", "left", "right", "front", "back", "vertical", "horizontal", "side", "bottom", etc. indicate directions or positional relationships based on the directions or positional relationships shown in the accompanying drawings. They are relational words determined only for the convenience of describing the structural relationships of the various parts or elements of the present invention, and do not specifically refer to any part or element in the present invention and should not be understood as limitations on the present invention.

[0059] In the present invention, terms such as "fixed connection", "connected", "connection", etc. should be understood in a broad sense, indicating that it can be fixedly connected, integrally connected or detachably connected; it can be directly connected or indirectly connected through an intermediate medium. Relevant scientific research or technical personnel in this field can determine the specific meanings of the above terms in the present invention according to specific circumstances, and they should not be understood as limiting the present invention.

[0060] Embodiment 1:

[0061] like Figure 1 As shown, this embodiment provides an efficient cross-document information extraction system based on a large language model, including the following steps:

[0062] (i) Information compression module: used to perform in-depth analysis and summary of multiple documents in various formats through a large language model (e.g., Qwen2) to obtain cross-document summary information.

[0063] The information compression module includes:

[0064] (1) Data preprocessing module: Preprocess multiple documents in various formats to obtain the text data of each document.

[0065] A) Metadata management unit: collects and manages metadata for each document in multiple documents of various formats, where the metadata for each document includes information such as the document's source, author, creation date, etc. These metadata are crucial for understanding the context of the data and conducting subsequent data analysis.

[0066] B) Document format conversion unit: Using natural language processing tools such as Word text recognizer and PDF extractor, each of multiple documents in multiple formats (such as PDF, HTML, and Word) is converted into easy-to-process text data. This conversion process not only helps improve the efficiency of data processing, but also enhances data consistency by eliminating format differences.

[0067] C) Feature Analysis Unit: Apply machine learning algorithms to perform mean analysis and maximum and minimum length statistics on the converted text data to understand the text data (such as the characteristics and distribution of the text data). This helps identify patterns and anomalies in the data and provides a basis for further data mining and knowledge discovery.

[0068] D) Data cleaning unit: Based on the understood text data, the text data is thoroughly cleaned, where data cleaning includes removing irrelevant information, noise filtering and data standardization. This helps to ensure the quality and accuracy of the text data and lays a solid foundation for subsequent analysis and processing.

[0069] E) Data storage unit: The cleaned data is stored in files of specific formats, such as JSON, XML, or CSV. These formats not only ensure the security of the data, but also improve the accessibility and readability of the data, facilitating subsequent data utilization and analysis.

[0070] F) Data consistency verification unit: In the process of converting multiple documents in various formats into text data and in the process of cleaning the stored text data, text data consistency verification is performed to ensure that data from different sources and formats remain consistent after integration, avoiding analysis errors caused by data inconsistency.

[0071] G) Implement Data Security Measures Unit: Implement strict data security measures during data storage and processing, including data encryption, access control, and audit logs to protect data from unauthorized access or tampering.

[0072] (2) A text segmentation module is used to segment the text data of each document into blocks to obtain multiple text blocks for each document.

[0073] The text segmentation module includes:

[0074] A) Efficient document loading and content extraction unit, used to identify the text data structure of each document and extract key content in a customized manner to ensure efficient acquisition and processing of information.

[0075] B) Recursive character text segmentation unit: Based on the extracted key content, the recursive algorithm is used to segment the text data into multiple structured text blocks according to natural language delimiters such as line breaks, spaces, punctuation, etc. This not only improves the flexibility of text data processing, but also lays the foundation for further semantic analysis and information extraction.

[0076] C) Dynamic text block size adjustment unit: Taking into account the contextual limitations of large language models, the dynamic text block size adjustment mechanism automatically adjusts the size of the text block according to the length and complexity of the text data to achieve optimal cutting of the text data. This not only meets the accuracy requirements, but also adapts to text data of different lengths, improving the adaptability and flexibility of processing.

[0077] D) Sliding window technology application unit: used to apply a sliding window on a text block to obtain broader contextual information while maintaining local information, thereby balancing the retrieval accuracy and the richness of contextual information, which helps to improve the accuracy of extracting summaries from text blocks, while also helping to retain contextual information in large block synthesis.

[0078] E) Adaptive text segmentation unit: It is used to dynamically adjust the segmentation strategy based on the contextual information of the text block to obtain semantic density and structural features. This enables the system to better handle text data of different types and styles, and improves the accuracy and efficiency of text segmentation.

[0079] F) Multi-dimensional text analysis unit: Perform multi-dimensional text analysis on text blocks, including but not limited to keyword extraction, topic identification, sentiment analysis, etc. These analyses provide rich information and perspectives for in-depth understanding and application of texts.

[0080] (3) Summary extraction module, which is used to extract summary information of each document based on multiple text blocks of each document.

[0081] A) Constructing effective prompt units: used to construct prompts similar to "help me extract summary information from the text" to guide the large language model to focus on the key information in the text. These prompts are customized according to the text content and context to ensure that the model can capture the most relevant summary information.

[0082] B) Use the context understanding unit: Understand the context information obtained by the sliding window technology application unit to ensure the coherence and relevance of the summary information and avoid information gaps or omissions.

[0083] C) Key Information Identification Unit: Through complex semantic analysis and pattern recognition technology, each text block is analyzed to identify key information such as topic sentences, important data, main points, etc., and a summary is constructed based on this.

[0084] D) Preliminary summary generation unit: used to generate a preliminary summary of each document based on the key information identified from each text block. In the process of generating the preliminary summary of each document, the main ideas and structure of the original text are retained as much as possible, while removing redundant and secondary information.

[0085] E) Summary coherence optimization unit: used to further optimize the generated preliminary summary to improve the summary coherence and generate an optimized summary. Specifically, the generated preliminary summary is optimized by adjusting the sentence order, adding transition words or phrases, etc. to ensure the logical fluency of the summary.

[0086] F) Summary conciseness and completeness balance unit: used to find the appropriate balance point between the length and information density of the optimized summary to ensure that the summary is concise while ensuring its completeness, that is, not omitting key information, and based on the balance point, to extract the summary information of each document.

[0087] (4) Information fusion module, which is used to fuse the summary information of each extracted document to generate cross-document summary information.

[0088] The information fusion module includes:

[0089] A) Hierarchical Attention Mechanism Unit: It is used to adopt multi-layer attention mechanism to capture key information at the word level for each document, and also perform importance evaluation at the sentence and paragraph level to more comprehensively understand the document structure and topic.

[0090] B) Context-aware embedding unit: It is used to dynamically adjust the representation of words or phrases according to the context through the dynamic embedding technology of the pre-trained language model to improve the accuracy of semantic understanding.

[0091] C) Long text processing capability: For long documents, the module adopts a strategy of block processing and global integration. It first processes the document in segments, and then integrates global information through a cross-segment attention mechanism to avoid information loss.

[0092] (5) Multilingual support module: used to translate the text data of each document into different languages ​​to obtain summary information of each document in different language environments, and to perform evaluation to optimize the summary information based on the evaluation results.

[0093] The multi-language support module includes:

[0094] A) Multilingual training model, which is used to train multilingual models (such as Qwen2, llama3) using a large amount of multilingual data, so that the multilingual model can understand the grammatical structure and semantic information of different languages. This can not only handle text translation, but also achieve cross-language understanding and generation in tasks such as text summarization and sentiment analysis.

[0095] B) Multilingual Model: The multilingual model has an efficient language detection algorithm built in. Through the trained multilingual model, the text language of each document's text data is automatically identified, and based on the identification results, the corresponding language processing and algorithm are automatically switched to process the text data. This automatic adaptation mechanism ensures the processing effect of texts in different languages, improving the flexibility of the system and user satisfaction.

[0096] C) Cross-language consistency evaluation unit: In a multilingual environment, it is crucial to ensure the consistency of text data between different languages. By comparing the summary information obtained from the same text data in different languages, the summary information is optimized to ensure consistency and accuracy. This not only involves direct translation comparison between languages, but also includes consideration of cultural differences and contextual understanding to improve the overall reliability of cross-language summaries. The details of the evaluation mechanism are as follows:

[0097] i) Multilingual Translation Engine Integration Unit: This is used to integrate the industry’s leading multilingual translation engines to ensure that text data can be accurately converted into the target language as the basis for evaluation. These engines are continuously optimized to handle complex language phenomena and provide high-quality translation comparisons for evaluation.

[0098] ii) Cultural sensitivity analysis unit: It is used to identify and compare specific expressions, metaphors or slangs in different language versions of text data that may arise due to differences in cultural backgrounds based on the pre-established cultural difference database.

[0099] iii) Contextual Understanding Enhancement Unit: It is used to evaluate the applicability of text data in different contexts through natural language processing (NLP) technology, especially the context-aware model Qwen2. The contextual understanding enhancement unit can analyze the context of the text data, including topics, emotional colors, social occasions, etc., to help evaluate whether the text data maintains the same contextual meaning and emotional tendency in different language versions.

[0100] iv) Manual review and feedback loop unit: The manual review team is responsible for reviewing the automated evaluation results, paying special attention to subtle cultural differences and contextual changes that are difficult for automated tools to capture. The feedback from manual review is incorporated into the continuous optimization of the evaluation system, forming a closed-loop feedback mechanism to continuously improve the accuracy and efficiency of the evaluation. The manual review team is composed of experts with multilingual capabilities and deep cultural background knowledge.

[0101] D) Language resource library construction module: used to build a rich language resource library, including vocabulary, grammar rules, language models, etc. These resources provide the system with necessary language knowledge, enabling the system to better understand and generate texts in various languages.

[0102] (6) Summary quality assessment module: used to evaluate the quality of the generated cross-document summary information and generate feedback information.

[0103] A) Automatic evaluation module: It is used to automatically evaluate the generated target summary based on evaluation indicators such as overlap, semantic similarity, coherence and diversity to generate automatic evaluation feedback, thereby comprehensively measuring the quality of the target summary.

[0104] B) Human evaluation feedback module: It is used to regularly collect human expert evaluation feedback to calibrate the parameters of the automatic evaluation module and ensure the reliability of the evaluation results.

[0105] C) Real-time feedback loop module: Based on automatic evaluation feedback and human expert evaluation feedback, the summary quality is continuously optimized through reinforcement learning or fine-tuning technology.

[0106] (ii) Event extraction module, which is used to extract event information from the generated cross-document summary information through large-scale language models (such as BERT, GPT series, etc.) and supervised sequence labeling models (such as BiLSTM-CRF).

[0107] Specifically, the event extraction module achieves accurate extraction of events in text by combining a large-scale language model and a supervised sequence labeling model.

[0108] The event extraction module includes:

[0109] (1) Model training module, used to train and integrate large-scale language models and supervised sequence labeling models.

[0110] Model training module, including:

[0111] A) Data collection unit, which is used to collect and organize historical cross-document summary information containing rich event information. These cross-document summary information should cover a variety of contexts and event types to ensure the generalization ability of the model. At the same time, the historical cross-document summary information is preprocessed, including word segmentation, part-of-speech tagging, named entity recognition, etc., to provide high-quality input for subsequent model training.

[0112] B) Large language model pre-training unit: The large-scale language model is pre-trained using the collected historical cross-document summary information to fine-tune the large-scale language model parameters so that it can better understand the event structure and contextual information in the cross-document summary information. This step aims to use the general knowledge of the language model to provide preliminary event candidates or event representations for event extraction.

[0113] C) A supervised sequence labeling model pre-training unit is used to construct a supervised sequence labeling model and train the constructed supervised sequence labeling model through the output data in the pre-training process of the large-scale language model.

[0114] Specifically, the supervised sequence labeling model is used to receive data output from large-scale language models (such as event mentions, event attributes, etc.), and learn how to accurately identify and classify events by labeling event boundaries and types in the output data of large-scale language models. During the training process, optimization strategies such as the cross entropy loss function are used to continuously adjust model parameters to improve the accuracy and efficiency of event extraction.

[0115] D) Model fusion and optimization unit: deeply integrate the pre-trained large-scale language model and the supervised sequence labeling model. Specifically, by designing a reasonable fusion mechanism (such as feature splicing, attention mechanism, etc.), the two models can work together to improve the performance of event extraction.

[0116] In addition, the validation set is used to tune the model to ensure that the model has good generalization ability in complex text environments.

[0117] (2) Event extraction module, which is used to extract event information from cross-document summary information through a deeply fused large-scale language model and a supervised sequence labeling model.

[0118] The event extraction module includes:

[0119] A) Text input unit, which is used to input cross-document summary information into the deeply fused large-scale language model and supervised sequence labeling model.

[0120] B) Event candidate generation unit, which is used to perform preliminary processing on cross-document summary information through a pre-trained large-scale language model to generate potential event candidates. These candidates may include key information such as event mentions and event attributes.

[0121] C) Event recognition and classification unit: used to further identify and classify event candidates through the pre-trained supervised sequence labeling model to extract complete event information. By labeling the event boundaries and types in the cross-document summary information, the supervised sequence labeling model can accurately extract complete event information from the cross-document summary information.

[0122] D) Event relationship construction unit: For cross-document summary information, the relationship between event information also needs to be considered. This can be achieved by analyzing the co-occurrence, temporal relationship and other features of event candidates in different documents, thereby building a more complete and accurate event information network.

[0123] (3) An event evaluation module, which is used to evaluate the performance of the deeply fused large-scale language model and supervised sequence labeling model based on the extracted temporal information.

[0124] A) Accuracy evaluation unit, used to calculate the matching degree between the extracted events and the manually labeled events, and evaluate the accuracy of the model. Common evaluation indicators include precision, recall, and F1 value.

[0125] Among them, manually labeled events refer to dividing the test set into a data set, first manually labeling some events as labels of the test set, and then matching the events obtained by the event extraction module with the manually obtained events.

[0126] B) Efficiency evaluation unit, which is used to evaluate the efficiency and speed of the deeply fused large-scale language model and supervised sequence labeling model in processing cross-document summary information, to ensure that they can meet the real-time requirements in practical applications.

[0127] C) Generalization ability evaluation unit: Use datasets that have not participated in the training of large-scale language models and supervised sequence labeling models to test the generalization ability of large-scale language models and supervised sequence labeling models, so as to evaluate the performance of large-scale language models and supervised sequence labeling models in different contexts and event types.

[0128] D) Error Analysis Unit: Conducts in-depth analysis on error events extracted by large-scale language models and supervised sequence labeling models to identify the causes and patterns of errors, and provide guidance for further optimization of large-scale language models and supervised sequence labeling models.

[0129] (3) The cross-document event collection module is used to intelligently identify and merge the extracted event information using deep semantic parsing and vectorized representation technology.

[0130] The cross-document event collection module includes:

[0131] (1) Deep semantic parsing and vectorized representation module: Using text embedding technology such as BGE, deep neural network learning is used to extract contextual information from event information, and complex natural language text is accurately converted into high-dimensional, dense semantic vectors.

[0132] This process not only captures the subtle differences between words, but also digs deep into the logical and semantic connections between sentences, paragraphs and even entire documents, laying a solid foundation for subsequent entity grouping and integration.

[0133] (2) Intelligent similarity evaluation and efficient grouping module: Based on the transformed semantic vector, the cosine similarity calculation method and the advanced clustering algorithm DBSCAN are used to intelligently identify and efficiently merge event information to generate merged event information.

[0134] Among them, intelligent identification and efficient merging can automatically detect and merge highly similar entity descriptions in events, effectively avoiding information redundancy while retaining the diversity and integrity of event information. By dynamically adjusting clustering parameters and similarity thresholds, the accuracy and flexibility of clustering results are ensured, providing users with a clearer and more orderly entity grouping view.

[0135] The intelligent similarity evaluation and efficient grouping module includes:

[0136] A) Semantic vector construction deepening unit: It uses the BGE-Embedding model to extract semantic features from events and generate high-dimensional semantic vectors.

[0137] Specifically, the BGE-Embedding model can capture subtle semantic differences between events by training on a large-scale corpus, providing a solid foundation for subsequent similarity calculations.

[0138] B) Cosine similarity calculation module: The weighted cosine similarity calculation method is used to assign different weights to features of different dimensions according to their importance. This optimization measure not only improves the accuracy of similarity evaluation, but also enhances the model's robustness to noise. Although traditional cosine similarity calculation can effectively measure the directional similarity between vectors, it may face problems such as high computational complexity and sensitivity to noise when processing high-dimensional semantic vectors.

[0139] C) Customized application unit of clustering algorithm DBSCAN: used to customize the application of clustering algorithm DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm. It is a density-based clustering algorithm that divides data points into core points, boundary points and noise points.

[0140] Specifically, the clustering algorithm DBSCAN can automatically identify and exclude noise points, and find clustering structures of any shape, which is very suitable for processing data sets with complex distribution characteristics. For specific application scenarios, we customized the clustering algorithm DBSCAN, including dynamically adjusting two key parameters: neighborhood radius (ε) and minimum number of points (MinPts). By monitoring the clustering effect in real time and adjusting these parameters based on feedback, we ensure the accuracy and flexibility of the clustering results.

[0141] D) Clustering parameter dynamic adjustment unit: used to dynamically adjust clustering parameters.

[0142] Specifically, in order to ensure that the clustering results can accurately reflect the intrinsic structure of the data, a dynamic adjustment strategy of clustering parameters based on the evaluation indicators of clustering results (silhouette coefficient, Calinski-Harabasz index, etc.) is used to automatically determine whether the current parameter settings are appropriate, and fine-tune them accordingly.

[0143] In addition, through the user feedback mechanism, users can fine-tune the clustering results according to their own needs, further improving the flexibility of the system and user experience.

[0144] (3) Cross-document event fusion and consistency maintenance module, which is used to fuse and maintain the merged event information to obtain the target event information.

[0145] Specifically, entity recognition, linking and matching technologies are used to intelligently identify and merge entities with the same or highly similar names from different documents. By building a global entity index and consistency check mechanism, the uniqueness and consistency of entity identity are ensured throughout the entire event collection and integration process.

[0146] Through sophisticated deduplication strategies, the quality and purity of the data are further improved, providing users with a more reliable event information foundation.

[0147] (4) Multi-dimensional event verification and information enhancement module, which is used to verify and enhance the target event information to obtain the event picture.

[0148] Specifically, in order to ensure the accuracy and authority of the target event information, a comprehensive entity verification process is carried out.

[0149] This process integrates external authoritative knowledge bases (such as Wikipedia, DBpedia, etc.), expert systems, and advanced automated context analysis technology. By performing multiple checks and comparisons on the identified entities, the misidentification rate is effectively reduced and the credibility of the information is improved.

[0150] At the same time, based on context analysis technology, it can automatically supplement or correct key attribute information of entities, such as time, location, participants, etc., making the event description more vivid, detailed and accurate. This function not only enhances the readability and value of information, but also provides users with a deeper insight into events.

[0151] (IV) Performance evaluation and optimization module: continuously optimizes the system through data feedback and algorithm iteration.

[0152] The performance evaluation and optimization module includes:

[0153] (1) Multi-dimensional performance indicator collection module: collects system performance indicators such as accuracy, recall rate, F1 score, robustness, scalability, precision, processing time, and resource consumption.

[0154] These indicators can help us fully understand the performance of the system in different aspects, so as to perform more detailed optimization.

[0155] (2) Automated hyperparameter adjustment module, which is used to apply machine learning-based hyperparameter optimization techniques to find the optimal configuration of the system.

[0156] Specifically, machine learning-based hyperparameter optimization techniques, especially Bayesian Optimization, were applied to find the optimal system configuration in a huge parameter space.

[0157] Specifically, the best system configuration can be found by:

[0158] In the context of intelligent search algorithms, grid search follows a formalized strategy, defines a grid in the parameter space, and searches for the optimal configuration by traversing all grid points. Its formal expression can be simplified to the parameter vector Each dimension of is discretized to form a Cartesian product space .

[0159] Random Search randomly selects parameter points, and its process can be regarded as In order to overcome the problem that the efficiency of grid search and random search decreases sharply when the dimension of parameter space increases.

[0160] Genetic Algorithms (GA) and Simulated Annealing (SA), GA iteratively improves the parameter solution set by simulating the natural selection process (such as selection, crossover, mutation and other operations). The core is the fitness function (usually related to model performance), which is used to evaluate the quality of each solution. SA simulates the energy reduction of the system during the physical annealing process, and accepts non-optimal solutions in a probabilistic manner, so that there is a chance to jump out of the local optimum and reach the global optimum.

[0161] The core of Bayesian optimization is to build a Gaussian Process (GP) model. ,in, is the mean function, is a covariance function (also called a kernel function) that describes the similarity between different hyperparameter configurations. Bayesian optimization maximizes an acquisition function (such as EI-Expected Improvement, PI-Probability of Improvement, etc.) to select the next evaluation point. The acquisition function combines the balance between exploration and utilization. Its formal expression, such as EI, can be written as:

[0162]

[0163] in, is an existing observation dataset, This is the maximum model performance currently known.

[0164] As the iteration proceeds, new observation points are added , To the dataset , the GP model is updated to more accurately predict the model performance in the unknown area and guide the search process to approach the optimal solution. This feature of Bayesian optimization enables it to find hyperparameter configurations close to the global optimal in fewer evaluations.

[0165] (3) Real-time monitoring and feedback module: used to monitor performance indicators and system health status, such as memory usage, CPU load, etc., to ensure stable operation of the system.

[0166] (4) Integrated A / B testing module: used to dynamically adjust the traffic distribution between the experimental group and the control group through the multi-armed bandit algorithm in A / B testing to maximize learning efficiency.

[0167] (5) Model version control and rollback module: used to implement system version control and record detailed information of each update so that it can be quickly rolled back to the previous stable version when problems occur in the new version.

[0168] (6) Cross-platform performance evaluation module: Considering that different platforms may have different impacts on system performance, a cross-platform performance evaluation should be conducted to ensure that the system can maintain optimal performance in various environments.

[0169] (7) Generalization ability test module: By testing the system on different data sets, its generalization ability is evaluated to ensure that the system not only performs well on the training data, but can also adapt to new, unseen data.

[0170] (8) Continuous Integration and Continuous Deployment (CI / CD) module: used to integrate performance evaluation and optimization processes into the CI / CD process to achieve automated testing, evaluation, and deployment, thereby improving development efficiency and system stability.

[0171] Through these optimization measures, the functions of each module in the system are more comprehensive, able to provide more personalized, efficient and accurate services, while enhancing the system's adaptability and user-friendliness.

[0172] Embodiment 2:

[0173] like Figure 2 As shown, this embodiment provides an efficient cross-document information extraction method based on a large language model, comprising the following steps:

[0174] S1: Use a large language model (e.g., Qwen2) to deeply analyze and summarize multiple documents in various formats to obtain summary information for each document.

[0175] S11: Preprocess multiple documents in multiple formats to obtain text data of each document.

[0176] S11-1: Collect and manage metadata of each document in a plurality of documents in a plurality of formats.

[0177] S11-2: Convert each of the plurality of documents in the plurality of formats into text data, and verify consistency of the text data.

[0178] S11-3: Used to apply machine learning algorithms to perform mean analysis and maximum and minimum length statistics on text data to understand text data;

[0179] S11-4: Based on the understood text data, perform a thorough data cleansing on the text data and verify the consistency of the text data.

[0180] S11-5: Storing the cleaned data in a file of a specific format.

[0181] S11-6: Real-time security protection measures during in-depth analysis and summarization of multiple documents in multiple formats.

[0182] S12: Slice the text data of each document into blocks to obtain multiple text blocks of each document.

[0183] S12-1: Identify the text data structure of each document to extract key content;

[0184] S12-2: Based on the extracted key content, the text data is segmented into multiple structured text blocks using a recursive algorithm;

[0185] S12-3: Automatically adjust the size of the text block according to the length and complexity of the text data.

[0186] S12-4: Apply sliding windows on text blocks to obtain broader contextual information.

[0187] S12-5: Dynamically adjust the segmentation strategy based on the acquired context information.

[0188] S12-6: Perform multi-dimensional text analysis on text blocks.

[0189] S13: Extract summary information of each document based on multiple text blocks of each document.

[0190] S13-1: Constructing prompts to guide large language models to focus on key information in the text.

[0191] Specifically, we construct prompts similar to “help me extract summary information from text” to guide the large language model to focus on the key information in the text. These prompts are customized according to the text content and context to ensure that the model can capture the most relevant summary information.

[0192] S13-2: Understand the contextual information obtained to ensure the coherence and relevance of the summary information and avoid information gaps or omissions.

[0193] S13-3: Through sophisticated semantic analysis and pattern recognition techniques, each text block of each document is analyzed to identify key information such as topic sentences, important data, main ideas, etc., and a summary is constructed based on this.

[0194] S13-4: Generate a preliminary summary of each document based on the identified key information.

[0195] In the process of generating a preliminary summary of each document, the main ideas and structure of the original text are retained as much as possible, while redundant and secondary information is removed.

[0196] S13-5: Further optimize the generated preliminary summary to improve summary coherence to generate an optimized summary.

[0197] Specifically, the generated preliminary summary is optimized by adjusting the sentence order, adding transition words or phrases, etc. to ensure the logical fluency of the summary.

[0198] S13-6: Finding a suitable balance point between the length and information density of the optimized summary, and extracting summary information of each text block based on the found balance point.

[0199] The purpose of finding the right balance is to ensure that the summary is concise while also ensuring its completeness, that is, not missing any key information.

[0200] The step S13 further includes:

[0201] The text data of each document is translated into different languages ​​to obtain summary information of each document in different language environments, and is evaluated to optimize the summary information based on the evaluation result.

[0202] S14: Fusing the extracted summary information of each document to generate cross-document summary information.

[0203] S15: Perform quality assessment on the generated cross-document summary information and generate feedback information.

[0204] S2: Extract event information from the generated cross-document summary information through a large-scale language model and a supervised sequence labeling model.

[0205] S21: Training and fusing large-scale language models and supervised sequence labeling models.

[0206] S21-1: Collect and organize historical cross-document summary information containing rich event information.

[0207] S21-2: Pre-training large-scale language models using collected historical cross-document summary information.

[0208] S21-3: Construct a supervised sequence labeling model, and train the constructed supervised sequence labeling model using the output data during the pre-training process of the large-scale language model.

[0209] S21-4: Deeply integrate the pre-trained large-scale language model and the supervised sequence labeling model.

[0210] S22: Extract event information from cross-document summary information by integrating a large-scale language model and a supervised sequence labeling model.

[0211] S22-1: Input cross-document summary information into the fused large-scale language model and supervised sequence labeling model.

[0212] S22-2: Preliminary processing of cross-document summary information through a pre-trained large-scale language model to generate potential event candidates.

[0213] S22-3: Further identify and classify event candidates through a pre-trained supervised sequence labeling model to extract complete event information.

[0214] S22-4: Build an event information network.

[0215] S23: Evaluate the performance of the fused large-scale language model and supervised sequence labeling model based on the extracted temporal information.

[0216] S3: Based on the extracted event information, deep semantic analysis and vectorized representation are performed to obtain the event picture.

[0217] S31: Using text embedding technology, deep neural networks are used to learn the contextual information in the extracted event information to generate high-dimensional and dense semantic vectors.

[0218] S32: Based on semantic vectors, the extracted event information is intelligently identified and efficiently merged through cosine similarity calculation method and advanced clustering algorithm to merge event information.

[0219] S33: used to merge and maintain the merged events to obtain target event information;

[0220] S34: Verify and enhance the target event information to obtain an event picture.

[0221] S4: Continuously optimize the system through data feedback and algorithm iteration.

[0222] S41: Collect system performance indicators to understand system performance.

[0223] S42: Apply machine learning-based hyperparameter optimization techniques to find the optimal configuration of the system.

[0224] Embodiment 3:

[0225] This embodiment provides a computer-readable storage medium, which includes a stored program, wherein when the program is running, the device where the computer-readable storage medium is located is controlled to execute the efficient cross-document information extraction method based on a large language model described in Example 2.

[0226] In this specification, the same or similar parts between the various embodiments can be referred to each other. In particular, for the terminal embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the description in the method embodiment.

[0227] In the several embodiments provided by the present invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of systems or units, which can be electrical, mechanical or other forms.

[0228] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0229] In addition, it should be noted that the flowchart in the accompanying drawings shows the method of the embodiment of the present disclosure. In the corresponding description in the flowchart or block diagram in the accompanying drawings, the operations or steps corresponding to different blocks may also occur in an order different from that disclosed in the description, and sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps can actually be performed substantially in parallel, and they can sometimes be performed in the opposite order, which can depend on the functions involved. Each block in the block diagram and / or flow chart, and the combination of blocks in the block diagram and / or flow chart, can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0230] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An efficient cross-document information extraction system based on a large language model, characterized in that: include: Information compression module, which is used to deeply analyze and summarize multiple documents in various formats through a large language model to obtain cross-document summary information; The event extraction module is used to extract event information from the generated cross-document summary information through a large-scale language model and a supervised sequence labeling model; The cross-document event collection module obtains the event picture through deep semantic analysis and vectorization representation based on the extracted event information; Among them, the information compression module includes: a data preprocessing module, which is used to perform data conversion, data cleaning, feature analysis, metadata management, consistency verification and take security measures on multiple documents in multiple formats.

2. The efficient cross-document information extraction system based on a large language model according to claim 1, characterized in that: The data preprocessing module comprises: A metadata management unit for collecting and managing metadata for each document in a plurality of documents in a plurality of formats; a document format conversion unit, configured to convert each of the plurality of documents in the plurality of formats into text data using a natural language processing tool; A feature analysis unit, used to apply a machine learning algorithm to perform mean analysis and maximum and minimum length statistics on the converted text data to understand the text data; A data cleaning unit, used for performing a thorough data cleaning on the text data based on the understood text data; A data consistency verification unit performs text data consistency verification in the process of converting multiple documents of various formats into text data and in the process of cleaning the stored text data; The Data Security Measures Implementation Unit is used to implement data security measures.

3. The efficient cross-document information extraction system based on a large language model according to claim 2 is characterized in that: The information compression module also includes a text segmentation module, a summary extraction module and an information fusion module; The text segmentation module is used to segment the text data of each document to obtain multiple text blocks of each document; The summary extraction module is used to extract summary information of each document from multiple text blocks of each document; The information fusion module is used to fuse the extracted summary information of each document to generate cross-document summary information.

4. The efficient cross-document information extraction system based on a large language model according to claim 3 is characterized in that: The text segmentation module comprises: An efficient document loading and content extraction unit, used to identify the text data structure of each document to extract key content; A recursive character text segmentation unit, for segmenting text data into a plurality of structured text blocks based on the extracted key content using a recursive algorithm; Dynamic text block size adjustment unit, used to automatically adjust the size of the text block according to the length and complexity of the text data; A sliding window technology application unit, used for applying a sliding window on a text block to obtain context information of the text block; The adaptive text segmentation unit is used to dynamically adjust the segmentation strategy based on the context information of the obtained text block to obtain the semantic density and structural features.

5. The efficient cross-document information extraction system based on a large language model according to claim 2, characterized in that: The information compression module also includes: a multi-language support module and a summary quality assessment module; The multi-language support module is used to evaluate the summary information obtained from the text data of each document in different language environments, and optimize the summary information based on the evaluation results; The summary quality assessment module is used to perform quality assessment on the generated cross-document summary information and generate feedback information.

6. The efficient cross-document information extraction system based on a large language model according to claim 1, characterized in that: The event extraction module includes: model training module, event extraction module and event evaluation module; The model training module is used to train and fuse large-scale language models and supervised sequence labeling models; The event extraction module is used to extract event information from cross-document summary information through a fused large-scale language model and a supervised sequence labeling model; The event evaluation module is used to evaluate the performance of the deeply fused large-scale language model and the supervised sequence labeling model based on the extracted time information.

7. The efficient cross-document information extraction system based on a large language model according to claim 1, characterized in that: The cross-document event collection module includes a deep semantic analysis and vectorization representation module, an intelligent similarity evaluation and efficient grouping module, a cross-document event fusion and consistency maintenance module, and a cross-document event fusion and consistency maintenance module; The deep semantic parsing and vectorization representation module is used to use text embedding technology to learn context information in the extracted event information through a deep neural network to convert the event information into a high-dimensional, dense semantic vector; Intelligent similarity evaluation and efficient grouping module, based on the transformed semantic vector, uses the cosine similarity calculation method and clustering algorithm DBSCAN to intelligently identify and merge event information to generate merged event information; The cross-document event fusion and consistency maintenance module is used to fuse and maintain the merged events to obtain the target event information; The multi-dimensional event verification and information enhancement module is used to verify and enhance the target event information to obtain an event picture.

8. An efficient cross-document information extraction method based on a large language model, characterized in that: The following steps are involved: Step 1: Use a large language model to deeply analyze and summarize multiple documents in multiple formats to obtain summary information of each document; the in-depth analysis and summary of multiple documents in multiple formats includes: data conversion, data cleaning, feature analysis, metadata management, consistency verification and security measures for multiple documents in multiple formats; Step 2: Extract event information from the generated cross-document summary information through a large-scale language model and a supervised sequence labeling model; Step 3: Based on the extracted event information, deep semantic analysis and vectorization representation are performed to obtain the event picture; Step 4: Continuously optimize the system through data feedback and algorithm iteration.

9. The efficient cross-document information extraction method based on a large language model according to claim 8, characterized in that: The step one comprises: Preprocess multiple documents in multiple formats to obtain text data for each document; Slice the text data of each document to obtain multiple text blocks of each document; Extract summary information of each document from multiple text blocks of each document; The summary information of each extracted document is fused to generate cross-document summary information.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein when the program is running, the device where the computer-readable storage medium is located is controlled to execute an efficient cross-document information extraction method based on a large language model as described in any one of claims 8 to 9.

Images