271 results about "Semantic relevance" patented technology

The invention relates to the technical field of wireless high-definition audio and video transmission, and discloses a plug-and-play wireless high-definition audio and video transmission method and system.The method comprises the steps that foreground space texture features and audio and voice segments are obtained, and an initial feature set containing space and semantic features is obtained through semantic correlation analysis; a unified representation vector is output through multi-modal feature fusion, and a dynamic importance score is obtained by determining time sequence consistency, calculating an importance weight and performing normalization; and when the score exceeds a threshold value, segmenting the video frame in real time to determine an attention focus area and optimize a boundary, thereby generating an attention weight matrix, preferentially allocating bandwidth and forming a partition differentiation compression result. And in combination with a network bandwidth state, protection is enhanced for a key stream, the priority and the bit rate are dynamically adjusted, and high-quality audios and videos are output through decoding and recombination. According to the invention, the bandwidth allocation and compression strategy can be dynamically optimized, the transmission quality of a key area is guaranteed when the bandwidth fluctuates, and the audio and video transmission efficiency and experience are improved.

The invention provides a private deployment-oriented multi-modal AI knowledge base construction system. The private deployment-oriented multi-modal AI knowledge base construction system comprises a knowledge storage module, an intelligent document loading module, a document partitioning engine module, a data enhancement engine module, a multi-language semantic vector alignment module and a private deployment module, the knowledge storage module comprises a knowledge authority management sub-module and a knowledge source management sub-module, the knowledge authority management sub-module is used for managing and storing knowledge from different sources, and the knowledge source management sub-module is used for managing electronic documents and multimedia documents; according to the method, intelligent identification, partitioning, vectorization and source file storage can be carried out on different types of electronic files, knowledge graph construction is carried out for specific fields, semantic relevance between texts and topics and key entities is fully considered, the method has wider applicability, higher robustness and controllability, the data leakage risk is effectively reduced, and the method is suitable for popularization and application. The method is suitable for enterprise sensitive data protection and personal user elastic computing power requirements.

Computer-implemented systems and methods implement semantic search in high-dimensional vector spaces, specifically tailored for use with large language models (LLMs). In particular, clustering is combined with Euclidean distance measurements to facilitate real-time vector searches. By implementing clustering, the invention reduces the computational complexity and costs associated with Euclidean distance calculations, which are typically more resource-intensive than other methods such as cosine similarity. This reduction is achieved by limiting the scope of distance calculations to within clusters, thereby avoiding the inefficiencies and diminished accuracy otherwise encountered by existing systems when using Euclidean distance in high-dimensional spaces. As a result, the invention retains the benefits of Euclidean distance, such as its superior granularity and precision in measuring semantic relevance, without succumbing to the usual drawbacks of high computational demands and poor scalability.

The invention discloses a retrieval enhancement generation method for multi-source heterogeneous data fusion, and relates to the technical field of cross-modal retrieval generation, and the method comprises the following steps: collecting structured data and unstructured data, respectively carrying out standardization processing and semantic cleaning, and outputting a multi-source heterogeneous data set; constructing double-view knowledge block representation by using a multi-source heterogeneous data set, and respectively generating a retrieval view and a generation view; carrying out weight adjustment and fusion on the retrieval view and the generation view according to the current query content, and outputting a fusion semantic vector; performing approximate vector matching retrieval on the knowledge fragments by using the fused semantic vector, performing semantic reordering on a retrieval result according to semantic correlation, and outputting an ordered knowledge fragment set; according to the method, the context adaptability and semantic expression capability of the generated input are enhanced, so that the accuracy of knowledge retrieval and the quality and pertinence of text generation are remarkably improved.

The invention discloses a knowledge graph-based old building reconstruction scheme automatic recommendation method. The method comprises the following steps of S1, obtaining and preprocessing old building house data; s2, extracting a semantic entity and attribute relationship from a transformation case library, a building specification library and a construction scheme library, and constructing a knowledge graph; s3, mapping the house data to knowledge graph entity nodes, and executing graph embedding to generate building semantic representation; s4, constructing a graph neural network, calculating node semantic relevancy, and generating a building state vector; s5, inputting the building state vector into the reinforcement learning decision network, and optimizing the strategy to obtain an optimal transformation action; s6, screening reconstruction measures from the knowledge graph according to the optimal reconstruction action, and performing scoring to form candidate schemes; and S7, sorting the candidate schemes, selecting the scheme with the highest score, and recommending and updating the knowledge graph. According to the method, intelligent generation and self-optimization recommendation of the old building reconstruction scheme are realized, and the reconstruction efficiency of the scheme is remarkably improved.

The invention relates to the technical field of knowledge maps, and discloses a knowledge map completion method based on semantic-structure multi-level fusion. According to the technical scheme, for a training set positive example triple, a structured rationality score and a semantic correlation score are fused to screen high-quality training samples; in order to complement a query search relation path, screening a guiding path through a uniqueness index; integrating the query text, the entity description, the neighbor facts and the guide path to construct an enhanced input prompt; dynamic structure embedding is generated by using a relational graph convolutional network, the dynamic structure is mapped by an adapter module and then injected into a large language model, and completion prediction is completed by combining a fusion input fine tuning model. According to the method, through multi-level fusion of semantics and structural information, intelligent sample screening, prompt enhancement construction and structural information dynamic injection are achieved, the accuracy, reasoning ability and training efficiency of a large language model in a knowledge graph completion task are effectively improved, and the method is especially excellent in performance in complex relation reasoning and inductive completion scenes.

The invention provides an event-driven sparse attention optimization method and system based on a brain-like computing chip, and belongs to the technical field of computers.The method comprises the steps that semantic analysis and feature extraction are conducted on input data and text information, and an input event sequence and a text event sequence are generated; calculating semantic relevancy between the input event sequence and the text event sequence, and generating an event importance evaluation matrix; calculating a sparseness parameter based on the complexity of the input data and the load of a brain-like calculation chip calculation node, and strengthening the sparseness parameter to obtain a target sparseness parameter; based on the target sparseness parameter and the event importance evaluation matrix, generating a sparse attention mask matrix; a sparse attention calculation task based on the sparse attention mask matrix is allocated to calculation nodes of the brain-like calculation chip for execution; and the output results of the calculation nodes are fused to generate a target output result. According to the invention, the accuracy and efficiency of the query result are improved.

The invention discloses a multi-modal data drawing logical relationship analysis method, electronic equipment and a medium, and the method comprises the steps: generating a node set based on drawing image data and text data; generating a cross-modal hyperedge set based on the spatial proximity relationship, the visual feature similarity and the semantic correlation between the node sets; generating a hypergraph embedding input representation based on the node set and the cross-modal hyperedge set; the hypergraph is embedded into the input representation input improved hypergraph self-attention network model, and a hyperedge logic relation type and a corresponding hyperedge confidence coefficient are generated; generating a graph structure result based on the hyperedge logic relationship type and the node set, wherein the graph structure result meets the structure legality requirement; and performing hyper-parameter automatic adjustment and convergence control on the atlas structure result based on hyper-edge confidence, and generating an optimal atlas analysis model and a structured output result. According to the method, the reliability and the quality of analysis of component nodes, logic edge relationships and semantic structures in the drawing are improved.

The invention relates to the technical field of computer intelligent operation and maintenance and artificial intelligence, in particular to an operation and maintenance scheme generation method, system and device based on cooperation of a large language model and a knowledge graph and a medium. The method comprises the following steps: denoising and abstracting multi-source heterogeneous operation and maintenance data by using a large language model, extracting key fact dimensions, further identifying a fault entity, mapping the fault entity to an operation and maintenance knowledge graph to position an initial anchor point, executing two-stage collaborative reasoning based on the anchor point, and generating a plurality of candidate traceability paths by alternately performing relationship exploration and entity exploration; and finally, reordering the candidate paths based on path length perception and semantic correlation, dynamically controlling the termination and understanding strategy of reasoning by using a comprehensive confidence score, outputting a final solution, effectively overcoming the problems of information redundancy interference reasoning and large model illusion in a complex operation and maintenance scene through the cooperation of a large model and a knowledge graph, and improving the reliability of the system. And the accuracy of root cause positioning and the executable performance of the solution are obviously improved.

The invention relates to the technical field of advertisement putting, in particular to a real-time dynamic advertisement putting optimization method and system based on reinforcement learning. The method comprises the following steps: collecting multi-dimensional attribute data of an advertisement putting environment and user historical behavior data, and extracting adaptation degree characteristics of advertisement contents and the environment and user behavior preference characteristics; constructing an advertisement effect prediction model based on reinforcement learning, generating an optimal advertisement content form and putting strategy, and performing advertisement putting; based on the environment adaptation degree data and the real-time feedback data, linkage analysis of the advertisement content, the environment matching degree and the advertisement effect is carried out, and the comprehensive efficiency of an advertisement strategy is calculated; and dynamically optimizing and adjusting the advertisement strategy according to the analysis result. According to the method, through semantic correlation analysis of the advertisement and the environment and comprehensive evaluation of the semantic correlation analysis result and the user behavior, accurate prediction before advertisement putting is realized, and the advertisement putting accuracy, the user experience and the conversion effect are effectively improved.

The invention discloses a retrieval enhancement generation method for a time-sensitive problem, which comprises the following steps of: mixed time perception retrieval: enhancing document retrieval by adding time constraint on the basis of semantic relevance, and guiding by a time card to ensure that the retrieved document not only conforms to the meaning of query, but also conforms to the semantic relevance; the time context is met; the progressive multi-step reflection comprises the following steps of: firstly, acquiring and evaluating an initial document set by applying mixed time perception retrieval; if a document is retrieved, generating a final answer by using a large language model; otherwise, entering a reflection stage, and summarizing useful time information in the retrieved document into a context; and merging document sets accumulated in all iterations to generate a final answer. According to the method, a new framework integrating dynamic knowledge updating and time reasoning into the retrieval and generation process is provided, and accurate and timely response can be made to time-related problems.

The embodiment of the invention provides a context construction method and device for an agent and a storage medium, and the method comprises the steps: obtaining the current input information of the agent in one interaction round, and carrying out the analysis to generate a current semantic representation; determining a semantic association degree between the current semantic representation and the context of the current dialogue task, and calculating a long-term value score of the current semantic representation; under the condition that the semantic association degree is greater than a first preset threshold value, storing the current semantic representation into a short-term memory library; under the condition that the long-term value score is greater than a second preset threshold value, storing the current semantic representation into a long-term memory library; when the intelligent agent needs to generate a response for the current input information, searching target memory content related to the semantics of the current input information from the short-term memory library and the long-term memory library; and combining the target memory content with the current input information to form prompt information, and inputting the prompt information into a large language model to generate a response for the current input information.

The invention provides a multi-modal retrieval enhancement generation method and system, and relates to the technical field of computer vision. According to the method, better cross-modal semantic alignment can be realized in an embedding space through a multi-modal vector model; the semantic correlation between the target knowledge retrieved by the data of different modalities in the user input request and the user input request is ensured, so that the target knowledge is more accurate, the accuracy of an output result generated by the large language model is improved, and the generation quality of the large language model is improved. Through a mode of vectorizing the structured description text of the dynamic data, cross-modal semantic alignment can be realized, semantic information of a complex mode can be ensured to be fully captured, high-dimensional original data of the dynamic data is prevented from being directly processed, the calculation overhead required when the dynamic data is coded and retrieved is greatly reduced, and the efficiency of encoding and retrieving the dynamic data is improved. Faster retrieval and generation are supported, the generation efficiency of the large language model is improved, and the method is suitable for a real-time or resource-constrained environment.

The invention discloses an image sensitive content auditing method and system, and the method comprises the steps: obtaining the original description of a to-be-audited image, and the original description comprises the text summarization of the visual content of the image and the text content in the image; extracting an entity area in the to-be-audited image, retrieving related external background information based on the to-be-audited image and the entity area, eliminating interference information by using a label of the entity area, and generating a retrieval enhancement result; semantic correlation screening based on sensitive topics is carried out on the retrieval enhancement results, the screened retrieval enhancement results are fused into the original description in an iteration mode, and final fusion description is generated; and constructing a training data set containing positive and negative judgment sample pairs based on the fusion description, performing fine tuning on the visual language model, and performing sensitive content auditing on a new image by using the fine-tuned visual language model. The method can effectively improve the recognition accuracy and reliability of the sensitive content of the image.

The invention discloses a self-adaptive multi-feature fusion hybrid retrieval sorting method and system, and belongs to the technical field of information retrieval. The method comprises the steps that after user query is received, a mixed retrieval process and an intention recognition process are executed in parallel; performing multi-dimensional feature extraction on the candidate documents obtained by the mixed retrieval, wherein the multi-dimensional feature extraction comprises semantic correlation features, keyword matching features, document authority features and timeliness features; and according to the identified query type, adaptively selecting a fusion weight, and carrying out weighted fusion on the multi-dimensional feature vector to calculate a final score and sort the final score. According to the method, the problem that weight distribution is rigid in traditional mixed retrieval is solved through an intention self-adaptive dynamic weight mechanism, meanwhile, by introducing multi-dimensional service features, the ranking result not only ensures the correlation, but also meets the quality requirement under a service scene, and the accuracy and practicability of a retrieval system are remarkably improved.

The invention provides an urban flood disaster emergency prediction method based on a retrieval enhancement generation technology. The urban flood disaster emergency prediction method comprises the steps of partitioning, vectorization, retrieval, evaluation and feedback optimization. According to the method, a unified fragment set is constructed by introducing three types of data organization modes of fixed partitioning, semantic partitioning and structured data, and comprehensiveness and fine-grained expression of the data can be considered. Through primary filtering based on metadata, a candidate space is quickly reduced, high calculation overhead caused by global traversal is avoided, and retrieval efficiency is improved; in the retrieval stage, two complementary mechanisms of sparse retrieval and dense retrieval are combined, the sparse retrieval can ensure the accuracy of keyword matching, the dense retrieval can capture semantic relevance, and after the sparse retrieval and the dense retrieval are fused, reordering processing is performed, so that the relevance and ordering quality of results can be remarkably improved. And by setting a Top-K mechanism, it is ensured that the finally returned candidate segments give consideration to correlation and coverage.

The invention provides a three-dimensional point cloud change detection method based on Siamese AdaptConv, and belongs to the field of intelligent scene reconstruction. The method comprises the following steps: firstly, acquiring three-dimensional point cloud data of a scene to be detected at two different time points, and constructing an adaptive neighborhood for each point; secondly, inputting the preprocessed three-dimensional point cloud data into a Siamese AdaptConv double-branch network, fusing the multi-scale features of each layer to obtain respective corresponding fused features, further calculating the change probability of the same position point, mapping the change probability into a candidate change point set, executing DBSCAN clustering, and outputting a candidate region; and finally, performing semantic verification on the candidate region, pre-constructing a geographic knowledge graph, and calculating semantic relevancy: when the semantic relevancy exceeds a set threshold value, considering that the three-dimensional point cloud change of the candidate region is consistent with the existing entity semantics in the knowledge graph, and judging that the change is reasonable. According to the invention, end-to-end multi-time sequence feature alignment and change identification are realized.

The embodiment of the invention provides a voxel-level semantic mapping method for a human visual cortex, and the method comprises the steps: building a corresponding relation between a visual stimulation feature and a human brain voxel response through a visual language model and a coding model; adopting a BLIP image description generation and word segmentation technology to extract a plurality of noun candidate tags from a human brain image sample set highly responsive to target human brain voxels; generating a plurality of image samples consistent with the plurality of noun-type candidate tags in terms of semantics in batches, and inputting the image samples into the coding model for response evaluation; and by sorting the voxel responses of the plurality of image samples, distributing semantic correlation scores for the plurality of noun type candidate tags. According to the voxel-level semantic mapping method for the human visual cortex provided by the embodiment of the invention, the label bias can be reduced by utilizing the visual language model by constructing a voxel-level semantic mapping mechanism running in an open vocabulary space. The embodiment of the invention further provides a voxel-level semantic mapping device and electronic equipment.

The invention provides a weak password detection method based on multi-modal feature fusion and dynamic behavior analysis. According to the method, multi-dimensional feature extraction is innovatively introduced, including password entropy, semantic relevance, user historical behaviors, system login frequency and the like, so that user feature adaptation is improved, and dependency on a static dictionary is reduced; in addition, through real-time interactive feedback, instant pushing of password strength evaluation and safety suggestions can be realized. Therefore, by means of the method, the problems that in an existing weak password detection technology, the static dictionary dependency is high, the user feature adaptation is insufficient, and real-time feedback is lacked can be solved, the weak password detection precision and defense efficiency of colleges and universities are remarkably improved, and an efficient and easy-to-deploy password security solution is provided for the education industry.

The invention provides a building material retrieval method, which comprises the following steps of: acquiring a building material retrieval request described by a natural language of a user, performing deep semantic analysis by using a large language model to extract key information such as names, specifications, performances, scenes and the like, and generating a keyword set; performing synonym, hypernym, hypernym, industry term and associated word expansion on the keywords based on material attributes and semantic correlation to form an expanded keyword set; comprehensively retrieving multi-dimensional attributes such as names, specifications, performance, manufacturers, prices and the like in the building material database to obtain preliminary results; and performing weighted scoring and sorting by using multiple factors such as a semantic matching degree, an attribute matching degree, user evaluation and a market trend, and outputting a visual chart. According to the method, the building material retrieval accuracy, efficiency and user experience can be improved, and diversified retrieval requirements of professional users in the building industry are met.

The invention provides a multi-modal false news detection method based on unsupervised clustering and frequency domain information, and relates to the technical field of image text data. The method comprises the following steps: based on multi-modal sample data, respectively extracting text features, visual features and image text features, obtaining text feature clustering labels according to text feature global semantic correlation, inputting the text features with the labels into an unsupervised clustering learning network, and obtaining an unsupervised clustering learning result; semantic consistency is enhanced through a bidirectional gating loop unit and a multi-head attention mechanism, and enhanced text features are obtained; extracting frequency domain features based on the visual features, and fusing the frequency domain features with the spatial domain features to obtain visual joint features; and after image text features are also enhanced by the unsupervised clustering learning network, the image text features are fused with text and visual features through an attention fusion module by taking the image text features as a bridge to obtain multi-modal fusion features, and the multi-modal fusion features are input into a full connection layer to complete false news classification. According to the method, the accuracy of false news detection is effectively improved.

Certain aspects of the disclosure provide a method of adaptive information retrieval based on both lexical and semantic relevance. In some aspects, the method includes identifying a plurality of documents based on a context of the query request, assigning an integrated score for each respective document of the plurality of documents based on a semantic score for the respective document and a lexical score for the respective document, and ranking each document of the plurality of documents based on the integrated score.

The invention discloses a question and answer interaction method, device and equipment based on artificial intelligence and a medium, and relates to the technical field of artificial intelligence in the professional service fields of finance, insurance, medical treatment, banking and the like, and the method comprises the steps: segmenting a historical dialogue into segments with coherent themes, and constructing the segments into memory units; generating a semantic embedding vector and a time decay weight for each unit to form a memory bank; retrieving related memory units from a memory bank based on user query, comprehensive semantic relevance, theme consistency and time decay weight; the query and related units are combined into a context input generative language model, and a topic consistency constraint is applied during the generation process to generate a topic consistent reply. By optimizing the memory granularity and enhancing the memory representation and timeliness, multi-dimensional accurate retrieval and constraint generation are realized, and the system improves the accuracy, coherence and service consistency of question and answer interaction.

The invention discloses a thermal power plant intelligent question and answer method based on multi-source corpus fusion, and aims to solve the problems that knowledge construction depends on manual annotation, the corpus format is single, the context understanding ability is insufficient and the credibility of a generated result cannot be guaranteed in an existing question and answer method. The method comprises the following steps: introducing a vectorization processing mechanism and a semantic block model generation module to perform unified embedding and compression expression on various types of texts such as thermal power plant operation regulations, equipment manual, operation records and the like, so as to realize accurate retrieval and structured answering based on semantic correlation; and an answer judgment mechanism is introduced to reduce the risk that a language model generates an illusion phenomenon, so that the accuracy and interpretability of question and answer output are improved, and the intelligent question and answer requirements of operators in the aspects of equipment diagnosis, regulation query, accident handling and the like are met.

The invention discloses a bastion host intelligent operation guidance method and system based on RAG, and the method comprises the following steps: carrying out the partitioning processing of a bastion host operation document, constructing an operation document knowledge base and a vector database, and improving the semantic consistency of the document content; creating a chat window, and proposing an original natural language operation query; according to the method, the original natural language operation query is optimized and reconstructed to obtain a semantic optimization result of frame structuring, and the original natural language operation query is converted into an operation query vector, so that oral language content and semantic ambiguity of the operation query are effectively reduced, and standardized query content is provided for generation of operation guidance; therefore, the accuracy of operation guidance generation can be improved. Similarity retrieval is carried out, through a two-step retrieval method, the retrieval difficulty is reduced, the retrieval effect is improved, and context information with high semantic correlation is provided for generation of operation guidance; and inputting the retrieval content and the operation query into a dialogue model, and generating an operation instruction and a command instruction.

The invention provides a knowledge editing method based on multilingual semantic retrieval, and belongs to the technical field of natural language process.The method achieves cross-lingual knowledge updating through the two stages of multilingual knowledge retrieval and context editing and comprises the steps that firstly, a multilingual retrieval model based on XLM-R is used for mapping queries and knowledge base entries to a shared semantic space, and then the shared semantic space is used for conducting context editing; semantic correlation is judged through a classifier, and target language knowledge is retrieved; and splicing the retrieval result and the query in a zero sample or small sample mode to generate a prompt template, and inputting the prompt template into a large language model to complete editing. The method is characterized in that monolingual limitation is broken through, collaborative updating of 12 languages is supported, a retrieval-editing decoupling architecture is adopted to achieve model irrelevant adaptation, and retrieval and context learning are fused to improve editing accuracy. The method can efficiently correct multi-language large model fact errors, is suitable for scenes such as search engines and intelligent customer services, avoids the cost of full-model retraining, and remarkably improves the efficiency and accuracy of cross-language knowledge updating.

The invention discloses a link prediction method based on conditional diffusion negative sampling, and relates to the technical field of computers. According to the invention, based on a pre-constructed commodity relation network, a commodity node having an interaction relation with a target user is used as a target node, and a diffusion source is determined according to a mean value of embedded representations of at least part of commodity nodes in a high-order neighbor node set of the target node in other commodity nodes related to the target node. Therefore, negative sample generation is carried out based on the diffusion source through the diffusion model, the generated negative sample is ensured to maintain semantic correlation with a user historical behavior path and have enough discrimination challenging, the negative sample is fused with user portrait and commodity attribute multi-dimensional business characteristics, the semantic coverage range of the negative sample is greatly expanded, and the user experience is improved. The exploration capability of potential interests of the user is improved, and the recommendation effect of link prediction based on the method is improved.

An improved context-aware adaptive contrastive decoding method, relating to the technical field of artificial intelligence. The method comprises: for relevant retrieved content obtained by a retrieval-augmented generation (RAG) framework, using a dynamic context-weighted adaptive contrastive decoding (DC-ACD) method to perform DC-ACD management on the relevant retrieved content: step 1, establishing a dynamic weight computation model for contextual semantic relevance, and ensuring, by means of the dynamic weight computation model for semantic relevance, that a higher weight is assigned to retrieved content context having a higher semantic similarity; step 2, performing multi-document fusion; step 3, evaluating context quality; step 4, computing a Token-level dynamic weight on the basis of whether prior knowledge of an intelligent language model and / or contextual information are / is relied on; and step 5, analyzing a final decoding probability distribution: combining all weight proportions to obtain the final decoding probability distribution.

The invention relates to the technical field of conversation type artificial intelligence, in particular to a multi-round conversation robot intention hit optimization method. The method comprises the following steps: after a user utterance is received, a system generates candidate intentions through a double-path mechanism; then, each candidate intention is evaluated through a multi-factor scoring algorithm, and the algorithm fuses four key signals: a graph transition probability, semantic correlation, a context coherence score calculated through a cross attention mechanism, and an out-of-domain penalty; and finally, making a decision according to the final scores of the candidate intentions: if the confidence coefficient of the candidate intention with the highest score is high enough and the candidate intention is obviously distinguished from other candidates, confirming the intention; otherwise, a targeted clarification problem is generated to solve ambiguity. According to the method, the accuracy and robustness of intention detection are remarkably improved through collaborative integration of dialogue streams, semantic matching and deep context analysis, and meanwhile, the user experience is improved through intelligent ambiguity processing.

The invention discloses a contract document analysis method and system based on multi-modal adaptive feature fusion. The method comprises the steps of obtaining an image and text data of a contract document, and performing preprocessing; performing multi-modal feature extraction on the preprocessed image and the contract text to obtain image features and text features; calculating semantic correlation between the image features and the text features, and performing adaptive weight fusion to obtain fusion features; capturing the context relevance of the contract text, and optimizing the position and logic consistency of the initial bounding box according to the fusion feature to obtain a prediction result; and binding the prediction result with the image features and the text features, and outputting a structured analysis report. By implementing the method provided by the invention, high accuracy and traceability required by contract document analysis based on multi-modal adaptive feature fusion can be fully met, and the technical problems of multi-modal feature fusion, bounding box generation accuracy, transparency of an analysis process and the like in the prior art are solved.