Search

CN-121980005-A - Reconsider case intelligent processing method for multi-mode fusion and large model

CN121980005ACN 121980005 ACN121980005 ACN 121980005ACN-121980005-A

Abstract

The invention discloses a reconsider case intelligent processing method of a multi-mode fusion and large model, and belongs to the technical field of legal government intelligent processing. The method comprises the steps of obtaining text, image and audio materials of reconsider cases, generating a standardized multi-modal data set binding a case number through standardized conversion and compliance verification, carrying out modal separation processing on the standardized multi-modal data set through a multi-modal analysis model to form a structured feature set, carrying out enhancement processing on the structured feature set, generating a fusion feature vector through cross-modal attention mechanism fusion, inputting the fusion feature vector into a fine adjustment large model to obtain case analysis results, generating reconsider processing opinion through combining similar historical cases, matching legal document template filling fields according to the obtained results, generating a standardized legal document through large model optimization, and meanwhile archiving data and manually modifying the log fine adjustment large model. The method improves the processing efficiency and the standardization level of reconsider cases of law and assists in accurate decision-making.

Inventors

  • LV DONGFANG
  • YANG GUANG
  • HAN CHUNYU
  • WANG DONGHE

Assignees

  • 北京万易信息技术有限公司

Dates

Publication Date
20260505
Application Date
20251225

Claims (9)

  1. 1. A reconsider case intelligent processing method of a multi-mode fusion and large model is characterized by comprising the following steps: s1, acquiring text materials, image materials and audio materials of reconsider cases, carrying out standardized conversion and compliance verification on various materials, and generating a standardized multi-mode data set binding unique case numbers; S2, carrying out modal separation processing on the standardized multi-modal data set by adopting a multi-modal analysis model to form a structural feature set; s3, after enhancement processing is carried out on each mode feature in the structural feature set, calculating the association weight among modes through a cross-mode attention mechanism and completing feature fusion to generate a fusion feature vector and a validity check result; s4, inputting the fusion feature vector into a large model for processing to obtain a case analysis result; s5, combining case analysis results and similar historical cases, and generating reconsider processing opinions through quantitative calculation of multidimensional decision factors; S6, processing the opinion and the structural feature set based on reconsider, matching legal document templates in a document template library, automatically filling dynamic fields, and generating a standardized legal document through large model optimization; s7, archiving reconsider full-flow data of the cases, constructing a feedback data set based on the manual modification log of the standardized law document, and performing incremental fine adjustment on the large model.
  2. 2. The method for intelligent processing of reconsider cases in a multi-modal fusion and large model according to claim 1, wherein the text-based material includes reconsider application, evidence list, legal representative person identity certificate and authorized attorney, the image-based material includes paper document scanning, live photo and document photo, and the audio-based material includes sound recording of hearing evidence and sound recording of applicant statement.
  3. 3. The method for intelligently processing reconsider cases with multi-mode fusion and large models according to claim 2 is characterized in that the standardization is converted into unified text materials into a pure text format of UTF-8 coding, special symbols and typesetting redundant marks are removed, the image materials are subjected to inclination correction after contrast enhancement by a CLAHE algorithm and are uniformly adjusted to 300DPI resolution, the audio materials are uniformly converted into WAV format files with 16kHz sampling rate and 16bit depth and are spliced in time sequence to generate audio indexes, the compliance verification is carried out through keyword library matching and audit model processing on various materials after the standardization conversion, and if any one of sensitive information, illegal content and content tampering suspicion exists in any one type of material data, the standardized converted material data is pushed to manual review.
  4. 4. The method for intelligent processing of reconsider cases with multi-modal fusion and large models according to claim 3, wherein the multi-modal analytical model includes BERT model, PP-OCRv model, resNet-50 model, legal scene ASR model and TextCNN model; The modal separation processing method comprises the steps of extracting case entities and dispute focus keywords from standardized text data through a BERT model, identifying text information from standardized image data through a PP-OCRv model, extracting visual features from standardized image data through a ResNet-50 model, obtaining audio transcription texts from standardized audio data through a legal scene ASR model, and analyzing emotion tendencies from the audio transcription texts by combining with a TextCNN model, wherein the structural feature set comprises the case entities, the dispute focus keywords, the text information, visual features, the audio transcription texts and the emotion tendencies.
  5. 5. The method for intelligently processing reconsider cases with multi-modal fusion and large models according to claim 4, wherein the process of enhancing is as follows: The case entity and the dispute focus keyword are respectively converted through a Word2Vec model to obtain a case entity Word vector and a dispute focus Word vector, and then feature stitching is carried out to obtain text modal enhancement features; Converting the text information into text information Word vectors through a Word2Vec model, performing dimension alignment on the text information Word vectors and visual features, and splicing the text information Word vectors and the visual features into image mode enhancement features through a full-connection layer; converting the audio transcription text into an audio Word vector through a Word2Vec model, and then carrying out feature stitching on the audio Word vector and emotion tendencies to generate audio mode enhancement features; inputting text mode enhancement features, image mode enhancement features and audio mode enhancement features into a cross-mode attention network to calculate inter-mode association weights, and carrying out weighted fusion on the mode enhancement features based on the inter-mode association weights to obtain a fusion result; Splicing the fusion result and the enhancement features of each mode, and carrying out BatchNorm layers of normalization processing to generate a fusion feature vector; and the validity checking result is obtained by calculating the information entropy and the feature variance of the fusion feature vector, and if any one of the information entropy <0.7 and the feature variance < a preset threshold exists, the method returns to S2 to trigger the multi-mode analysis model to re-optimize the model processing parameters.
  6. 6. The method for intelligently processing reconsider cases by using a multi-mode fusion and large model according to claim 5, wherein the large model is a basic large model finely tuned by a reconsider case corpus, the basic large model comprises GPT-4, LLaMA-70B, qwen-72B and Baichuan2, and the case analysis results comprise case compliance initial review conclusions, core disputes focus, fact and law basis mapping relation and evidence chain assessment reports.
  7. 7. The method for intelligently processing reconsider cases with multi-modal fusion and large models according to claim 6, wherein the multi-dimensional decision factors comprise case compliance scores, evidence chain integrity scores, law basis matching scores and similar case similarities, the similar historical cases search historical reconsider cases with the similarity between the historical reconsider cases in a historical reconsider case library and the fusion feature vectors of the current reconsider cases being more than or equal to 85% through a cosine similarity algorithm, and the fusion feature vectors of 3-5 historical reconsider cases with the similarity are pushed.
  8. 8. The method for intelligent processing of reconsider cases in a multimodal fusion and large model according to claim 7, wherein the legal document template library includes, but is not limited to, an acceptance notice template, a non-acceptance notice template and a reconsider decision template, the dynamic fields of the legal document template include, but are not limited to, applicant and applicant basic information, reconsider request and dispute focus, and the large model optimization includes legal expression specification checksum logical consistency optimization.
  9. 9. The method for intelligently processing reconsider cases with multi-modal fusion and large models according to claim 8, wherein the full-flow data is stored in an encrypted manner by using an AES-256 algorithm.

Description

Reconsider case intelligent processing method for multi-mode fusion and large model Technical Field The invention relates to the technical field of intelligent legal government affair processing, in particular to a reconsider case intelligent processing method of a multi-mode fusion and large model. Background In the technical field of intelligent legal government affair processing, reconsider case processing is a key link for guaranteeing the validity of administrative behaviors, and the problems of complicated flow, dependence on experience, nonstandard multi-mode material processing and low standardization degree in the traditional manual processing are required to be solved. Along with the increase of reconsider cases, the case materials present the multi-mode concurrent characteristics of texts, images and audios, and higher requirements are put forward on the automation and the precision of the processing, so that the intelligent transformation is realized by relying on multi-mode fusion and large model technology. According to the method and system for intelligent analysis of the mediation cases based on the AI large model, the automation of legal case processing is promoted by multi-mode fusion of large model processing mediation case multi-type evidence, but the method and system still have defects when applied to reconsider case scenes, namely, firstly, the scene suitability is insufficient, the prior art multi-focus mediation, lities and other case types are not customized according to the special compliance initial examination requirements of reconsider cases, administrative laws are subjected to custom design according to the applicable rules and multi-mode material type characteristics, secondly, the multi-mode data processing lacks standardization and integrity, part of technology does not cover specialized processing of audio modes, or unified multi-mode data conversion and compliance checking mechanism is not established, so that data foundation is not standardized, sensitive information and tampering content screening are not needed, the accuracy of subsequent processing is influenced, thirdly, feature analysis logic is not perfect, and more stays on the foundation feature splicing layer, the feature analysis rules and the administrative law are not built up in a cross-structure mode, the requirements of the strict feature analysis rules are not met, the requirements of the strict analysis rules are not met, the quality of the law case is not met, and the requirements of the strict analysis rules are not met, and the quality of the law case is not optimized, and the requirements of the strict are not-optimized, and the requirements of the quality of the law case are not met, and the requirements of the requirements are met, and the requirements of the quality of the analysis rules are not met, and the requirements of the requirements are met. The defects limit the processing efficiency and the processing accuracy of reconsider cases together, and the intelligent development requirements of legal government affairs are difficult to meet, so that improvement is needed. Disclosure of Invention The invention aims to provide a reconsider case intelligent processing method of multi-mode fusion and large models, which aims to solve the problems in the background technology. In order to achieve the above object, the present invention provides the following technical solutions: a reconsider case intelligent processing method of a multi-mode fusion and large model comprises the following steps: s1, acquiring text materials, image materials and audio materials of reconsider cases, carrying out standardized conversion and compliance verification on various materials, and generating a standardized multi-mode data set binding unique case numbers; S2, carrying out modal separation processing on the standardized multi-modal data set by adopting a multi-modal analysis model to form a structural feature set; s3, after enhancement processing is carried out on each mode feature in the structural feature set, calculating the association weight among modes through a cross-mode attention mechanism and completing feature fusion to generate a fusion feature vector and a validity check result; s4, inputting the fusion feature vector into a large model for processing to obtain a case analysis result; s5, combining case analysis results and similar historical cases, and generating reconsider processing opinions through quantitative calculation of multidimensional decision factors; S6, processing the opinion and the structural feature set based on reconsider, matching legal document templates in a document template library, automatically filling dynamic fields, and generating a standardized legal document through large model optimization; s7, archiving reconsider full-flow data of the cases, constructing a feedback data set based on the manual modification log of the standardized law document, and performing increment