221 results about "Context vector" patented technology

The invention relates to the technical field of natural language processing, in particular to a semantic synthesis and text vectorization-based data lake metadata governance method, which comprises the following steps of: performing vector coding of field content keywords and adjacent context words on the basis of field names, data types and description information of data sources in a data lake; and extracting semantic co-occurrence groups of the fields in the differentiated contexts, identifying distances between semantic vectors, and judging whether the distances are in a similar field grouping range or not to obtain a field semantic collection quantity. According to the method, a semantic co-occurrence relation is extracted through field keywords and context vector coding, a semantic hierarchical structure is constructed, the field division precision and abstract ability are improved, the consistency is judged in combination with a context semantic structure, a high-frequency alternative path is adjusted to optimize a semantic structure, and word vector similarity and a structure retention rate are fused to realize field merging; the continuity and the accuracy of a treatment structure are improved, and the intelligence and the consistency of metadata treatment in the data lake are enhanced.

A multimodal intelligent agent system for dynamic environmental monitoring and user-centered support, consisting of: a multimodal sensor module configured to continuously acquire environmental and behavioral data from multiple input modalities, including at least one visual sensor, at least one acoustic sensor, at least one environmental conditions sensor, and at least one proximity or motion detection sensor, each generating modality-specific data streams representing visual images, audio waveforms, physical environmental parameters, and motion signatures within a monitored environment; a data preprocessing and fusion subsystem that is operationally coupled with the multimodal sensor module and configured to normalize, temporally align, and transform the modality-specific data streams into high-dimensional feature embeddings using a variety of encoders, wherein the visual encoder uses convolutional or vision transformer architectures, the audio encoder uses a spectral-temporal feature extractor, and the sensor encoder transforms raw analog data into context vectors suitable for multimodal alignment; a multimodal processing unit consisting of a transformer-based large language model (LLM) trained on paired multimodal datasets and configured to perform semantic fusion, context abstraction, and inference across the aforementioned aligned multimodal feature embeddings to generate a contextual understanding of environmental and behavioral states; an adaptive agent controller coupled to the multimodal inference processing unit and configured to instantiate, manage, and terminate a variety of task-specific intelligent agents, each agent being a software unit configured to perform a specialized function selected from meeting summarization, behavioral analysis, misplaced object detection, or environmental anomaly identification, with the agents dynamically interacting with the inference engine to retrieve contextually relevant multimodal embeddings for task execution; a personalization and adaptive learning subsystem consisting of a user preference database and a neural memory structure configured to update and refine model parameters based on user-specific interaction history, thereby enabling personalized output generation, prioritization of recommendations, and long-term behavioral adaptation; and An output generation interface is operationally connected to the adaptive agent controller and configured to produce multimodal output in textual, visual, and auditory form. The interface is capable of displaying human-readable summaries, notifications, and visual reconstructions of identified entities or environmental states.

Disclosed in the present invention is a non-autoregressive Transformer-based modeling method for a 4-level pulse amplitude modulation high-speed transmitter. The method involves establishing a deep learning model having an encoder-decoder architecture to predict the behavior of a 4-level pulse amplitude modulation transmitter. An encoder processes unordered non-sequential inputs, including an input signal parameter and link parameters, to generate a context vector and then transmit same to a decoder. The decoder uses both the context vector generated by the encoder and a transmitter output signal sequence to generate a categorical probability distribution for each point in the sequence one by one. The model is trained using a random masking strategy, and inference is performed by means of non-autoregressive decoding and filtering, so that the model can perform parallel prediction on an output sequence, and perform a filtering process to predict an output signal. Compared to traditional simulation methods, the present invention achieves a significant acceleration effect, particularly when processing multi-link systems.

The invention discloses a sound interaction intention recognition and intelligent decision-making method based on an AI large model, and relates to the technical field of intelligent voice interaction, and the method comprises the steps: collecting a voice signal through an acoustic sensor, carrying out the noise reduction processing and acoustic feature extraction, capturing a text instruction, and carrying out the semantic segmentation and text feature extraction, splicing the acoustic feature vector and the text feature vector to form a multi-modal data packet; retrieving a historical memory library based on the enhanced fusion feature vector to generate a memory context vector, identifying the category of a deliberate map through a two-stage intention reasoning model, analyzing operation parameters, and outputting a structured intention instruction; and performing parameter legality verification, equipment state verification and security risk assessment on the structured intention instruction, and packaging the structured intention instruction into an executable instruction set after correcting abnormal parameters. According to the method, through double screening of the frequency band energy ratio and the lexical item importance score, the acoustic-text feature scale difference is reduced.

The invention discloses a knowledge graph enhanced reasoning framework based on a graph neural network and a large language model. The framework comprises a vectorization representation module, a graph neural network coding module, a large language model interaction module and a joint reasoning module. The vectorization representation module is used for aligning the structured knowledge graph and the unstructured text and converting the structured knowledge graph and the unstructured text into unified vectorization representation so as to provide a basis for subsequent processing; the graph neural network coding module obtains graph structure characterization by utilizing low-dimensional embedding of graph neural network learning data; the big language model interaction module performs semantic completion on the knowledge graph entity relationship by means of a big language model to generate an enhanced context vector; and the joint reasoning module fuses the results of the two modules and outputs an enhanced knowledge graph tetrad. Through multi-module cooperation, the accuracy and integrity of knowledge graph reasoning are remarkably improved, and the method has wide application prospects in the fields of intelligent questions and answers, intelligent decisions and the like.

The invention provides a wound surface grading model construction method and a wound surface self-analysis system, and the method comprises the steps: processing an obtained wound surface image, clinical text description data and structured background information data through a specific feature extraction strategy, and obtaining an image feature vector; clinical text description data and structured background information data are subjected to cleaning, word segmentation, entity labeling and other operations, a wound semantic evolution graph is constructed, related vectors are fused to obtain text embedding, wound background context vectors are generated according to the structured background information data, dynamic weighting, splicing and fusion are performed on image and text feature vectors, and the image and the text feature vectors are subjected to image fusion. According to the method, the fusion feature vector is obtained, the corresponding grading model is constructed according to the fusion feature vector, more accurate, comprehensive and explainable grading evaluation can be performed on different types of wounds, and a self-analysis system designed for the constructed model can quickly and accurately judge the wound image condition shot by a user.

The present application provides a large language model-based query statement generation method, apparatus, and device, and a medium. The large language model comprises a plurality of encoders. The method comprises: acquiring a query text; performing vectorization processing on the query text to obtain a target vector representation corresponding to the query text; inputting the target vector representation into the plurality of encoders, to perform encoding processing on the target vector representation by means of weight matrices of the encoders, so as to obtain context vectors corresponding to the encoders, wherein different encoders have different weight matrices; and generating a target query statement on the basis of the context vectors corresponding to the encoders. A model is allowed to focus on different information features in different attention heads, independently capture different aspects of a query text, and process information at multiple abstraction levels, thereby better capturing complex and abstract semantic relationships, effectively distinguishing the importance of information when processing complex and fuzzy questions, improving the accuracy of query statement generation.

The invention discloses a real-time translation recognition system under a cloud service framework, belongs to the technical field of real-time translation, and solves the problems that an existing translation system is insufficient in real-time performance, poor in scene adaptability, weak in privacy protection, slow in model evolution and the like. The dynamic model management engine obtains adaptive slices from a model slice factory according to scenes, equipment states and network quality and distributes the adaptive slices to edges, and the adaptive slices are distributed to a cloud-side collaborative reasoning system; the cloud-side collaborative reasoning system comprises a cloud-side collaborative reasoning system, a cloud-side collaborative reasoning system, a cloud-side collaborative reasoning system and a cloud-side collaborative reasoning system; the multi-modal perception engine fuses audio, images and dialogue history to generate a structured context vector and improve translation context fitting degree, the cloud edge cooperation engine takes an edge model as a core, processes different complexity tasks in combination with a cloud end, and constructs a data closed loop by incremental learning and a federation engine to realize model optimization and privacy protection; according to the system, the real-time performance, accuracy and safety are improved through cloud edge collaboration, dynamic adaptation and continuous learning, and the system is suitable for multi-scene real-time translation.

The invention provides a digital human AGI dialogue system based on a cloud side-end collaborative architecture, and relates to the technical field of digital humans, the system is characterized in that a sensing module, a processing module, a decision module, a rendering module and an output module are deployed at a side end, and a decision module, a driving module and a rendering module are deployed at a cloud end; the sensing module collects and preprocesses an input signal of a user; the processing module is connected with the sensing module and is used for extracting features of the input signals and generating context vectors; the decision-making module is connected with the processing module, and generates a decision-making result containing an answer text and an emotion label according to the context vector; the driving module is connected with the decision module, generates an audio stream and a phoneme sequence according to the answer text, and calculates skeleton driving parameters and mouth shape driving parameters of the digital human; the rendering module is connected with the driving module to generate a rendered picture; and the interaction module is connected with the driving module and the rendering module, and aligns the rendered picture and the audio stream to obtain an output result. And low time delay and high performance are realized by adopting cloud edge collaboration.

The invention provides a method, a device and equipment for generating a multi-round dialogue corpus and training and testing a large language model. The method for generating the multi-round dialogue corpus comprises the following steps: acquiring first question information; vectorizing the first question information to obtain a first question vector; querying a target preceding text vector of which the similarity with the first question vector meets a preset similarity condition from a vector database in which a plurality of groups of data pairs are stored; any group of data pair comprises a preceding text vector generated according to a preceding text of a question of the user in a historical interaction process with the language model, and a following text of the question in the historical interaction process; obtaining question generation constraint information based on the question following text corresponding to the target preceding text vector; calling the large language model to generate second question information after the first question information by taking the question generation constraint information as a constraint condition; and generating a multi-round dialogue corpus based on the first question information and the second question information.

The invention discloses a model training method and device based on a reattention mechanism. In the scheme, an embedded vector sequence is input into a multi-head attention layer of a pre-training language model to extract global semantic features, and a global context vector is generated based on the global semantic features; based on a sparse Softmax function and the global context vector, a sparse gating network calculates dynamic weights corresponding to multiple attention heads, and constructs an attention weight vector; performing weighted fusion on the context vectors output by the attention heads and the corresponding attention weight vectors by a multi-head attention layer, and performing linear transformation on the fused context representation based on a fixed weight matrix to obtain a self-attention value; and adjusting parameters of the pre-training language model according to the self-attention value through a preset loss function to obtain a target pre-training language model. According to the method, the sparse gating mechanism is introduced to dynamically adjust the weight of the multiple attention heads, and the model prediction precision is remarkably improved.

The invention relates to the technical field of artificial intelligence, and discloses an intelligent dosing method and system for lead zinc ore flotation to improve the efficiency, stability and economic benefits of a flotation process. The method comprises the following steps: calculating a first cross-modal feature and a second cross-modal feature, fusing the first cross-modal feature and the second cross-modal feature with a bubble image feature and an element grade feature through residual connection, and then processing to obtain a multi-modal feature sequence; inputting the multi-modal feature sequence of each time step into an LSTM network to obtain a corresponding LSTM hidden state sequence; calculating the time attention weight of the hidden state sequence of each time step in the time window, and performing weighted summation on the corresponding LSTM hidden state sequence according to the time attention weight to obtain a context vector in the time window; and finally, sharing the weighted and summed context vector serving as an input to a full connection layer corresponding to the two detection heads so as to respectively obtain probability distribution of predicted medicament types and a dose value of each medicament.

The invention discloses a semantic extension matching method and system based on a domain synonym library, and relates to the technical field of data processing, the method comprises the following steps: obtaining a query keyword input by a user, and generating a query context vector according to the query keyword and user context information; if the query keyword does not belong to the category word in the platform service category system, determining an extended synonym set matched with the query keyword from a pre-constructed field synonym library; for each extended synonym in the extended synonym set, calculating a correlation score between the extended synonym and the query context vector, and performing weighting processing to obtain a corresponding weighted extended synonym; and according to each weighted extended synonym and the query keyword, generating an extended query index for search matching. Therefore, dynamic expansion combining the field scene and the synonym library is realized, multi-dimensional semantic expansion can be carried out when complex and non-standardized user query is processed, and the intelligent level of the system is improved.

The invention relates to a request result generation method and device, a large model reasoning architecture, a vector database, equipment, a storage medium and a program product. The method comprises the following steps: unloading key value cache data obtained in a pre-filling stage to a vector database decoupled from a large language model for storage, and executing context vector retrieval and attention score calculation in a decoding stage in a text generation task by adopting the vector database decoupled from the large language model, therefore, the large language model can quickly reuse the target context vector obtained by the vector database and the attention score of the target context vector to carry out text generation reasoning, and a request result corresponding to the user request data is generated. By adopting the method, the data processing amount of the large language model can be greatly reduced, and the long context reasoning cost of the large language model is reduced.

The invention relates to the technical field of energy prediction, in particular to an energy short-term load prediction method and system based on SE-Block improved Transform. The method comprises the steps of performing reversible normalization preprocessing based on acquired multi-element load sequence data; carrying out feature extraction and fusion on the preprocessed data by utilizing improved cross-scale interaction Patching, wherein the feature extraction and fusion comprise multi-scale feature extraction, cross-scale interaction alignment, residual error correction and dynamic fusion; and performing feature screening on the fused features based on a channel attention mechanism, wherein the feature screening comprises feature response based on improved SE-Block and non-linear interaction of context vectors. Aiming at the non-stationarity of the actual load caused by the influence of meteorological conditions and user behaviors, the model accurately depicts the fluctuation details of the load curve by automatically eliminating the noise interference among multiple variables, and the robustness of the model in the multi-element load prediction of the integrated energy system is reflected.

A system and method for enhancing query responses from large language models without retraining by converting a query into a query vector; using a proximity metric to measure a proximity from the query vector to a plurality of vector embeddings stored in a vector database; ranking the plurality of vector embeddings based on proximity to the query vector; mapping the query to a homogenized context vector from a plurality of homogenized context vectors; using an augmented proximity metric to convert the proximity to an augmented proximity for each vector embeddings; performing an augmented ranking to refine the vector embeddings to those most relevant to the query; creating a prompt for a large language model comprising the query and the text data corresponding to refined vector embeddings as context; and feeding the prompt to the large language model to generate a response to the query.

The invention discloses a retrieval method based on a semantic enhanced knowledge graph, which relates to the technical field of information, and comprises the following steps: receiving a natural language query of a user, and carrying out deep analysis on the query, including named entity recognition and linking, relationship extraction and query intention classification; and based on an analysis result, extracting a related local sub-graph from the knowledge graph, generating a query context vector, and generating dynamic semantic embedding for the sub-graph through a query-perceived graph attention network to obtain a dynamic enhanced semantic graph. According to the retrieval method based on the semantic enhancement knowledge graph, the retrieval precision and the recall rate are remarkably improved, the limitation of static knowledge representation is solved through a dynamic semantic enhancement mechanism of query intention perception, so that the local semantic representation of the knowledge graph is highly aligned with the query intention of a specific user; and the ability of understanding and answering complex, fuzzy, ambiguous and multi-hop queries is improved.

The invention discloses an attention-based context-aware sparse hybrid expert model routing method, which comprises the following steps of: encoding prompt information input by a user to obtain context vector representation of the prompt information, and introducing a multi-head attention mechanism to obtain multi-scale semantic interaction information; further, constructing a gating network based on attention output, and dynamically selecting Top-K expert networks for reasoning; by introducing a self-adaptive neighbor attention weight and a fusion gating mechanism, expert dispatching and weight fusion of a token level are realized; combining with the sub-output of each expert network, and aggregating according to the weight to obtain the final model output; according to the method, the understanding ability of the model for different semantic contexts is enhanced through a multi-expert structure and a dynamic routing mechanism, and the method is adaptive to multiple rounds of token generation processes, so that the accuracy and diversity of generated texts can be improved.

The invention discloses a multi-role configuration and effective judgment method based on semantic arbitration, and belongs to the technical field of artificial intelligence, and the method comprises the steps: setting a multi-organization role and an arbitration role; receiving and vectorizing a test description text, distributing the test description text as a first clinical test context vector to multiple organizations, randomly assigning a first speaking party and setting an empty discussion pool; the first speaking party generates speaking content based on the clinical test context vector, vectorizes the speaking content and writes the speaking content into a discussion pool, calculates the similarity between a speaking content vector and the clinical test context vector, and calculates an arbitration score in combination with multi-organization attributes; the arbitration role executes three-state judgment according to the arbitration score and the parameters, an arbitration decision is generated, and an effective speaking content vector is determined; fusing, updating and writing the effective vector and the clinical test context vector into a discussion pool; discussion is terminated according to preset conditions, and based on the final vector and the discussion pool, effective judgment is output and a log is recorded; according to the method, a multi-organization discussion scene can be effectively simulated, and the clinical test discussion efficiency and accuracy are improved.

The invention provides a 3D modeling feature extraction method and system combined with natural language processing, and relates to the technical field of game development.The method comprises the steps that firstly, text description data containing morphological attributes, spatial relations and surface features of a target object is obtained, multi-level semantic analysis is conducted on the text description data, and a key semantic feature set is extracted; comprising global and local semantic vectors and dynamic context vectors, then generating intermediate representation data comprising geometric contour features, material distribution features and illumination response features of the target object based on the key semantic feature set, and executing three-dimensional topology reconstruction processing according to the intermediate representation data to generate initial 3D model data; and finally, performing multi-stage optimization processing on the initial 3D model data to obtain an optimized 3D model feature set containing geometric smoothness, texture continuity, illumination rendering and physical collision features, thereby improving 3D modeling efficiency and accuracy, and providing an efficient modeling scheme for game development and the like.

The invention relates to the technical field of artificial intelligence, and discloses an abstract generation method, device and equipment based on a large language model and a storage medium, which are applied to a text abstract generation scene of financial businesses, and the method comprises the following steps: obtaining a to-be-processed text, and preprocessing the to-be-processed text to generate a sentence sequence and a chunk sequence; performing vectorization and attention score calculation on the sentence sequence and the chunk sequence to screen out an initial candidate sentence sequence; based on the similarity between the initial candidate abstract fragments and items in a preset knowledge base, target knowledge base items are screened out, and the target knowledge base items are spliced with the candidate abstract fragments to obtain a target candidate sentence sequence; and converting and splicing context vectors in the target candidate sentence sequence to generate a spliced sequence, dynamically adjusting the attention weight of each vector in the spliced sequence, generating target attention information, and performing text abstract extraction through a preset abstract extraction model to generate target abstract information. According to the invention, the abstract generation accuracy is improved.

The invention discloses a front-end cache management method, system and device for a lightweight large-model dialogue state and a medium, belongs to the technical field of front-end cache management of a large-model dialogue system, and aims at solving the technical problem of how to overcome the defects that in a traditional scheme, long context cache is low in efficiency, storage redundancy and insufficient in dynamic semantic adaptation capacity, and the large-model dialogue state cannot be managed easily. In order to realize dialogue context volume compression, improve semantic similar request hit rate and reduce cross-end synchronization delay, the adopted technical scheme is as follows: data acquisition and preprocessing: capturing user interaction behaviors in real time through front-end burying points, and performing preprocessing operation on the acquired user behavior data; semantic normalization processing: performing embedded vector conversion and semantic clustering on the text input by the user to generate a unique semantic identifier and a context vector; querying and updating the multi-level cache; and dynamic collaborative updating: dynamically adjusting the cache based on the cache hit rate, the response delay and the user feedback, and optimizing the cache effect in real time.

The invention provides an intelligent teaching method and system fusing a learning track and an attention mechanism, and belongs to the technical field of education, and the method comprises the steps: obtaining a question text inputted by a user and historical learning behavior data of the user; using a current problem encoder to construct an input problem text into a current problem semantic vector; using a user memory encoder to construct the historical learning behavior data into a user memory vector sequence, and updating a user memory state based on the user memory vector sequence; based on the semantic vector of the current problem and the updated user memory state, a context fusion module of a dynamic memory attention mechanism is adopted to generate a memory context vector, and meanwhile, a double-flow user behavior encoder is adopted to obtain double-flow representation; and fusing the double-flow representation and the memory context vector, and obtaining a final teaching answer through a decoder and an output generator. According to the invention, deep modeling and long-term memory integration of the learning track of the user are realized, so that more efficient and personalized teaching services are provided.

The invention discloses a capsule network sentiment analysis method for teaching evaluation. Firstly, teaching evaluation texts are collected, aspect items are extracted through data preprocessing, an aspect category-emotion two-tuple is generated through manual annotation, and a teaching evaluation data set is constructed. Secondly, splicing each evaluation text and all aspect categories, and inputting the spliced evaluation text and all aspect categories into a pre-training language model for encoding to obtain high-dimensional context vector representation; thirdly, text features highly related to a specific aspect are extracted through a cross attention mechanism, and modeling and classification of category sentiment polarity of all aspects are achieved through a capsule network and a dynamic routing mechanism; and finally, judging the existence of aspect categories and the sentiment polarity of the aspect categories through a multi-task classifier, and outputting a plurality of aspect-sentiment two-tuples contained in the sentences. The sentiment analysis accuracy and interpretability in a multi-aspect and multi-sentiment polarity coexistence scene in a teaching evaluation text are effectively improved, and the method is suitable for intelligent analysis of large-scale education evaluation data.

The invention discloses a subject entity labeling method and system fusing image recognition and a knowledge graph, and relates to the technical field of image recognition and natural language processing. The method comprises the following steps: carrying out preprocessing and image-text association on multi-source heterogeneous subject data; detecting a visual entity in the image through an improved YOLO model, and extracting and linking a text entity in combination with a subject dictionary and a knowledge graph; cross-modal collaborative disambiguation is realized by calculating the semantic similarity of visual candidate entities and text context vectors; multi-modal entities are combined, relation reasoning and enrichment labeling are carried out in a knowledge graph, and a deep labeling result containing the entities and a semantic relation network of the entities is generated; the problems of difficulty in multi-source data fusion, inaccurate professional entity recognition and difficulty in semantic ambiguity elimination are effectively solved, the depth and accuracy of subject knowledge semantic understanding are remarkably improved, and key technical support is provided for intelligent education application.

The invention relates to the technical field of syndrome evolution prediction, in particular to a syndrome evolution prediction method, system and device based on multi-modal data driving and a medium, the method comprises the steps that multi-modal syndrome data is acquired and preprocessed, and the multi-modal syndrome data comprises clinical symptom data, tongue condition feature data and pulse condition feature data; extracting bidirectional time sequence characteristics in the preprocessed multi-modal syndrome data through a bidirectional long-short-term memory network, and generating a hidden state sequence; weighting the hidden state sequence by using an attention mechanism to obtain context vectors of contribution weights of different time steps; and inputting the context vector into a hidden Markov model of which the state transition matrix is constrained and corrected by the traditional Chinese medicine theory, performing syndrome state reasoning and evolution trend prediction, and outputting a prediction result. The objective of the invention is to improve the prediction precision and interpretability of syndrome evolution.

The invention provides a prototype pseudo instance enhanced video anomaly detection method and system, and the method comprises the steps: carrying out the feature extraction of a video, obtaining a depth feature, and obtaining a feature sequence based on the depth feature; inputting the feature sequence into a prototype interaction layer, and calculating through a normal prototype set to obtain cosine similarity; enhanced features are obtained through cosine similarity; performing calculation through a classifier to obtain an abnormal score; screening the videos through the abnormal scores to obtain an index set of extreme instances; based on the index set of the extreme instances, performing low-dimensional mapping on the enhanced features corresponding to the extreme instances to obtain depth feature embedding representation; constructing a loss function based on the depth feature embedding representation; training the model in combination with a loss function; and obtaining a detection result through the trained model. According to the method, feature context vectors are dynamically generated through a learnable normal prototype set and an attention mechanism, structured normal knowledge is embedded into features, and a robust normal reference is established.

A translation method is provided, including: encoding to-be-processed text information to obtain a source vector representation sequence, the to-be-processed text information belonging to a first language; obtaining a source context vector corresponding to a first instance according to the source vector representation sequence, the source context vector indicating to-be-processed source content in the to-be-processed text information at the first instance; determining a translation vector according to the source vector representation sequence and the source context vector; and decoding the translation vector and the source context vector, to obtain target information of the first instance, the target information belonging to a second language.

Embodiments of the present disclosure provide a method, an apparatus and an electronic device for determining a word representation vector, and a computer-readable storage medium, which belong to a field of processing natural languages. The method includes obtaining a set of glyph units of a text; obtaining a context vector of the text based on the set of glyph units; and predicting next word of the text based on the context vector. The method for determining the word representation vector of the present disclosure can effectively obtain a corresponding set of glyph units even for hieroglyphics in which hyperbolic characters are prone to appear or languages evolved from the hieroglyphics, thereby improving an accuracy of determining the word representation vector.

The invention discloses a knowledge graph generation method and system based on RAG deduction, relates to the technical field of knowledge graph generation, and obtains a preliminary candidate text set by extracting an expected entity category, a relationship category and a field context in a knowledge graph generation task, constructing a task structure template and performing dense retrieval in combination with keywords. Then, entity and relation coverage is analyzed, missing information is recognized, retrieval is supplemented, and a text set is optimized; fusing the text set and obtaining state information, calculating a fusion quality disqualification coefficient, and if the coefficient is smaller than a threshold value, generating a structured triple set based on a fusion context vector; otherwise, re-fusing until the generation is completed. According to the method, a fusion mechanism process can be effectively detected, potential problems can be timely identified and corrected, risks of entity errors, relation mismatching and the like in knowledge graph generation are reduced, and the accuracy and usability are improved.