Multi-format document intelligent retrieval and semantic association system driven by large model

Abstract

The invention relates to the technical field of artificial intelligence and judicial informatization, and particularly discloses a multi-format document intelligent retrieval and semantic association system driven by a large model. Comprising a multi-modal document intelligent analysis module, an intention-driven semantic retrieval module, a knowledge graph enhanced association recommendation module, a retrieval result visualization and interaction module, a reinforcement learning-driven system optimization module and a multi-format document data storage module. According to the method, the multi-format judicial document is intelligently analyzed through a large model technology; the query intention of the user is accurately understood by means of a semantic retrieval technology Semantic association among the documents is deeply mined through the knowledge graph technology; a retrieval result is visually displayed through a visualization and interaction interface; and the retrieval strategy and model performance are continuously optimized according to user feedback by relying on a reinforcement learning algorithm, so that the retrieval efficiency and semantic association capability of the multi-format document in the judicial field are effectively improved, and the development of judicial informatization is promoted.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and judicial information technology, and in particular to a large-model-driven intelligent retrieval and semantic association system for multi-format documents. Background Technology

[0002] In the fields of smart prisons and judicial information technology construction, the need for intelligent management of heterogeneous documents is becoming increasingly urgent. With the development of business operations, various document formats such as paper records, PDF contracts, and Excel reports are generated. Efficient retrieval and semantic association mining of these documents have become crucial for improving the document management level of the judicial industry.

[0003] Traditional document intelligent retrieval and semantic association systems, in heterogeneous document processing, achieve rapid location of electronic documents through keyword retrieval, promote the digitization of paper documents through OCR technology, and initially explore document associations through manual annotation combined with rule matching.

[0004] However, traditional document intelligent retrieval and semantic association systems have many shortcomings, hindering the intelligent upgrade of document management in the judicial industry. First, semantic understanding is insufficient. Traditional keyword matching technology struggles to parse the deep semantics of multi-format documents (paper scans, PDFs, Word documents, etc.). Faced with complex document content, it cannot accurately capture the meaning, resulting in a high false negative rate and failing to meet the needs of precise retrieval. Second, structured processing is lacking. OCR tools can only recognize characters and cannot preserve the original document structure (such as table rows and columns, chapter hierarchy), leading to semantic breaks during retrieval and affecting the accuracy and coherence of the results. In addition, knowledge association mining is inefficient. Relying on manual index building, it cannot automatically discover implicit relationships between entities such as "personnel-case-regulation" through semantic reasoning, limiting the efficiency of knowledge reuse, making it difficult to build an efficient knowledge network, and failing to fully explore the value of documents. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a large model-driven intelligent retrieval and semantic association system for multi-format documents, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a large model-driven intelligent retrieval and semantic association system for multi-format documents, including a document type identification unit, a paper document processing unit, an electronic document processing unit, and a unified format conversion unit; Multimodal document intelligent parsing module: It is used to receive multi-format document input, parse the multi-format documents through the LayoutLM model, and generate a unified intermediate format document containing text content, logical structure and position information; The intent-driven semantic retrieval module is used to convert documents in a unified intermediate format into high-dimensional vectors through a large model encoder and store them in a document vector database. It receives natural language queries from users, identifies the type of demand through an intent classifier, generates user query vectors, and matches user query vectors with the document vector database to generate preliminary search results. The knowledge graph-enhanced association recommendation module is used to extract key entities from the initial search results, query related nodes in the pre-built judicial domain knowledge graph, calculate the entity association weight coefficient, filter related documents with entity association weight coefficients greater than the preset association weight coefficient threshold, and generate a recommendation list of related documents. The search results visualization and interaction module is used to display the recommended list of related documents in a graphical format, mark key entities and the relationships between entities, and provide a user feedback interface. The reinforcement learning-driven system optimization module is used to collect user feedback data in the results visualization and interaction module, and dynamically adjust the parameters of the large model and the retrieval strategy through reinforcement learning algorithms. Multi-format document data storage module: used to store unified intermediate format documents, high-dimensional vectors, legal knowledge graphs, and user feedback data in a large model database.

[0007] Preferably, the multimodal document intelligent parsing module includes a document type recognition unit, a paper document processing unit, an electronic document processing unit, and a unified format conversion unit; The document type recognition unit is used to receive multi-format document input and automatically recognize the document type; The paper document processing unit is used to preprocess the paper scanned document identified by the document type recognition unit. The preprocessing includes noise reduction and tilt correction, and the text, table area and seal position are identified by the LayoutLM model to generate a JSON file with structure tags. The electronic document processing unit is used to extract text and formatting marks from the electronic document identified by the document type recognition unit; The unified format conversion unit is used to convert the JSON file and the text and format tags into a unified intermediate format document containing text content T, logical structure S, and position information P. .

[0008] Preferably, the execution steps of the intent-driven semantic retrieval module are as follows: S11: Obtain the generated unified intermediate format document from the multimodal document intelligent parsing module. The unified intermediate format document D is split into n text blocks according to its logical structure. The resulting unified intermediate format document ; Through large model semantic encoder Convert each text block into a high-dimensional vector. ,in, The vector representing the i-th text block. This represents a large model semantic encoder. This represents the encoder parameters, i = 1, 2, 3, ..., n, where i represents the number of each text block; For each text block vector Weighted average to generate document vectors The generated document vector Vd is stored in the document vector database, and an efficient retrieval index is built. S12: Receive user input natural language query and perform basic text preprocessing, including word segmentation and standardization, to generate preprocessed user input natural language query text Q; Through large model semantic encoder The user-input natural language query text Q is converted into a high-dimensional vector, and the user query vector is generated by identifying the demand type through an intent classifier. ,in, Represents the user query vector; S13: Calculate document vectors using cosine similarity. and user query vector The similarity is used to match the user query vector with the document vector database based on cosine similarity, generating preliminary search results.

[0009] Preferably, the specific details of matching the user query vector with the document vector database are as follows: For each document vector in the document vector database, the cosine similarity algorithm is used to calculate its similarity to the user query vector. The similarity is calculated as follows: First, the calculated similarity is compared with a preset similarity threshold, and documents with similarity greater than the preset similarity threshold are selected. Then, the selected documents are sorted in descending order of similarity. Finally, the top m documents with the highest similarity after descending order are selected as the preliminary search results and returned.

[0010] Preferably, the specific content of extracting key entities from the preliminary search results is as follows: The document set d receiving the initial search results is used to perform word segmentation and part-of-speech tagging on each document, and a feature vector matrix F is constructed with dimensions of . N represents the total number of candidate entities, K represents the total number of feature dimensions, and each row of the matrix represents a candidate entity and each column represents a feature dimension. Calculate the importance score S for each candidate entity, rank the candidate entities according to the importance score S, and select the top A entities with the highest scores as the key entity set E. For each entity e in the key entity set E, calculate its weight coefficient in the document set d. .

[0011] Preferably, the specific content of querying related nodes in the pre-constructed judicial domain knowledge graph is as follows: Construct a knowledge graph G in the judicial field. For each entity e in the key entity set E, locate the corresponding node v(e) in the knowledge graph G. If entity e cannot be accurately matched in the knowledge graph G, a fuzzy matching strategy is adopted. For each successfully matched node v(e), perform graph traversal. The fuzzy matching strategy is as follows: calculate the semantic similarity between entity e and all nodes in knowledge graph G, select the node with the highest similarity and greater than the preset threshold as v(e), and if the similarity of all nodes is less than the preset threshold, create a new node v(e) and add it to knowledge graph G. The execution of the graph traversal involves: setting a maximum traversal depth H and a maximum number of nodes M; starting from node v(e), traversing the knowledge graph G using a breadth-first strategy; recording all node sets N(e) visited during the traversal, with a constraint |N(e)|≤M; and merging the node sets of all key entities to obtain the set of associated nodes. .

[0012] Preferably, the calculation of the entity association weight coefficient is as follows: Map each entity e to a d-dimensional vector v(e). For any two entities... and entity , computational entity and entity direct semantic similarity By constructing a path enhancement factor P, the entity is calculated. To the entity Path strength ; Calculate the entity association weight coefficient, filter related documents whose entity association weight coefficient is greater than the preset association weight coefficient threshold, and generate a recommended list of related documents.

[0013] Preferably, the specific content of the generated list of recommended related documents is as follows: For the user query document q and the candidate related document set B in the preliminary search results, calculate the association weight coefficient R(q,b) between each candidate document b∈B and the query document q, construct the candidate document association weight coefficient matrix M, and filter all candidate documents that satisfy the condition M[d]>T according to the preset association weight coefficient threshold T to form the related document set L. Sort the documents in the related document set L in descending order of association weight coefficient to generate a recommendation list of related documents, where M[d] represents the association weight coefficient matrix between the candidate document b and the query document q.

[0014] As described above, the large-model-driven intelligent retrieval and semantic association system for multi-format documents provided by this invention has at least the following beneficial effects: This invention provides a large-model-driven intelligent retrieval and semantic association system for multi-format documents. It receives and parses multi-format documents to generate a unified intermediate format, achieving standardized processing of document information. A large-model encoder converts the documents into high-dimensional vectors, and combines this with user query intent for precise matching, generating preliminary search results. Based on these preliminary results, it mines related information from a legal knowledge graph to generate a recommended list of related documents, enhancing the richness and relevance of the search results. The recommended list is then displayed graphically, intuitively presenting key entities and their relationships, and providing a user feedback interface to enhance the user experience. The large-model parameters and retrieval strategies are dynamically adjusted based on user feedback data, enabling continuous system optimization. Finally, the unified intermediate format documents, high-dimensional vectors, the legal knowledge graph, and user feedback data are stored in a large-model database, providing comprehensive data support for the system. This invention, through the deep integration of artificial intelligence large-model technology and legal expertise, achieves efficient intelligent parsing of multi-format documents, precise semantic retrieval, knowledge graph-based recommendation, and continuous system optimization based on user feedback. This invention can automatically process judicial documents of different formats, extract key information and construct a unified intermediate format, and then use a large model for semantic understanding and vector representation. Combining intent classification and knowledge graph technology, it provides users with highly relevant and insightful search results, enhances user experience through a visual interface and interactive functions, and continuously optimizes search strategies and model performance using reinforcement learning algorithms, providing strong technical support for judicial informatization. Attached Figure Description

[0015] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the structure of the large-model-driven intelligent retrieval and semantic association system for multi-format documents of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 Please see Figure 1 As shown, this invention provides a large-model-driven intelligent retrieval and semantic association system for multi-format documents, including a multimodal document intelligent parsing module, an intent-driven semantic retrieval module, a knowledge graph-enhanced association recommendation module, a retrieval result visualization and interaction module, a reinforcement learning-driven system optimization module, and a multi-format document data storage module.

[0019] Multimodal document intelligent parsing module: It is used to receive multi-format document input, parse the multi-format documents through the LayoutLM model, and generate a unified intermediate format document containing text content, logical structure and position information; In this embodiment, it should be specifically noted that the multimodal document intelligent parsing module includes a document type recognition unit, a paper document processing unit, an electronic document processing unit, and a unified format conversion unit; The document type recognition unit is used to receive multi-format document input and automatically recognize the document type; It should be noted that the document type recognition unit supports batch uploading of documents in multiple formats, including but not limited to paper scans, PDFs, Word documents, and Excel files.

[0020] The paper document processing unit is used to preprocess the paper scanned document identified by the document type recognition unit. The preprocessing includes noise reduction and tilt correction, and the text, table area and seal position are identified by the LayoutLM model to generate a JSON file with structure tags. It's important to clarify that JSON (JavaScript Object Notation) files are a text-based, lightweight data-interchange format that organizes and stores data using a specific syntax. JSON files are language-independent; different programming languages ​​can easily parse and generate JSON-formatted data, making it an ideal choice for data exchange between different systems and applications.

[0021] It should be noted that the LayoutLM model is a multimodal pre-trained model specifically designed for document understanding tasks, proposed by Microsoft Research Asia. It integrates text, layout, and image modal information, learns general feature representations of documents through pre-training on large-scale document data, and then fine-tunes it on labeled data for specific tasks to adapt to different document understanding scenarios, such as document classification, information extraction, and table recognition.

[0022] The electronic document processing unit is used to extract text and formatting marks (such as heading levels and table structures) from the electronic document identified by the document type identification unit. The unified format conversion unit is used to convert the JSON file and the text and format tags into a unified intermediate format document containing text content T, logical structure S, and position information P. .

[0023] In this embodiment, it should be specifically explained that the preprocessing of the paper scanned document identified by the document type recognition unit is as follows: The document type recognition unit generates corresponding paper scan images from the scanned paper documents. Gaussian filtering algorithm is applied to remove noise from the paper scan images, and Hough transform detection technology is used to correct the tilt of the paper scan images. An adaptive histogram equalization algorithm is used to improve the clarity of the paper scan images. By preprocessing the paper scans, the OCR accuracy can be improved.

[0024] It should be noted that the multimodal document intelligent parsing module can overcome the limitations of traditional OCR's character-level parsing and achieve integrated modeling of document "text content - logical structure - semantic relationships".

[0025] The intent-driven semantic retrieval module is used to convert documents in a unified intermediate format into high-dimensional vectors through a large model encoder and store them in a document vector database. It receives natural language queries from users, identifies the type of demand through an intent classifier, generates user query vectors, and matches user query vectors with the document vector database to generate preliminary search results. In this embodiment, it should be specifically explained that the execution steps of the intent-driven semantic retrieval module are as follows: S11: Obtain the generated unified intermediate format document from the multimodal document intelligent parsing module. The unified intermediate format document D is split into n text blocks according to its logical structure. The resulting unified intermediate format document ; It should be noted that logical structure refers to the hierarchical organization of a document, including chapter divisions (such as first-level headings and second-level headings), paragraph breaks, and structured content boundaries (such as independent text blocks for tables and clauses).

[0026] Through large model semantic encoder (e.g., RoBERTa) converts each text block into a high-dimensional vector. ,in, The vector representing the i-th text block (dimension d is fixed by the model, such as 768). This represents a large model semantic encoder. This represents the encoder parameters, i = 1, 2, 3, ..., n, where i represents the number of each text block; For each text block vector Weighted average to generate document vectors The generated document vectors Vd are stored in a document vector database (such as Milvus or Pinecone), and an efficient retrieval index (such as the FAISS index or the HNSW index) is built. The document vector The expression is: ,in, , This indicates the weight of each text block.

[0027] It should be noted that the weight of each text block is determined by the logical structure S of each text block (e.g., the title has a higher weight). This embodiment does not impose specific limitations, and those skilled in the art can freely set it according to the logical structure.

[0028] S12: Receive user input natural language query and perform basic text preprocessing, including word segmentation and standardization, to generate preprocessed user input natural language query text Q; Through large model semantic encoder The user's input natural language query text Q is converted into a high-dimensional vector, and the intent classifier is used to identify the type of request (such as "case search" or "regulatory query") to generate the user query vector. ,in, Represents the user query vector; S13: Calculate document vectors using cosine similarity. and user query vector Based on the similarity of cosine similarity, the user query vector is matched with the document vector database to generate preliminary search results; The formula for calculating the cosine similarity is: ,in, This represents the vector dot product, which is the document vector... and user query vector Multiply the values ​​of corresponding dimensions and then add them together. Document vector and user query vector The product of moduli.

[0029] In this embodiment, it should be specifically explained that the specific content of matching the user query vector with the document vector database is as follows: For each document vector in the document vector database, the cosine similarity algorithm is used to calculate its similarity to the user query vector. The similarity is calculated as follows: First, the calculated similarity is compared with a preset similarity threshold, and documents with similarity greater than the preset similarity threshold are selected. Then, the selected documents are sorted in descending order of similarity. Finally, the top m documents with the highest similarity after descending order are selected as the preliminary search results and returned.

[0030] It should be noted that, in a specific embodiment, the first m documents are the top 10 documents with the highest similarity.

[0031] It should be noted that the document vector and the user query vector are transformed using the same encoder, and the document vector and the user query vector belong to the same dimensional vector space.

[0032] The knowledge graph-enhanced association recommendation module is used to extract key entities from the initial search results, query related nodes in the pre-built judicial domain knowledge graph, calculate the entity association weight coefficient, filter related documents with entity association weight coefficients greater than the preset association weight coefficient threshold, and generate a recommendation list of related documents. In this embodiment, it should be specifically explained that the specific content of extracting key entities from the preliminary search results is as follows: The document set d receiving the initial search results is used to perform word segmentation and part-of-speech tagging on each document, and a feature vector matrix F is constructed with dimensions of . N represents the total number of candidate entities, K represents the total number of feature dimensions, each row of the matrix represents a candidate entity, and each column represents a feature dimension, which includes word frequency features, position features, length features, and domain features; The term frequency feature refers to the normalized frequency of an entity's appearance in a document; the position feature refers to the paragraph position of an entity in a document (such as the hierarchical encoding of the beginning, body, and end); the length feature refers to the character length of the entity name (adapting to the difference between "long legal citations" and "short party names"); and the domain feature refers to the judicial domain label to which the entity belongs (such as the independent encoding of "contract disputes" and "criminal defense").

[0033] Calculate the importance score S for each candidate entity using the formula: S = W·F, where W is the feature weight vector. Sort the candidate entities according to their importance scores S, and select the top A entities with the highest scores as the key entity set E. For each entity e in the key entity set E, calculate its weight coefficient in the document set d. ; The key entities include, but are not limited to, accident names and legal provisions.

[0034] The formula for calculating the weighting coefficient is as follows: ,in, This represents the weight coefficient of entity e in document d. This represents the importance score of entity e in document d. This represents the sum of importance scores for all key entities in document d, used for normalization. This represents the set of key entities in the document collection d.

[0035] It should be noted that, to provide a specific example, we assume that the document set d receiving the preliminary search results is "Complaint Regarding a Housing Lease Contract Dispute in 2024", and the key entity set is... Among them, entities =House rental contract, physical =Rent arrears, physical store =Contract termination; If entity The importance score in document d is 0.15, entity The importance score in document d is 0.09, entity If the importance score in document d is 0.12, then the sum of the importance scores of all key entities in document d is... Calculate the entity The weight coefficient in document d is approximately 0.333, calculated using the following formula: .

[0036] In this embodiment, it should be specifically explained that the specific content of querying related nodes in the pre-constructed judicial domain knowledge graph is as follows: Construct a knowledge graph G in the judicial field, where nodes represent legal entities and edges represent relationships between entities. For each entity e in the key entity set E, locate the corresponding node v(e) in the knowledge graph G. If entity e cannot be accurately matched in the knowledge graph G, a fuzzy matching strategy is adopted. For each successfully matched node v(e), perform graph traversal. The fuzzy matching strategy is as follows: calculate the semantic similarity between entity e and all nodes in knowledge graph G, select the node with the highest similarity and greater than the preset threshold as v(e), and if the similarity of all nodes is less than the preset threshold, create a new node v(e) and add it to knowledge graph G. The execution of the graph traversal involves: setting a maximum traversal depth H and a maximum number of nodes M; starting from node v(e), traversing the knowledge graph G using a breadth-first strategy; recording all node sets N(e) visited during the traversal, with a constraint |N(e)|≤M; and merging the node sets of all key entities to obtain the set of associated nodes. ; In this embodiment, it should be specifically explained that the calculation of the entity association weight coefficient is as follows: Map each entity e to a d-dimensional vector v(e). For any two entities... and entity , computational entity and entity direct semantic similarity By constructing a path enhancement factor P, the entity is calculated. To the entity Path strength ; The path strength The methods for obtaining it are as follows: If there exists from arrive The path, then , where P represents all possible paths, len(P) is the path length, rel(P) is the path reliability, and β is the path length decay factor; If there is no [from] arrive The path, then ; Calculate the entity association weight coefficient, filter the associated documents whose entity association weight coefficient is greater than the preset association weight coefficient threshold, and generate a recommended list of associated documents; The formula for calculating the entity association weight coefficient is as follows: ,in, Represents the balance parameters. Representing entities and entity The entity association weight coefficient.

[0037] In this embodiment, it should be specifically noted that the specific content of the generated list of recommended related documents is as follows: For the user query document q and the candidate related document set B in the preliminary search results, calculate the association weight coefficient R(q,b) between each candidate document b∈B and the query document q, construct the candidate document association weight coefficient matrix M, and filter all candidate documents that satisfy the condition M[d]>T according to the preset association weight coefficient threshold T to form the related document set L. Sort the documents in the related document set L in descending order of association weight coefficient to generate a recommendation list of related documents, where M[d] represents the association weight coefficient matrix between the candidate document b and the query document q; The formula for the correlation weight coefficient R(q,b) is as follows: ,in, This represents the set of key entities in the query document q. This represents the set of key entities in candidate document b. Representing entities The weight coefficients in the query document q. Representing entities The weighting coefficients in candidate document b. Representing entities and entity The entity association weight coefficient.

[0038] The search results visualization and interaction module is used to display the recommended list of related documents in a graphical format, mark key entities and the relationships between entities, and provide a user feedback interface. The reinforcement learning-driven system optimization module is used to collect user feedback data in the results visualization and interaction module, and dynamically adjust the parameters of the large model and the retrieval strategy through reinforcement learning algorithms. In this embodiment, it should be specifically explained that the execution method of the reinforcement learning-driven system optimization module is as follows: Obtain user feedback data, which includes explicit feedback data and implicit feedback data; wherein, explicit feedback data includes user ratings, tags, and comments on search results; and implicit feedback data includes user click behavior, dwell time, and browsing depth. The acquired user feedback data is preprocessed, including: removing outliers, standardizing numerical ranges, and merging multiple feedbacks from the same user on the same document; Construct a user feedback feature vector by converting the preprocessed user feedback data into a numerical feature vector, which includes click features, time features, rating features, and interaction features; Define the system state space, where the system state is composed of the current retrieval parameter configuration and the user feedback feature vector; Define the system action space, which includes a set of adjustable operations on the large model parameters and retrieval strategies; the large model parameters include: semantic encoding weights, entity recognition thresholds, and association calculation parameters; the retrieval strategies include: query expansion methods, result ranking weights, and filtering conditions. Construct a reward function to calculate the reward value of system actions based on user feedback feature vectors; the reward value calculation formula is: reward value = w1 × click rate + w2 × average dwell time + w3 × explicit rating - w4 × bounce rate, where w1, w2, w3 and w4 are weight coefficients; A reinforcement learning algorithm is used to train the policy network. The input is the system state, and the output is the optimal action selection. Based on the action output by the policy network, the parameters of the large model and the retrieval strategy are adjusted: if the action points to the adjustment of the large model parameters, the corresponding model parameter values ​​are updated; if the action points to the adjustment of the retrieval strategy, the corresponding retrieval strategy configuration is updated. The adjusted parameters and strategies are applied to the system to generate new search results; user feedback data obtained from the new search results is recorded to form a new system state; the system parameters and strategies are continuously optimized until the system performance reaches a stable state or meets the preset optimization goals.

[0039] Multi-format document data storage module: used to store unified intermediate format documents, high-dimensional vectors, legal knowledge graphs, and user feedback data in a large model database.

[0040] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0041] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A large-model-driven intelligent retrieval and semantic association system for multi-format documents, characterized in that: include: Multimodal document intelligent parsing module: It is used to receive multi-format document input, parse the multi-format documents through the LayoutLM model, and generate a unified intermediate format document containing text content, logical structure and position information; Intent-driven semantic retrieval module: It is used to convert documents in a uniform intermediate format into high-dimensional vectors through a large model encoder and store them in a document vector database. It receives natural language queries from users, identifies the type of demand through an intent classifier, generates user query vectors, and matches user query vectors with the document vector database to generate preliminary search results. The knowledge graph-enhanced association recommendation module is used to extract key entities from the initial search results, query related nodes in the pre-built judicial domain knowledge graph, calculate the entity association weight coefficient, filter related documents with entity association weight coefficients greater than the preset association weight coefficient threshold, and generate a recommendation list of related documents. The search results visualization and interaction module is used to display the recommended list of related documents in a graphical format, mark key entities and the relationships between entities, and provide a user feedback interface. The reinforcement learning-driven system optimization module is used to collect user feedback data in the results visualization and interaction module, and dynamically adjust the parameters of the large model and the retrieval strategy through reinforcement learning algorithms. Multi-format document data storage module: used to store unified intermediate format documents, high-dimensional vectors, legal knowledge graphs, and user feedback data in a large model database.

2. The large-model-driven intelligent retrieval and semantic association system for multi-format documents according to claim 1, characterized in that: The multimodal document intelligent parsing module includes a document type recognition unit, a paper document processing unit, an electronic document processing unit, and a unified format conversion unit; The document type recognition unit is used to receive multi-format document input and automatically recognize the document type; The paper document processing unit is used to preprocess the paper scanned document identified by the document type recognition unit. The preprocessing includes noise reduction and tilt correction, and the text, table area and seal position are identified by the LayoutLM model to generate a JSON file with structure tags. The electronic document processing unit is used to extract text and formatting marks from the electronic document identified by the document type recognition unit; The unified format conversion unit is used to convert the JSON file and the text and format tags into a unified intermediate format document containing text content T, logical structure S, and position information P. .

3. The large-model-driven intelligent retrieval and semantic association system for multi-format documents according to claim 1, characterized in that: The execution steps of the intent-driven semantic retrieval module are as follows: S11: Obtain the generated unified intermediate format document from the multimodal document intelligent parsing module. The unified intermediate format document D is split into n text blocks according to its logical structure. The resulting unified intermediate format document ; Through large model semantic encoder Convert each text block into a high-dimensional vector. ,in, The vector representing the i-th text block. This represents a large model semantic encoder. This represents the encoder parameters, i = 1, 2, 3, ..., n, where i represents the number of each text block; For each text block vector Weighted average to generate document vectors The generated document vector Vd is stored in the document vector database, and an efficient retrieval index is built. S12: Receive user input natural language query and perform basic text preprocessing, including word segmentation and standardization, to generate preprocessed user input natural language query text Q; Through large model semantic encoder The user-input natural language query text Q is converted into a high-dimensional vector, and the user query vector is generated by identifying the demand type through an intent classifier. ,in, Represents the user query vector; S13: Calculate document vectors using cosine similarity. and user query vector The similarity is used to match the user query vector with the document vector database based on cosine similarity, generating preliminary search results.

4. The large-model-driven intelligent retrieval and semantic association system for multi-format documents according to claim 3, characterized in that: The specific details of matching user query vectors with the document vector database are as follows: For each document vector in the document vector database, the cosine similarity algorithm is used to calculate its similarity to the user query vector. The similarity is calculated as follows: First, the calculated similarity is compared with a preset similarity threshold, and documents with similarity greater than the preset similarity threshold are selected. Then, the selected documents are sorted in descending order of similarity. Finally, the top m documents with the highest similarity after descending order are selected as the preliminary search results and returned.

5. The large-model-driven intelligent retrieval and semantic association system for multi-format documents according to claim 1, characterized in that: The specific steps for extracting key entities from the preliminary search results are as follows: The document set d receiving the initial search results is used to perform word segmentation and part-of-speech tagging on each document, and a feature vector matrix F is constructed with dimensions of . N represents the total number of candidate entities, K represents the total number of feature dimensions, and each row of the matrix represents a candidate entity and each column represents a feature dimension. Calculate the importance score S for each candidate entity, rank the candidate entities according to the importance score S, and select the top A entities with the highest scores as the key entity set E. For each entity e in the key entity set E, calculate its weight coefficient in the document set d. .

6. The large-model-driven intelligent retrieval and semantic association system for multi-format documents according to claim 1, characterized in that: The specific content of querying related nodes in the pre-constructed judicial domain knowledge graph is as follows: Construct a knowledge graph G in the judicial field. For each entity e in the key entity set E, locate the corresponding node v(e) in the knowledge graph G. If entity e cannot be accurately matched in the knowledge graph G, a fuzzy matching strategy is adopted. For each successfully matched node v(e), perform graph traversal. The fuzzy matching strategy is as follows: calculate the semantic similarity between entity e and all nodes in knowledge graph G, select the node with the highest similarity and greater than the preset threshold as v(e), and if the similarity of all nodes is less than the preset threshold, create a new node v(e) and add it to knowledge graph G. The execution of the graph traversal involves: setting a maximum traversal depth H and a maximum number of nodes M; starting from node v(e), traversing the knowledge graph G using a breadth-first strategy; recording all node sets N(e) visited during the traversal, with a constraint |N(e)|≤M; and merging the node sets of all key entities to obtain the set of associated nodes. .

7. The large-model-driven intelligent retrieval and semantic association system for multi-format documents according to claim 1, characterized in that: The specific calculation of the entity association weight coefficient is as follows: Map each entity e to a d-dimensional vector v(e). For any two entities... and entity , computational entity and entity direct semantic similarity By constructing a path enhancement factor P, the entity is calculated. To the entity Path strength ; Calculate the entity association weight coefficient, filter related documents whose entity association weight coefficient is greater than the preset association weight coefficient threshold, and generate a recommended list of related documents.

8. The large-model-driven intelligent retrieval and semantic association system for multi-format documents according to claim 7, characterized in that: The specific content of the generated list of recommended related documents is as follows: For the user query document q and the candidate related document set B in the preliminary search results, calculate the association weight coefficient R(q,b) between each candidate document b∈B and the query document q, construct the candidate document association weight coefficient matrix M, and filter all candidate documents that satisfy the condition M[d]>T according to the preset association weight coefficient threshold T to form the related document set L. Sort the documents in the related document set L in descending order of association weight coefficient to generate a recommendation list of related documents, where M[d] represents the association weight coefficient matrix between the candidate document b and the query document q.

Images