Multi-agent cooperative judicial text semantic mapping based on model enhancement

Abstract

The invention relates to the field of crossing of artificial intelligence and legal science and technology, and discloses multi-agent cooperative judicial text semantic mapping based on model enhancement, and the method comprises the steps: constructing a data processing pipeline, collecting multi-source heterogeneous legal data by using a dynamic crawler, and cleaning to generate a standardized resource library; constructing a cross-language semantic mapping model based on low-rank adaptive fine tuning, injecting professional legal knowledge and optimizing an attention architecture to support long text processing; deploying a translation agent as a scheduling center, cooperating with the term knowledge base and the evaluation feedback model, executing an iterative closed-loop link including term constraint generation, multi-dimensional evaluation and reflective correction, and generating a target law semantic mapping translation; and constructing a Chinese and foreign laws semantic mapping relationship library to carry out structured association storage. According to the method, the problems that multi-source data is difficult to obtain, professional terms are inaccurate to translate and cross-law semantic alignment is missing are solved, and high-precision semantic mapping and logic integrity of legal texts are realized.

Description

Technical Field

[0001] This invention relates to the intersection of artificial intelligence and legal technology, specifically to model-enhanced multi-agent collaborative semantic mapping of judicial texts. Background Technology

[0002] Currently, with the acceleration of global economic integration, the demand for transnational legal dispute resolution and international legal cooperation is experiencing explosive growth. In-depth research into foreign legal systems and accurate understanding of heterogeneous judicial precedents have become core aspects of comparative law research and international legal practice. Against this backdrop, the connotations of legal concepts differ significantly between different legal systems. Overcoming the dual barriers of language and culture to achieve accurate semantic mapping and knowledge sharing between Chinese and foreign legal texts has become a crucial issue that urgently needs to be addressed in the field of legal technology.

[0003] To address the aforementioned needs for cross-language legal text processing, existing technologies typically employ a model of machine translation systems assisted by human review. Traditional solutions are mostly based on statistical machine translation or general neural machine translation architectures, utilizing bilingual corpora to train sequence-to-sequence models to complete the conversion from source language to target language. In recent years, some technical solutions have begun to explore the introduction of general large language models, leveraging their powerful generative capabilities to directly perform end-to-end translation and interpretation of the input legal text. At the data acquisition level, the conventional approach is to write general web crawler scripts to batch-fetch and store data from publicly available legal databases or judgment document websites based on static pages.

[0004] However, existing technologies still have many shortcomings when dealing with high-precision judicial scenarios. First, modern legal databases widely employ dynamic rendering and anti-crawling strategies, making it difficult for conventional crawlers to capture deep data loaded asynchronously via AJAX. Furthermore, the lack of adaptive waiting and deduplication mechanisms leads to incomplete data collection, frequent IP blocking, and low signal-to-noise ratios in the imported data. Second, legal documents are typically lengthy and logically coherent. General models, limited by fixed context window lengths, often require text truncation. This mechanical segmentation disrupts the integrity of the legal logic chain, and the general framework lacks low-rank fine-tuning for legal terminology, easily leading to semantic illusions when dealing with specific concepts such as res judicata. Third, existing single-generative translation lacks terminology constraints and self-reflection mechanisms, failing to simulate the multi-round review process of experts. This results in frequent inconsistencies in terminology or logical breaks during cross-legal semantic alignment, making it difficult to meet the stringent requirements of judicial practice for rigor.

[0005] Therefore, this invention provides a model-enhanced multi-agent collaborative judicial text semantic mapping to address the shortcomings of existing technologies. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a model-enhanced multi-agent collaborative legal text semantic mapping, which solves the problems of difficulty in acquiring multi-source heterogeneous legal data, inaccurate translation of professional terms, and lack of cross-legal semantic alignment in existing technologies.

[0007] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a model-enhanced multi-agent collaborative semantic mapping method for judicial texts, comprising the following steps: A multi-source heterogeneous legal data processing pipeline is constructed. A dynamic crawler engine is used to collect multi-dimensional raw legal data, which is then cleaned and aligned to generate a standardized legal case resource library. A cross-lingual semantic mapping model based on low-rank adaptive fine-tuning is constructed. A pre-trained large model is selected as the base, and its parameters are fine-tuned using a parallel legal corpus, while legal terminology knowledge is injected to form a generative base with legal semantic understanding capabilities. A translation agent is deployed as the scheduling center. This agent extracts the text to be processed from the standardized legal case resource library, calls the cross-lingual semantic mapping model as the generation core, and coordinates with a professional terminology knowledge base and an evaluation feedback model to execute an iterative closed-loop chain including terminology constraint generation, multi-dimensional evaluation, and reflective correction to generate semantic mapping translations for the target legal system. A semantic mapping relationship library between Chinese and foreign legal systems is constructed, and the semantic mapping translations are structurally associated and stored with the text to be processed. Professional data tables are derived based on field mapping technology.

[0008] By adopting the above technical solutions, this invention solves the standardization problem of heterogeneous data through a data pipeline, establishes the semantic foundation of the legal field using a fine-tuning model, and realizes full-process control from terminology constraints to reflective correction through an iterative closed-loop link constructed by a translation agent, effectively improving the accuracy and logical rigor of cross-legal text translation.

[0009] Preferably, the construction of the multi-source heterogeneous legal data processing pipeline specifically includes: in the data acquisition stage, deploying a dynamic crawler engine incorporating headless browser technology to process dynamically rendered pages by simulating human operation behavior; integrating an adaptive waiting algorithm to adjust the waiting time by dynamically monitoring the loading status of page document object model elements; integrating a request traffic smoothing scheduling model to simulate human reading intervals by scheduling request frequencies; in the data cleaning stage, using regular expressions to clean the pipeline and eliminate format noise in unstructured text; in the data quality control stage, integrating data quality monitoring tools to monitor field missing rates, and using the SimHash algorithm to calculate text fingerprints, and judging and eliminating duplicate cases by calculating Hamming distance.

[0010] By adopting the above technical solutions, the limitations of dynamic webpage rendering and anti-crawler mechanisms can be effectively overcome, ensuring the integrity and stability of data collection, and improving the signal-to-noise ratio of the data entering the database through fingerprint deduplication technology.

[0011] Preferably, the construction of the cross-lingual semantic mapping model based on low-rank adaptive fine-tuning specifically includes: optimizing the architecture of the pre-trained large model by replacing the multi-head attention mechanism with a grouped query attention mechanism to reduce the memory usage of key-value pairs in the inference process; introducing a dual-block attention mechanism by dividing the input sequence into multiple local blocks and combining intra-block attention calculation with cross-block information interaction to support input sequences that exceed the original context window limit of the pre-trained large model; constructing a legal parallel corpus using a hierarchical progressive data strategy, with data sources including written legal codes and bilingual comparative case data; freezing the backbone parameters of the pre-trained large model and injecting a trainable low-rank decomposition matrix into the attention module of the Transformer layer for fine-tuning.

[0012] By adopting the above technical solutions, the inference resource consumption is optimized by using a grouped query attention mechanism, and the context truncation problem caused by long legal documents is solved by using a dual-block attention mechanism, thus ensuring the integrity of the legal logic chain.

[0013] Preferably, the parameter fine-tuning and injection of legal terminology knowledge specifically includes: setting the rank of the low-rank decomposition matrix to a preset low-rank value, and the low-rank decomposition matrix covering the query matrix, key matrix, value matrix, and output projection matrix in the Transformer layer; during the fine-tuning data construction stage, encoding key-value pair information from the legal terminology knowledge base into prompt words, concatenating terminology constraint information at the input end, and forcing the cross-language semantic mapping model to focus on fixed translations of terms during the training stage; implementing a progressive, layered fine-tuning process, and successively training the cross-language semantic mapping model using general legal corpora, parallel corpora of written codes, and complex case data.

[0014] By adopting the above technical solutions, the efficient injection of legal expertise is achieved, which not only avoids model overfitting, but also enhances the model's sensitivity and compliance with specific legal terms through prompt word engineering and hierarchical training.

[0015] Preferably, the translation agent, in collaboration with its components, generates a semantically mapped translation of the target legal system through an iterative closed-loop mechanism comprising seven iterative steps: Original text submission and preprocessing step: the translation agent receives the input text and performs format cleaning and segmentation; Terminology knowledge retrieval and constraint generation step: the translation agent calls the professional terminology knowledge base, identifies legal concepts, extracts standard translations, and generates terminology constraint prompts; First-round enhanced translation step: the translation agent concatenates the original text with the terminology constraint prompts and calls the cross-language semantic mapping model to generate a preliminary translation; Multi-dimensional quality assessment step: the translation agent submits the preliminary translation to the evaluation feedback model, which scores and diagnoses the translation based on evaluation indicators and legal logic rules; Feedback generation and strategy formulation step: the translation agent generates optimization strategy feedback based on the diagnostic results; Second-round refined optimization step: the translation agent reconstructs prompt words based on the optimization strategy feedback, guiding the cross-language semantic mapping model to perform reflective translation correction; Result verification and database entry step: the final optimized translation is formatted and written into the Chinese-foreign legal system semantic mapping relation database.

[0016] By adopting the above technical solutions, a translation and proofreading workflow was constructed. Through multiple rounds of iteration and feedback correction mechanisms, the logical loopholes and terminology inconsistencies that may occur in a single translation were resolved.

[0017] Preferably, the evaluation feedback model adopts a large language model with a high number of parameters. The multi-dimensional quality assessment steps specifically include: diagnosing the translation from three dimensions: terminology accuracy, syntactic and logical integrity, and cross-legal semantic equivalence; using the ROUGE series of indicators as the core evaluation criteria, and quantitatively evaluating the semantic mapping quality by calculating the N-gram grammatical overlap and the longest common subsequence between the preliminary translation and the standard reference translation; the translation agent has self-reflection capabilities, adjusts the sentence structure according to the context logic, and eliminates ambiguity by adding qualifiers. The adoption of the ROUGE series of indicators as the core evaluation criteria specifically includes: calculating the ROUGE-N indicator, which is calculated by taking the number of N-grams shared by the preliminary translation and the standard reference translation as the numerator, and the total number of N-grams in the standard reference translation as the denominator, and dividing the two to obtain the ratio, which is used to measure the coverage of legal terms; and calculating the ROUGE-L indicator, which is calculated based on the length of the longest common subsequence between the preliminary translation and the standard reference translation. First, the ratio of the length of the longest common subsequence to the length of the standard reference translation is calculated as the recall rate; second, the ratio of the length of the longest common subsequence to the length of the preliminary translation is calculated as the precision rate; and finally, the weighted harmonic mean of the recall rate and the precision rate is calculated as the final score.

[0018] By adopting the above technical solution, an objective and quantitative evaluation index system was introduced. Through the calculation of N-gram overlap and longest common subsequence, the quality of the translation in terms of terminology coverage and syntactic structure was accurately measured, providing data support for reflective revision.

[0019] Preferably, the construction of the semantic mapping relationship database for Chinese and foreign legal systems specifically includes: constructing a core semantic mapping data table, wherein the primary key identifier of the data table adopts an integer data type; the data table includes a document type field for identifying document dimensions, a country classification field for identifying legal system background, and a data source field for identifying specific source databases; the data table includes a source language field and a target language field for identifying language correspondence, and a source text field and a translated text field for storing long-form content, wherein both the source text field and the translated text field adopt a text data type; based on the core semantic mapping data table, a case table, a legal provision table, a litigation document table, an international treaty table, a judgment document table, and a judicial opinion table are generated through field mapping technology.

[0020] By adopting the above technical solutions, structured storage and multi-dimensional association of cross-language legal texts have been achieved, supporting efficient retrieval and downstream applications.

[0021] Preferably, the method further includes a traffic-based dynamic scaling and continuous integration mechanism: an inference service framework combined with pagination attention technology is used to manage the key-value pair cache in the video memory, supporting parallel loading of multiple models; an asynchronous message queue is introduced to manage the data interaction between the translation agent and each component; a continuous integration pipeline is designed, which triggers an offline fine-tuning process to generate new low-rank adapter weights when new corpus or terminology database updates are obtained, and the model version is switched through grayscale environment testing and blue-green deployment strategies.

[0022] By adopting the above technical solutions, the system's high-concurrency processing capability and memory utilization have been improved, and the continuous iteration and smooth upgrade of model capabilities have been ensured.

[0023] Secondly, the present invention provides a model-enhanced multi-agent collaborative judicial text semantic mapping system, comprising: The data processing module is configured to build a multi-source heterogeneous legal data processing pipeline, using a dynamic crawler engine to collect legal data and perform cleaning and alignment; the semantic mapping model construction module is configured to build a cross-language semantic mapping model based on low-rank adaptive fine-tuning, fine-tuning the parameters of a pre-trained large model and injecting legal expertise; the translation agent scheduling module is configured to deploy translation agents, coordinating user input, a professional terminology knowledge base, the cross-language semantic mapping model, and an evaluation feedback model, executing an iterative closed-loop chain that includes terminology constraint generation, multi-dimensional evaluation, and reflective correction; and the data storage module is configured to build a semantic mapping relationship library of Chinese and foreign legal systems, storing semantic mapping translations and generating professional data tables.

[0024] By adopting the above technical solutions, data processing, model building, intelligent scheduling and storage management are organically combined in a modular manner, supporting high-quality semantic mapping of cross-language judicial texts.

[0025] This invention provides a model-enhanced multi-agent collaborative semantic mapping of judicial text. It offers the following advantages: 1. This invention improves the quality and efficiency of acquiring multi-source heterogeneous legal data by constructing a data processing pipeline integrating headless browser technology, adaptive waiting algorithms, and SimHash deduplication mechanisms. This technical solution effectively overcomes the limitations of dynamic rendering and anti-crawler mechanisms in overseas legal databases, solving the problem of data collection failures caused by network fluctuations. Simultaneously, through a three-layer cleaning architecture and fingerprint deduplication, it reduces format noise and redundant data in unstructured text, improving the signal-to-noise ratio of the legal corpus and providing a high-quality data foundation for subsequent model training.

[0026] 2. This invention introduces grouped query attention and bi-block attention mechanisms into the pre-trained model, combined with low-rank adaptive fine-tuning technology, to achieve efficient processing of long legal texts and deep transfer of professional semantics. This architectural optimization not only reduces memory usage during inference and improves system throughput, but also breaks through the context window limitation of traditional models, enabling it to process complete long judgments or legal codes at once, avoiding the break in logical chains caused by text truncation, and ensuring the integrity and accuracy of legal semantic understanding.

[0027] 3. This invention addresses the issues of terminological inconsistency and logical illusion in traditional machine translation by deploying a translation agent to execute a seven-step iterative closed-loop process that includes terminological constraints, multi-dimensional evaluation, and reflective correction. This scheme utilizes a multi-agent collaborative mechanism to simulate the translation, review, and correction workflow of human legal experts. By introducing a professional terminology knowledge base and real-time feedback from a high-parameter evaluation model, the model is forced to self-reflect and perform logical correction during the generation process, thereby achieving a high degree of alignment in terminological accuracy and logical rigor between cross-legal texts. Attached Figure Description

[0028] Figure 1 This is a diagram illustrating the overall framework of the cross-language legal knowledge system of Chinese and foreign legal systems in this invention. Figure 2 This is a schematic diagram of the multi-agent collaborative translation workflow based on large model enhancement according to the present invention; Figure 3 This is a schematic diagram of the semantic mapping core entity logic model design of the present invention; Figure 4 This is a schematic diagram illustrating the construction of the extraterritorial legal knowledge base and the derivation of professional data tables according to the present invention. Detailed Implementation

[0029] The technical solutions in 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.

[0030] This embodiment discloses a model-enhanced multi-agent collaborative judicial text semantic mapping system and method, which aims to solve the problems of difficulty in obtaining multi-source heterogeneous legal data, inaccurate translation of professional terms, and lack of semantic alignment across legal systems.

[0031] First refer to Figure 1 The cross-language legal knowledge system for Chinese and foreign legal systems mainly consists of a data processing module on the left (responsible for the collection and cleaning of multi-source heterogeneous data), a core cross-language semantic mapping module in the middle (shown in the dotted box, including translation agent, LoRA fine-tuning, and evaluation feedback mechanism), and a case structured information extraction and visualization application module on the right. These modules work together to ultimately construct the foreign legal knowledge and case database at the bottom. The following sections provide a detailed explanation of each subsystem, focusing on specific technical details.

[0032] like Figure 1As shown on the left, this embodiment first constructs a data processing pipeline consisting of three automated stages: crawling, cleaning, and alignment, aiming to solve the challenges of acquiring and standardizing multi-source, heterogeneous legal data. This data layer covers seven authoritative global legal databases, including the World Institute of Legal Information, the Courts Electronic Records Repository (PACER) of the United States, the Supreme Court of Canada case database, the Court of Justice of the European Union case database (CJEU), the Federal Supreme Court of Germany case database, the Supreme Court of Singapore case database, and the website of the Supreme Court of the United Kingdom. The collected data covers six dimensions: cases, judgments, litigation documents, laws and regulations, international treaties, and judicial opinions, spanning from 1877 to 2025. By the time this embodiment was completed, the system had successfully acquired and processed a resource database containing 52,495 multilingual legal data entries, including high-quality Chinese-English and Chinese-German bilingual data, laying a solid data foundation for subsequent model fine-tuning and semantic mapping.

[0033] In the data acquisition phase, this embodiment designed and deployed a lightweight dynamic crawler engine based on SeleniumWebDriver. Addressing the dynamic rendering characteristics commonly used in the aforementioned legal databases (such as AJAX asynchronous loading and JavaScript dynamic rendering), this engine innovatively introduces headless browser technology to simulate real human operation. To overcome anti-crawler mechanisms, the system integrates an adaptive waiting algorithm and a smooth request traffic scheduling model: the former adjusts the waiting time by dynamically monitoring the loading status of page DOM elements, effectively solving the problem of collection failures caused by network fluctuations; the latter avoids IP blocking risks by intelligently scheduling request frequencies to simulate human reading intervals. Through this mechanism, the system can penetrate the AJAX dynamic loading layer, automatically handle JavaScript rendering logic, and possess a certain ability to bypass CAPTCHA interception, achieving fully automated collection from multiple databases, including the U.S. Electronic Records Database (data from 2016-2022) and the German Federal Supreme Court Case Database (data from 2017-2023).

[0034] In the data cleaning and standardization phase, this embodiment establishes a three-layer data processing architecture. The first layer is the metadata extraction layer, which uses a high-precision combination of XPath and CSS selectors to accurately extract metadata such as the title, case number, date, and court level of the judgment document from the original HTML page. The second layer is the cleaning and filtering layer, which designs a dedicated regular expression cleaning pipeline to eliminate HTML tags, garbled characters, redundant whitespace, and formatting noise in unstructured text, ensuring the purity of the text content. The third layer is the structured storage layer, which designs a standardized legal case resource database schema containing 18 core fields to address the heterogeneity of data from different sources. The system uses a heterogeneous data table structure to achieve independent storage of multi-source cases, but logically achieves unified retrieval through unified field mapping. Finally, all collected data is integrated into a standard format including key attributes such as case ID, document type, country classification, data source, data year, original / translated language, and content.

[0035] In terms of system architecture and performance optimization, this embodiment adopts a lightweight architecture design with a front-end and back-end separation. The front-end is built on Vue.js to create a visual console, while the back-end uses the Flask framework to provide API services and is deployed using Docker containerization technology, achieving resource isolation and optimization. To cope with high-concurrency data collection requirements, the system introduces an asynchronous task queue mechanism to decouple data collection tasks from retrieval services, supporting dynamic scaling of the system. In actual campus server environment testing (configured as regular computing nodes), this architecture maintained system stability with an average response time of 2.4 seconds under the pressure of 300 concurrent requests, effectively reducing the memory consumption of a single-node crawler and improving data throughput.

[0036] Finally, in the data quality control stage, this embodiment introduces a strict integrity check and deduplication mechanism. On the one hand, the Apache Griffin data quality monitoring tool is integrated to monitor the field missing rate of the incoming data in real time, setting a threshold of 0.2%. Once the missing rate exceeds this value, an alarm is triggered and the collection task is rescheduled. On the other hand, to solve the problem of duplicate collection that may exist in multi-source data, the Sim Hash algorithm is used to calculate text similarity. The system calculates the Sim Hash fingerprint of each case content and determines duplication by calculating the Hamming distance; in this embodiment, a Hamming distance less than or equal to 3 is set as a duplicate case, thereby automatically deduplicating it. Through the above-mentioned full-process data governance, a semantic mapping basic library containing more than 20,000 high-quality Chinese and foreign bilingual comparison data was finally built.

[0037] In this embodiment, the model layer, as the core computational unit for cross-language legal text semantic mapping, constructs a composite technical architecture consisting of a pre-trained large model base, low-rank adaptive fine-tuning (LoRA), and an inference acceleration engine. This aims to address the problems of inaccurate translation of legal terminology, loss of logic in long and complex sentences, and difficulties in cross-legal system semantic alignment in general-purpose large models. By constructing a three-level optimization system of LoRA fine-tuning, knowledge injection, and quality assessment, a deep transfer from general language capabilities to legal semantic understanding is achieved. Regarding the selection of the base model, after comparative testing of mBART-50, NLLB-200, and the Qwen series models, Qwen-72B was ultimately selected as the base pre-trained model. To adapt to the characteristics of lengthy and highly logically connected legal texts (such as judgments and legal codes), this embodiment has made targeted architectural optimizations to the basic model: On the one hand, Grouped Query Attention (GQA) is adopted to replace the traditional multi-head attention mechanism, reducing the memory usage of key-value (KV) cache during inference and thus improving inference throughput; on the other hand, Dual Chunk Attention (DCA) is introduced to extend the model's context window length to 131,072 tokens. This improvement enables the model to read and process complete long legal case files or complex transnational contract texts at once, avoiding the contextual semantic breaks caused by truncated text in traditional models and ensuring the integrity of the legal logic chain.

[0038] The data foundation for model fine-tuning comes from a high-quality parallel legal corpus, employing a hierarchical and progressive data strategy. Core data is built upon the German Civil Code (BGB), the French Civil Code, and authoritative case law databases, specifically including 2412 Chinese-German legal provisions, 2345 Chinese-English legal provisions, and over 20,000 high-quality Chinese-English and Chinese-German bilingual data points cleaned from the US Court Electronic Records Database and the European Court of Justice Case Law Database. All training data underwent a rigorous standardization and cleaning process, including removing format noise from unstructured text, standardizing legal citation formats using regular expressions, and using the Sim Hash algorithm for deduplication (Hamming distance ≤ 3 is considered duplicates), ensuring a high signal-to-noise ratio for the input data to the model.

[0039] To retain the general language capabilities of Qwen-72B while incorporating legal expertise, this embodiment employs LoRA (Low-Rank Adaptation) technology for efficient parameter fine-tuning. Specifically, the backbone parameters of the pre-trained model are frozen, and a trainable low-rank factorization matrix is ​​injected only into the attention module of the Transformer layer. Regarding key hyperparameter configuration, the rank (Rank, r) is set to 8. Experimental data shows that this setting yields the optimal ROUGE-LPrecision metric for the legal terminology mapping task, avoiding overfitting while ensuring sufficient parameter space. The weight matrix covers all weight matrices in the Transformer layer, including the query matrix (…). Key matrix ( ), value matrix ( ) and output projection matrix ( The full-coverage strategy ensures that the model comprehensively adjusts the semantic space to adapt to complex concept mappings. The training parameters are set with a learning rate of 5e-5, a batch size of 128, and the AdamW optimizer. The training objective is to enable the model to learn accurate mapping relationships by minimizing the cross-entropy loss of predicting the next token, especially strengthening its sensitivity to specific terms such as ResJudicata (resignative force) and Rechtskraft (legal effect).

[0040] To overcome potential issues of forgetting or terminology inconsistencies that may arise from simple fine-tuning, the model layer also implements knowledge injection and terminology enhancement strategies. During the fine-tuning data construction phase, key-value pair information from the legal terminology knowledge base is encoded into the Prompt, for example, by concatenating [Terminology Constraint] Rechtskraft to legalforce at the input, forcing the model to focus on the fixed translations of specific terms during training. Simultaneously, a progressive, layered fine-tuning process is adopted, using general legal texts, specific legal codes, and complex case judgments. First, general legal corpora are used to adapt the model to legal language; then, highly specialized parallel corpora such as the German Civil Code are used for fine-tuning; finally, complex case data is used to enhance the model's logical reasoning ability.

[0041] After model training, to meet the high-concurrency, low-latency service requirements of judicial scenarios, this embodiment designs a high-performance inference deployment scheme. The inference service is built based on the vLLM framework, utilizing its core Paged Attention technology to manage the KVCache in GPU memory through paging, similar to the virtual memory management mechanism of an operating system, effectively solving the GPU memory fragmentation problem. Compared to the traditional Hugging Face Transformers inference method, this scheme supports continuous batch processing of multiple requests, increasing the throughput of legal text translation by 2-4 times while ensuring semantic coherence. In terms of resource scheduling, the model is deployed in Docker containers, supporting dynamic scaling based on traffic. In actual testing on an NVIDIA Tesla V100 environment, this architecture achieved a response time in seconds when processing legal texts with an average length of 500 tokens, and supports parallel loading of multiple models to handle concurrent requests from different language pairs.

[0042] The fine-tuned model constructed based on the above implementation method was rigorously validated on an independent test set (412 articles in Chinese-German and 345 articles in Chinese-English). Experimental results show that the Qwen-72B fine-tuned model achieved a ROUGE-LF-measure of 0.9273 and a ROUGE-1 Precision of 0.9272 in the German legal text translation task. Even in resource-constrained scenarios, the Qwen-7B fine-tuned model, through LoRA fine-tuning with r=8, also achieved an excellent ROUGE-LF-measure of 0.9563 (German). Qualitative analysis shows that the model successfully corrected the semantic biases of the un-fine-tuned model when handling complex concepts such as contract termination conditions and the reversal of the burden of proof. It can accurately reconstruct long passive German sentences into active voice sentences that conform to English legal conventions, fully validating the effectiveness of the model layer construction scheme.

[0043] In this embodiment, the translation agent is designed as the intelligent scheduling hub of the entire cross-language judicial text semantic mapping system, and its core architecture is built on the FastGPT platform.

[0044] See attached document Figure 2This agent is not merely a simple task executor, but a collaborative controller responsible for coordinating data and control flows among user input, a specialized terminology knowledge base, a finely tuned large language model (Qwen series), and an evaluation feedback model (DeepSeek-70B). Through this centralized scheduling mechanism, the translation agent organically connects disparate technical components, forming a closed-loop chain of retrieval, generation, evaluation, and correction, thereby resolving common issues in traditional single-transaction machine translation such as inconsistencies in terminology, logical breaks, and legal illusions. The workflow is designed to simulate the translation review process of human legal experts through machine intelligence's self-reflection and multiple iterations, ensuring a high degree of legal semantic accuracy and logical rigor in the final output translation.

[0045] Specifically, the workflow of the translation agent is meticulously divided into a seven-step iterative mechanism, covering the entire lifecycle management from data input to final storage.

[0046] The first step is the submission and preprocessing of the original text: The system receives the legal text data to be translated (such as complex judgment paragraphs or specific code provisions) through the front-end standard interface. The Agent first performs format cleaning and segmentation processing on the input text. The second step is terminology knowledge retrieval and constraint generation: Before translation, the Agent will proactively call the legal terminology knowledge base, use semantic retrieval technology to identify key legal concepts (such as Rechtskraft, ResJudicata) from the original text, extract their corresponding standard translations in the target legal system, and generate terminology constraint prompts. The third step is the first round of enhanced translation: The agent concatenates the original legal text with the generated terminology constraint information to construct a composite input instruction that includes contextual background and mandatory terminology mapping. This instruction is then sent to the Qwen-72B model, which has been fine-tuned by LoRA, to perform the first round of automated translation and generate a preliminary translation.

[0047] Subsequently, the workflow enters the critical quality control and optimization phase.

[0048] The fourth step is multi-dimensional quality assessment: The Agent submits the first-round generated translation along with the original text to a large evaluation model with a large number of parameters (DeepSeek-70B is used in this embodiment). The evaluation model strictly scores and diagnoses the translation based on the ROUGE index and preset legal logic rules from three dimensions: terminology accuracy, syntactic logic integrity and cross-legal semantic equivalence. The fifth step is feedback generation and strategy formulation: Based on the diagnostic results of the evaluation model, the Agent automatically generates specific optimization strategy feedback, such as pointing out specific problems such as subject ambiguity caused by passive voice or specific terms not following knowledge base constraints; The sixth step is the second round of fine-tuning: The agent reconstructs the prompt words based on the generated optimization strategy, guiding the fine-tuning model to perform a second round of reflective translation. In this process, the model corrects the flaws in the first round of translation, such as reconstructing the long passive sentences commonly found in German legal texts into active voice that conforms to English legal conventions, while ensuring strict alignment of core terms; The seventh step is result verification and database entry: After two rounds of iteration, the final selected translation is paired with the original text and formatted and written into the semantic mapping relation database, completing a complete semantic mapping task.

[0049] To support the efficient operation of the aforementioned complex workflow, the backend service architecture employs a high-performance inference engine and message queue mechanism. For underlying computing resource management, this embodiment uses the vLLM framework combined with Paged Attention technology to build the inference service. This technology effectively solves the memory fragmentation problem by performing paging management on the key-value cache (KVCache) in GPU memory, similar to operating system virtual memory, supporting parallel loading of multiple models and high-concurrency requests. The translation agent interacts with the fine-tuning model and knowledge base services (both deployed in Docker containers) via a RESTful API. To ensure the reliability and low latency of data transmission between multiple components, the system introduces RabbitMQ as an asynchronous message queue, ensuring that tasks in each step can be scheduled in an orderly manner without blocking in high-concurrency scenarios (such as batch processing tens of thousands of cases). Furthermore, the database layer employs a database sharding strategy combined with Elasticsearch full-text search, ensuring fast access and fuzzy matching capabilities for massive amounts of legal text.

[0050] This translation agent workflow also possesses advanced capabilities for self-reflection and dynamic correction, which is particularly evident when handling complex legal provisions. Taking the translation of Article 2164 of the German Civil Code as an example, traditional fine-tuning models might mechanically translate things into expressions with temporal ambiguity. However, the translation agent, by introducing a reflective mechanism, can identify logical flaws that might result from a literal translation. In the workflow, the agent proactively adjusts sentence structure based on contextual logic (such as the temporal relationship between the testator and the estate), eliminates ambiguity by adding precise qualifiers (such as damage to the subject matter), and utilizes the glossary's binding function to avoid conceptual confusion commonly found in general models (such as distinguishing between Schadensersatz and Entschädigung). This mechanism allows the system to achieve an equivalent transmission of legal intent through contextual logic compensation when perfect corresponding terms are unavailable, effectively controlling the generation of legal illusions and achieving an optimal balance between professionalism and readability.

[0051] Finally, to ensure the continuous evolution of the translation agent's capabilities, this embodiment designs an independent continuous integration pipeline for data acquisition, model retraining, and deployment switching. When the system acquires new high-quality legal corpora or terminology updates through the crawler engine, it automatically triggers an offline fine-tuning process to generate new LoRA adapter weights. The new model version is first A / B tested in a gray-scale environment. By comparing the ROUGE metrics of the old and new models on the test set, the performance improvement is verified. Once verified, the system uses a blue-green deployment strategy to seamlessly switch to the new model version without interrupting online services. This mechanism ensures that the translation agent can continuously iterate itself as legal data accumulates, allowing each submodule to update independently without hindering each other, thus laying a solid foundation for the long-term operation and technological evolution of the cross-language mapping library of Chinese and foreign legal systems.

[0052] In this embodiment, given the high requirements for terminological accuracy and logical integrity in legal text translation, traditional dichotomous accuracy metrics (i.e., judging only correctness or error) are insufficient to quantify the fine-grained matching degree between the translation and the standard answer. Therefore, the system adopts the ROUGE (Recall-Oriented Understudy for Gisting Evaluation) series of metrics as the core evaluation standard. The ROUGE metrics quantify the semantic mapping quality from three dimensions: recall, precision, and F-measure, by calculating the N-gram overlap and longest common subsequence between the legal translation (Hypothesis, hereinafter referred to as the generated translation) generated by the calculation model and the manually verified standard reference translation (hereinafter referred to as the reference translation). Specifically, it includes two calculation methods: ROUGE-N and ROUGE-L.

[0053] ROUGE-N calculation formula and implementation details: The ROUGE-N metric is primarily used to assess the co-occurrence of the generated and reference translations at the N-gram level, where N represents the length of a continuous word sequence (e.g., N=1 represents a unary word, N=2 represents a bigram). This metric effectively measures the coverage of legal terms and phrases.

[0054] In practical calculations, the general formula for ROUGE-N is expressed as follows: ; Among them, the denominator Indicates the total number of N-grams in all reference translations; molecule This indicates the number of N-grams shared by the generated translation and the reference translation (i.e., the number of matched N-grams).

[0055] To more comprehensively evaluate model performance, this embodiment further breaks down ROUGE-N into recall, precision, and F-score for calculation: Recall rate (): This measures how much information in the reference translation is correctly translated by the model. The calculation formula is the number of N-grams shared by the generated translation and the reference translation divided by the total number of N-grams in the reference translation: ; Precision ): This measures the proportion of correctly translated content in the model-generated translation. The calculation formula is the number of N-grams shared by the generated translation and the reference translation divided by the total number of N-grams in the generated translation. ; F-measure, Precision (MR) is the harmonic mean of precision and recall, used to comprehensively reflect translation quality. It is usually taken as... (i.e., F1 value), the calculation formula is as follows: ; ROUGE-L calculation formula and implementation details: The ROUGE-L metric is calculated based on the Longest Common Subsequence (LCS). Unlike N-grams, LCS does not require word continuation, but the original relative order must be maintained. This characteristic makes it particularly suitable for assessing logical coherence and sentence structure in legal texts (e.g., the logical relationship of "if...then..." in legal provisions), and it can capture semantic structural similarity at the sentence level.

[0056] Let the reference translation be Its length is The generated translation is Its length is . Represents a sequence and The length of the longest common subsequence. The specific calculation of ROUGE-L includes recall. Accuracy and F value The formula is as follows: Recall rate The length of the longest common subsequence is determined by the ratio of the length of the reference translation. ; Accuracy The length of the longest common subsequence is determined by the ratio of the length of the generated translation. ; F value The weighted harmonic mean is also used for calculation. In the experimental setup of this embodiment, in order to balance the impact of recall and precision, it is usually set to... Preset hyperparameters (such as The final ROUGE-L score is... : ; Using the formula described above, this embodiment can calculate the similarity score (ranging from 0 to 1) between the generated legal translation and the standard translation in terms of lexical overlap and structural order. A higher score indicates more accurate semantic mapping. In actual testing, the focus is on ROUGE-1 (measuring lexical and terminological coverage) and ROUGE-L (measuring syntactic structure and logical coherence), serving as the technical basis for determining whether the model achieves an accuracy exceeding 75% and for verifying semantic equivalence.

[0057] This embodiment employs a hybrid storage architecture combining relational databases and a full-text search engine, aiming to balance the structured management of legal text metadata with the efficient retrieval needs of massive amounts of unstructured text. The database layer not only carries the raw data collected from seven authoritative extraterritorial legal databases but also serves as the final destination for multi-agent collaborative translation workflows, storing semantically aligned mappings between Chinese and foreign legal systems.

[0058] See attached document Figure 3To support the construction of a semantic mapping relation database for Chinese and foreign legal systems, the system has built a logical model with semantic mapping as the core entity. This model aims to comprehensively characterize the correspondence between cross-language legal texts through ten key attributes. Logically, the core data table is named `semantic_mapping`, which adopts a highly standardized field structure. The primary key identifier `Id` uses the INT data type as the unique identifier for the entity. It leverages the database's auto-incrementing feature to support expansion to millions of records while ensuring a compact 4-byte storage space. This facilitates the construction of an efficient B+ tree primary key index, ensuring the uniqueness and traceability of data under massive storage conditions. In terms of document and data classification, the instrument_type field is defined as VARCHAR(50) to distinguish six document dimensions, including case documents, litigation documents, judgment documents, legal treaties, and judicial opinions. This length setting can accommodate Chinese descriptions while avoiding the space waste that may be caused by fixed-length CHAR types. The data_type field, as an auxiliary field, further refines the specific form of the data and supports precise retrieval based on document type. The country_category field is defined as VARCHAR(50) to identify the legal system background to which the legal text belongs, such as the civil law system, the common law system, and the specific country to which it belongs. In terms of data traceability and time dimension, the data_source field is defined as VARCHAR(100). Considering that the names of institutions such as the World Institute of Legal Information and the case database of the Federal Supreme Court of Germany are relatively long, this length is sufficient to completely record the specific source database of the data to ensure the traceability of the data. The data_year field adopts the INT type, which not only saves storage space but also facilitates numerical comparison queries based on time ranges to support the study of the historical evolution of different legal systems. As the core value carrier of the entire database, the language and content mapping part includes the source_language and target_language fields, both defined as VARCHAR(20), used to identify the language correspondence between the original text and the translation, supporting bidirectional mapping retrieval of Chinese-English, Chinese-German and other language pairs; the Original_content and Translations_content fields are both of type TEXT, and a single TEXT field can store up to 65KB of content, which fully meets the storage requirements of a single legal provision or case summary, ensuring the complete preservation of long and difficult sentences and complex logical structures.

[0059] At the physical storage level, this embodiment uses a MySQL database with the InnoDB storage engine, combined with Elasticsearch to achieve high-performance retrieval. InnoDB is chosen as the default storage engine primarily because of its support for transactions and row-level locking. Since the system is expected to store over 20,000 multilingual aligned data entries with an average row size of approximately 10 to 15 KB, the total storage space requirement is approximately 200 to 300 MB. InnoDB's row-level locking mechanism can effectively support high-concurrency writes from the translation agent and concurrent read requests from front-end users, avoiding data contention. Regarding the index optimization system, a B+ tree clustered index is built based on the primary key ID to ensure the continuity of physical data storage and improve the speed of ID-based retrieval. For frequently queried fields such as instrument_type, country_category, and the combination of source_language and target_language, composite auxiliary indexes are built to accelerate multi-dimensional filtering queries. Full-text indexes are configured for the Original_content and Translations_content fields. To further improve fuzzy matching and semantic retrieval capabilities, the system synchronizes this text data to the Elasticsearch cluster, utilizing inverted index technology to achieve millisecond-level cross-language legal terminology retrieval. Furthermore, considering the continuous growth of data volume, the physical table design adopts a horizontal partitioning strategy, partitioning tables based on country_category or data_source, storing data from different sources or legal systems in different physical partitions. This not only improves query efficiency for specific legal systems but also facilitates the implementation of independent data backup and lifecycle management strategies.

[0060] See attached document Figure 4Based on the core `semantic_mapping` table mentioned above, the system automatically derives and generates six major categories of specialized data tables through field mapping technology to support the construction of a database of foreign legal knowledge and cases. These specialized tables have customized fields for different legal documents: the case table covers case data from various legal systems such as English-French, German-French, European and American, and Canadian law, including specific fields such as case retrieval, judgment results, and parties involved, to support in-depth research on case law; the legal provisions table specifically stores the content of codified laws such as the German Civil Code, including fields such as legal provision title, legal provision content, and effective date, ensuring the timeliness of regulations; the litigation documents table includes fields such as document retrieval, litigation document citation, and litigation online language environment, used for analyzing litigation strategies and document norms; the international treaty table records treaty titles, signatories, and treaty content, serving international law research; and the judgment documents and judicial opinion tables respectively include fields such as legal content, case type, Chinese procedural rules, and judicial opinion content, used to extract judicial judgment logic and academic viewpoints. Before the data enters the storage layer, a strict data cleaning and integrity verification mechanism is implemented to ensure the high quality of the data entering the database. First, a three-layer data processing architecture was established to map heterogeneous data crawled from seven major databases into a standardized legal case resource library containing 18 core fields. Second, Apache Griffin was deployed for data quality monitoring, setting a field missing rate threshold of less than 0.2%; data exceeding this threshold was intercepted and alerted. Finally, the SimHash algorithm was introduced to calculate text fingerprints, with a Hamming distance of less than or equal to 3 considered as a duplicate case, effectively eliminating redundant data from multiple sources and ensuring the uniqueness of each record in the semantic mapping library. Through the above design, the database and storage layers not only achieved structured storage of 52,495 multilingual legal data entries, but also, through the translation agent to the mapping model to the relational database, accumulated over 20,000 high-quality Chinese-English and Chinese-German bilingual comparison data entries, providing a solid data foundation for upper-layer applications.

Claims

1. A model-enhanced multi-agent collaborative semantic mapping method for judicial texts, characterized in that: Includes the following steps: A multi-source heterogeneous legal data processing pipeline is constructed, a dynamic crawler engine is used to collect multi-dimensional raw legal data, and the raw legal data is cleaned and aligned to generate a standardized legal case resource library. A cross-linguistic semantic mapping model based on low-rank adaptive fine-tuning is constructed. A pre-trained large model is selected as the base, and the parameters of the pre-trained large model are fine-tuned using a legal parallel corpus and legal professional terminology knowledge is injected to form a generative base with legal semantic understanding capabilities. The translation agent is deployed as the scheduling center. The translation agent extracts the text to be processed from the standardized legal case resource library, calls the cross-language semantic mapping model as the generation core, and coordinates with the professional terminology knowledge base and evaluation feedback model to execute an iterative closed-loop link including terminology constraint generation, multi-dimensional evaluation and reflective correction to generate semantic mapping translations of the target legal system. A semantic mapping relation library of Chinese and foreign legal systems is constructed, and the semantic mapping translations are structurally associated and stored with the text to be processed. Based on field mapping technology, professional data tables are generated.

2. The model-enhanced multi-agent collaborative judicial text semantic mapping method according to claim 1, characterized in that, The construction of the multi-source heterogeneous legal data processing pipeline specifically includes: In the data collection phase, a dynamic crawler engine incorporating headless browser technology is deployed to process dynamically rendered pages by simulating human operation behavior. An integrated adaptive waiting algorithm is used to adjust the waiting time by dynamically monitoring the loading status of page document object model elements; An integrated request traffic smoothing scheduling model is used to simulate human reading intervals by scheduling request frequency. In the data cleaning stage, regular expressions are used to clean up formatting noise in unstructured text. In the data quality control stage, integrated data quality monitoring tools are used to monitor field missing rates, and the SimHash algorithm is used to calculate text fingerprints. Duplicate cases are identified and removed by calculating Hamming distance.

3. The model-enhanced multi-agent collaborative judicial text semantic mapping method according to claim 1, characterized in that, The construction of the cross-linguistic semantic mapping model based on low-rank adaptive fine-tuning specifically includes: The architecture of the pre-trained large model is optimized by replacing the multi-head attention mechanism with a group query attention mechanism, thereby reducing the cache memory usage of key-value pairs during the inference process. A dual-block attention mechanism is introduced, which divides the input sequence into multiple local blocks and combines intra-block attention calculation with cross-block information interaction to support input sequences that exceed the original context window limit of the pre-trained large model. A hierarchical and progressive data strategy was adopted to construct a parallel legal corpus, with data sources including written legal codes and bilingual case data; Freeze the backbone parameters of the pre-trained large model and fine-tune it by injecting a trainable low-rank decomposition matrix into the attention module of the Transformer layer.

4. The model-enhanced multi-agent collaborative judicial text semantic mapping method according to claim 3, characterized in that, The parameter fine-tuning and the injection of legal terminology specifically include: The rank of the low-rank decomposition matrix is ​​set to a preset low-rank value, and the low-rank decomposition matrix covers the query matrix, key matrix, value matrix and output projection matrix in the Transformer layer. In the fine-tuning data construction stage, key-value pair information from the legal terminology knowledge base is encoded into prompt words, and terminology constraint information is concatenated at the input end, forcing the cross-language semantic mapping model to focus on the fixed translation of terms during the training stage. A progressive, hierarchical fine-tuning process was implemented, and the cross-linguistic semantic mapping model was trained using general legal corpora, parallel corpora of written laws, and complex case data.

5. The model-enhanced multi-agent collaborative judicial text semantic mapping method according to claim 1, characterized in that, The translation agent, in collaboration with its components, generates a semantically mapped translation of the target legal system through an iterative closed-loop mechanism, specifically comprising a seven-step iterative process: The original text submission and preprocessing steps involve the translation agent receiving the input text and performing format cleaning and segmentation. In the terminology knowledge retrieval and constraint generation step, the translation agent calls the professional terminology knowledge base, identifies legal concepts and extracts standard translations, and generates terminology constraint prompts. In the first round of enhanced translation steps, the translation agent concatenates the original text with the terminology constraint prompts and calls the cross-language semantic mapping model to generate a preliminary translation; The multi-dimensional quality assessment steps involve the translation agent submitting the preliminary translation to the evaluation feedback model, which scores and diagnoses the translation based on evaluation indicators and legal logic rules. The feedback generation and strategy formulation steps involve the translation agent generating optimized strategy feedback based on the diagnostic results. In the second round of refined optimization, the translation agent reconstructs prompt words based on the optimization strategy, guiding the cross-language semantic mapping model to perform reflective translation correction. The results verification and database entry steps involve formatting the final preferred translation and writing it into the Chinese and foreign legal semantic mapping relationship database.

6. The model-enhanced multi-agent collaborative judicial text semantic mapping method according to claim 5, characterized in that, The evaluation feedback model employs a large language model with a high number of parameters, and the multi-dimensional quality assessment steps specifically include: The translation was diagnosed from three dimensions: terminological accuracy, syntactic and logical integrity, and cross-legal semantic equivalence. The ROUGE series of indicators are used as the core evaluation criteria. The semantic mapping quality is quantitatively evaluated by calculating the N-gram overlap and the longest common subsequence between the preliminary translation and the standard reference translation. The translation agent possesses self-reflection capabilities, adjusts sentence structure based on contextual logic, and eliminates ambiguity by adding qualifiers.

7. The model-enhanced multi-agent collaborative judicial text semantic mapping method according to claim 6, characterized in that, The adoption of the ROUGE series of indicators as the core evaluation criteria specifically includes: The ROUGE-N index is calculated by dividing the number of N-grams shared by the preliminary translation and the standard reference translation by the total number of N-grams in the standard reference translation, and is used to measure the coverage of legal terms. The ROUGE-L metric is calculated using a formula involving the longest common subsequence length of the preliminary translation and the standard reference translation. The ratio of the longest common subsequence length to the standard reference translation length is calculated as the recall rate, and the ratio of the longest common subsequence length to the preliminary translation length is calculated as the precision rate. Finally, a weighted harmonic mean of the recall rate and the precision rate is calculated.

8. The model-enhanced multi-agent collaborative judicial text semantic mapping method according to claim 1, characterized in that, The construction of the semantic mapping relation library between Chinese and foreign legal systems specifically includes: Construct a core semantic mapping data table, wherein the primary key identifier of the data table adopts an integer data type; The data table includes a document type field for identifying document dimensions, a country classification field for identifying legal system background, and a data source field for identifying specific source libraries; The data table includes source language and target language fields for identifying language correspondences, and original text and translated text fields for storing long content. Both the original text and translated text fields use text data types. Based on the core semantic mapping data table, case tables, legal provisions tables, litigation document tables, international treaty tables, judgment document tables, and judicial opinion tables are generated through field mapping technology.

9. The model-enhanced multi-agent collaborative judicial text semantic mapping method according to claim 1, characterized in that, The method also includes a traffic-based dynamic scaling and continuous integration mechanism: A reasoning service framework combined with pagination attention technology is used to manage the key-value pair cache in video memory, supporting parallel loading of multiple models; An asynchronous message queue is introduced to manage the data interaction between the translation agent and its components. The design of the continuous integration pipeline triggers an offline fine-tuning process to generate new low-rank adapter weights when new corpora or terminology database updates are acquired, and the model version is switched through gray-scale environment testing and blue-green deployment strategies.

10. A model-enhanced multi-agent collaborative judicial text semantic mapping system, characterized in that, include: The data processing module is configured to build a multi-source heterogeneous legal data processing pipeline, using a dynamic crawler engine to collect legal data and perform cleaning and alignment. The semantic mapping model building module is configured to build a cross-language semantic mapping model based on low-rank adaptive fine-tuning, fine-tuning the parameters of a pre-trained large model and injecting legal expertise. The translation agent scheduling module is configured to deploy translation agents, coordinate user input, a professional terminology knowledge base, the cross-language semantic mapping model, and an evaluation feedback model, and execute an iterative closed-loop link that includes terminology constraint generation, multi-dimensional evaluation, and reflective correction. The data storage module is configured to build a semantic mapping relationship library between Chinese and foreign legal systems, store semantic mapping translations, and generate specialized data tables.

Images