117 results about "Semantic difference" patented technology

The invention discloses a retrieval enhancement question and answer generation method and system based on a large language model in combination with a multi-agent cooperation mechanism. The method comprises the steps of S1, text preprocessing and semantic difference enhancement and amplification; s2, checking and complementing user questions; s3, evaluating problem complexity and selecting a generator; s4, text retrieval and answer generation; the system comprises an information enhancement processing module, a query understanding optimization module, a question complexity intelligent evaluation module and a collaborative answer generation module, and is used for realizing the method. According to the method, a multi-agent collaborative RAG framework of an information enhancement agent, an interaction analysis agent and a problem complexity evaluation agent is introduced, and semantic difference enhancement and amplification, query integrity verification and complementation and a dynamic generation strategy based on problem complexity are performed after hierarchical document partitioning are combined; and the performance of the system in the aspects of semantic distinguishing capability, query understanding precision, generation efficiency and reliability is comprehensively improved.

The invention discloses an Internet of Things real-time task distribution system based on AI scheduling, and the system comprises a data collection module which is used for generating a task input vector, and constructing a node resource vector set; the semantic modeling module is used for inputting the task input vector into an improved StructBERT model to construct a task semantic vector; the intention analysis module is used for constructing a historical task intention vector library and a semantic conflict graph; the comparison judgment module is used for executing task conflict judgment and scheduling priority adjustment operation according to the task aggregation degree and the semantic difference vector; and the scheduling execution module is used for executing a scheduling calculation operation, generating a task allocation result and updating the historical task intention vector library and the semantic conflict graph. The method has the advantages of high task semantic understanding precision, high conflict judgment efficiency and high scheduling decision intelligent degree, and is suitable for a real-time task allocation scene in a multi-node heterogeneous Internet of Things environment.

The invention discloses a semantic unification and anomaly traceability analysis method and system for heterogeneous program logs, and the method comprises the steps: firstly collecting original logs in various formats, and packaging the original logs into standard objects in a unified manner; thirdly, identifying fields through a rule and a semantic model, uniformly mapping and labeling semantic tags, eliminating semantic differences, and constructing log data with uniform semantics; automatically constructing a cross-system call chain based on the request identifier and the like; secondly, extracting log features and comparing the log features with a normal behavior model to carry out anomaly detection, and generating an abnormal event; combining time, calling and dependency to construct an anomaly propagation path, and reversely backtracking and positioning an anomaly source; abnormal priority scores are calculated according to the propagation range, the influence degree and the like, and sorting and visual display are carried out; through unified semantic modeling and automatic association analysis of heterogeneous logs, accurate tracing and intelligent positioning of cross-service link anomalies are realized, the operation and maintenance efficiency is effectively improved, and the dependence on artificial experience is reduced.

The invention provides a natural image matting method and system based on text and boundary information aggregation, and relates to the technical field of image processing. Splicing the original color image and the corresponding ternary image, extracting initial features, and obtaining enhanced fusion features through multi-scale Laplacian high-frequency extraction and cosine similarity weighted fusion; based on the enhanced fusion feature and the ternary image, generating a gated ternary fusion feature fused with priori knowledge through a trans-attention mechanism; global coding modeling is carried out on the gated three-value fusion features to obtain deep features; text prompt and multi-scale boundary information are introduced based on deep features, adaptive up-sampling is guided through a cross-attention mechanism, semantic difference consistency constraint is adopted between decoding layers, consistency constraint is implemented from pixel appearance, high-level semantics and color component dimensions, and finally a transparency image is output through a prediction header to obtain an image matting result. And the fidelity and the boundary accuracy of high-frequency details in a matting result are effectively improved.

The invention discloses a semantic alignment method and device for heterogeneous data and a computer program product, and the method comprises the steps: firstly carrying out semantic enhancement processing on a heterogeneous data field, and obtaining enhanced field representation information; and then coding the first heterogeneous data field and the second heterogeneous data field which belong to different data sources respectively by utilizing a pre-constructed domain language coding model with the capability of accurately understanding specific domain specific terms and business contexts to obtain semantic vectors of the first heterogeneous data field and the second heterogeneous data field so as to capture subtle semantic differences between the first heterogeneous data field and the second heterogeneous data field. Then, after the semantic similarity of the first heterogeneous data field and the second heterogeneous data field is calculated according to the semantic vectors of the first heterogeneous data field and the second heterogeneous data field, a target interpretation template is obtained through matching from an interpretation template library in combination with enhanced field representation information; the method and the device are used for generating a semantic alignment result and an interpretation of a first heterogeneous data field and a second heterogeneous data field, so that the semantic alignment efficiency and accuracy of the heterogeneous data fields can be effectively improved.

The invention provides a semantic communication method and system based on multi-modal perception, and the method comprises the steps: obtaining multi-modal information, carrying out the semantic extraction of the multi-modal information, and generating a corresponding current semantic representation; performing semantic association on the current semantic representation and historical semantic information to obtain context enhanced semantic information; determining a semantic difference degree or a semantic change trend between the current semantic representation and historical semantic information based on the context enhanced semantic information to generate a key frame semantic description; semantic processing and semantic completion are carried out on the semantic description of the key frame; and according to a set context constraint condition, performing conditional semantic restoration on the complemented semantic description of the key frame by using a multi-modal model, and outputting auxiliary decision information or prompt information based on a restoration result. According to the method, the multi-modal data transmission bandwidth requirement is reduced, and meanwhile, the robustness and continuity of semantic communication in a complex channel environment are improved.

The invention provides a supplier grading and classifying method based on multi-dimensional feature fusion, and relates to the technical field of supply chain management. The method comprises the steps of obtaining structured indexes and unstructured text data of suppliers; structural features are extracted through a multi-layer perceptron, and semantic features are extracted through Transform; a cross-modal attention mechanism is adopted, and deep fusion is carried out with the structured features as queries and the semantic features as key values to obtain interaction features; performing dynamic modeling on the historical time sequence data through bidirectional LSTM to generate time sequence comprehensive features; inputting a multi-task hierarchical classifier, combining ordinal regression and contrast learning loss optimization, and outputting categories and grades; according to the management priority, the weight is dynamically adjusted through attention bias, and grading flexible adaptation is achieved; and incremental learning fine tuning is carried out by using actual service data through closed-loop feedback. The method solves the problems that multi-source heterogeneous features are difficult to fuse, and feature space semantic distinguishing and grading standards lack dynamic adaptation, and is applied to intelligent evaluation of industrial internet suppliers.

The invention particularly relates to a multi-modal data processing and retrieving method. The multi-modal data processing and retrieval method comprises the following steps: respectively carrying out depth feature extraction on image data and text data to generate an image vector and a text vector; carrying out interaction on the image vector and the text vector, and capturing semantic association information among multiple modes; mapping the vectors after interaction to a unified semantic space to realize semantic alignment among different modes; dividing the unified semantic space vector into fragments, storing the fragments in distributed nodes, and establishing a vector index; and generating a query vector by using an image or text queried by a user, carrying out parallel calculation on the similarity with a storage vector, and returning a retrieval result according to the similarity. According to the multi-modal data processing and retrieval method, the problems that the semantic difference between different modal data such as images and texts is large, the retrieval efficiency is low and storage is difficult to expand are solved, the accuracy and efficiency of multi-modal data retrieval are remarkably improved, and the method has remarkable technical advantages and wide application scenes.

This invention relates to the field of software development technology and discloses a document processing method, apparatus, device, and medium. The method includes: acquiring multiple heterogeneous data sources of a software project, wherein the heterogeneous data sources include at least source code and software documents associated with the source code; constructing a semantic difference graph describing the semantic relationships between the heterogeneous data sources based on the multiple heterogeneous data sources; assessing the synchronization risk value between the software documents and the current source code based on the semantic difference graph; if the synchronization risk value exceeds a preset threshold, determining the target content in the software documents that needs to be updated synchronously based on the semantic difference graph; generating differential update information for updating the target content based on the current source code and the semantic difference graph; and triggering an update processing flow for the software documents based on the differential update information. This method can be applied to document processing scenarios in fintech and healthcare, improving the accuracy and efficiency of document processing.

The invention discloses a query selection-based ship sign accurate detection method under a complex background. Aiming at the characteristic that the ship brand target scale is variable, a scale-sensitive foreground selector is introduced to separate foreground and background queries, only the selected foreground query is processed layer by layer in a Transform encoder, and foreground features are subjected to enhanced expression. In order to compensate for the semantic difference between the background query and the foreground query which are not subjected to self-attention calculation, a semantic reconstruction module is designed to bridge the encoder and the decoder. Additional explicit reference point supervision is provided in a decoder to enhance the validity of reference point generation. Finally, the whole model has good ship sign detection performance under a complex background.

The invention provides an identification system and method for intrinsic semantic difference learning, relates to the technical field of wine body identification and identification, and solves the problem that weak but essential differences between highly simulated adulterated wines and true wines are difficult to effectively identify from highly simulated adulterated wines. In the system, an instrument module forms unified wine body sample pair multi-source data for detected wine bodies and reference wine bodies, and a feature embedding module converts the unified wine body sample pair multi-source data into advanced semantic features and fuses the advanced semantic features; the intrinsic semantic extraction module is used for respectively extracting intrinsic semantic features reflecting a detected wine body and a reference wine body, the true and false difference syndrome extraction module is used for obtaining key difference features between the detected wine body and the reference wine body, and the causal diagram reasoning flavor extraction module is used for respectively extracting a component coupling relationship reflecting the detected wine body and the reference wine body; and finally, the true and false wine inference module identifies the true and false of the detected wine body to obtain an identification result. According to the method, multi-modal heterogeneous data are fused, true and adulterated wines can still be accurately distinguished under the condition that components are highly similar, and a more essential and reliable identification effect is achieved.

The invention provides a test case processing method and device, and the method comprises the steps: recognizing the semantic difference between m modified code files and corresponding m unmodified code files, so as to obtain M to-be-tested methods, and the to-be-tested methods are methods containing the semantic difference; n candidate test cases are determined from the multiple test cases of the test case set, the N candidate test cases are used for testing the M to-be-tested methods, and the candidate test cases cover at least one to-be-tested method in the M to-be-tested methods. According to the scheme of the embodiment of the invention, the method needing to be tested can be identified, so that the test efficiency is improved.

The invention discloses a personalized image generation method and system based on maximum difference anchoring and mask attention guidance. In order to solve the problems of concept excessive generalization and background interference distortion of an existing text-to-image diffusion model in a user-defined concept and general concept combined scene, an end-to-end dynamic constraint framework is constructed. The method comprises the following steps: firstly, generating a candidate set related to a user concept through a multi-modal large model; a plurality of anchor point concepts with the maximum semantic difference are selected through K-means clustering; secondly, designing a double-stage training mechanism, wherein in the first stage, attention loss guided by an SAM mask is adopted to suppress background interference, and in the second stage, maximum differentiation anchor point constraint loss is introduced to optimize embedded distribution; and finally, a high-quality image fusing the custom concept and the general concept is generated. According to the method provided by the invention, unique features of a user-specified concept can be better reserved in the generated image, meanwhile, the deficiency or distortion of a general concept is avoided, and the consistency and the structural accuracy of the generated image are improved.

The invention discloses a pregnancy assisting process management system and method based on intelligent interaction, and relates to the technical field of data processing. The intelligent interaction-based pregnancy assisting process management method comprises the steps of S1, collecting pregnancy assisting stage data and historical symptom data of a pregnancy assisting process, performing preprocessing, generating semantic vectors in combination with natural language processing, and constructing a pregnancy assisting management database; s2, analyzing semantic meanings of the symptom words in different pregnancy-assisting stages through pregnancy-assisting stage data, and adjusting the matching precision of the symptoms and the pregnancy-assisting stages; s3, quantifying the context matching degree by analyzing the comprehensive semantic representation and stage semantic difference of the symptom words; and S4, fusing multi-dimensional stage semantics, feature data, evaluation symptoms and pregnancy assisting stage matching degrees to carry out stage classification analysis, and realizing stage attribution optimization and problem type classification. The problems that in the pregnancy assisting process, symptoms and different stages are difficult to accurately match, contextual understanding is insufficient, and the problem classification error is large are solved.

The invention relates to the technical field of material science data extraction, in particular to an end-to-end material property relation extraction method of a large language model, which comprises the steps of literature data preprocessing, dynamic data classification, context sensing retrieval, staged relation extraction and structured output. The method is characterized in that a semantic comprehension capability of a large language model (LLM) is combined with a retrieval enhancement generation (RAG) framework. According to the method, cross-domain migration can be realized without fine adjustment of the model by utilizing a collaborative architecture of a large language model and retrieval enhancement generation (RAG), the technical bottleneck that a traditional method needs repeated training is solved, a multi-query difference retrieval strategy is adopted, and the literature recall rate at10 is increased to 100% through a semantic differentiation query generation technology, so that the method is remarkably superior to a traditional single-query retrieval method, and the method has the advantages that the method is simple and convenient to operate and high in efficiency. And a staged and classified prompt framework is innovatively designed.

The invention provides a multi-modal fusion-driven ship navigation target detection method and device, and the method comprises the steps: extracting the feature vectors of ship vision, marine radar, AIS and ship navigation data, taking the vector of the marine radar as a dominant vector, taking the other feature vectors as supplementary fusion, and achieving the modal complementation. And mapping the fusion feature vector to a bird's-eye view semantic space to obtain bird's-eye view features, decoding the bird's-eye view features, and inputting the decoded bird's-eye view features into a detection head for target detection to obtain a target detection result. In this way, through multi-modal fusion complementation and semantic space mapping, the problem that existing multi-modal data fusion only stays at a data layer for direct splicing or shallow feature level fusion can be avoided, and the space-time isomerism and semantic difference between multi-modal data are overcome. The consistency and robustness of target detection in the ship navigation process are improved under the complex sea condition, and the reliability and accuracy of the detection result are ensured.

The invention discloses a multi-dialect large model adaptation method and system oriented to a Gluggin language, and the method comprises the steps: obtaining corpus data and model constraint information of each Gluggin dialect region, building a dialect mapping matrix, and positioning key semantic difference points; characterization offset features are extracted, contrast training sample pairs are constructed, and vector space analysis is implemented to determine a priority adjustment layer; identifying dialect difference features in the model constraint information to generate layered coding parameters, generating a layered representation structure through collaborative matching of a priority adjustment layer and the layered coding parameters, and determining a weight adjustment boundary; determining a parameter distribution density by combining a weight adjustment boundary and a priority coefficient, generating an optimal migration weight, extracting a negative migration node, converting the negative migration node into a dialect boundary mark, and constructing a gradient propagation path; and detecting the confusion expression set, generating representation decoupling configuration, executing progressive collaborative optimization based on the representation decoupling configuration and the gradient propagation path to output a multi-dialect adaptation scheme, and realizing accurate adaptation and collaborative optimization of the multi-dialect of the Gu-language family.

This invention belongs to the technical field of backdoor defense in neural networks. To improve the security of existing graph neural networks in backdoor attack scenarios, it proposes a graph neural network backdoor defense method and system based on dual semantics and neighborhood anomalies. The method involves constructing feature semantic representations and structural semantic representations of nodes in the graph data, and modeling these representations. It compares the representation-level and relation-level differences between the feature semantic representations and structural semantic representations of nodes; analyzes the degree of representational deviation or relational consistency of nodes relative to their neighboring nodes; fuses the cross-semantic difference information of nodes with local anomaly scores to obtain a comprehensive anomaly score for each node; and constructs node training weights based on the comprehensive anomaly score to train the model. This invention improves the security and robustness of graph neural networks in backdoor attack scenarios.

The invention discloses a small sample industrial anomaly detection method based on prompt learning. The method comprises the steps that (1) word splitting is carried out on a normal prompt text and an abnormal prompt text used for industrial image anomaly detection to obtain character sequences, word boundary marks are attached to the tail of the character sequences to form initial sub-word units, and the abnormal prompt text is an industrial image judgment text to be subjected to anomaly detection; performing adaptive byte pair code merging operation on the sub-word units to obtain merged sub-word units; 2) coding the sub-word units to obtain a normal prompt text context feature vector and an abnormal prompt text context feature vector; the method comprises the steps of (1) obtaining a context feature vector, (2) carrying out importance evaluation and semantic non-linear expression enhancement to obtain a context feature vector, (4) obtaining an abnormal confidence coefficient and a semantic difference score between the abnormal confidence coefficient and the context feature vector according to a difference value, and carrying out calculation to obtain a comprehensive abnormal score, and (5) judging whether categories described by an industrial image and an abnormal prompt text are consistent or not according to the comprehensive abnormal score.

A method, apparatus, and electronic device for detecting abnormal input to an intelligent agent based on a large model are disclosed, relating to the fields of large models, intelligent agents, and artificial intelligence. The method includes: acquiring input content for the intelligent agent; wherein the intelligent agent is associated with a first large model, which outputs model processing results based on the input content; performing semantic understanding on the input content and the service scope description information corresponding to the intelligent agent based on a second large model, and determining the abnormal detection result corresponding to the input content based on the semantic differences between the input content and the service scope description information; wherein the service scope description information corresponding to the intelligent agent is determined based on the intelligent agent's configuration file, which at least describes the definition of the service functions provided by the intelligent agent. This effectively identifies abnormal inputs using semantic transformation techniques such as synonym replacement, sentence reconstruction, and the addition of interfering words, thereby improving the accuracy and reliability of abnormal detection of the intelligent agent's input content.

The invention belongs to the technical field of artificial intelligence, and relates to a scientific and technological knowledge extraction method and system based on context awareness of a large language model. The method comprises the following steps: 1) based on query provided by a user, performing deep semantic analysis and context matching analysis on each information source in an initial information source set obtained by retrieval by utilizing a large language model so as to execute fine-grained content screening, and obtaining filtered content; 2) scoring the contribution degree of the filtered content to the knowledge container by adopting a scoring function fusing semantic difference degree and information gain; 3) pruning the initial information source set based on a scoring result to form fusion content; 4, the knowledge container and the fusion content are fused based on a large language model to obtain a final knowledge base, the problem of information overload of traditional retrieval is solved, the defects that the context of existing LLMs is limited, updating lags and the like are overcome, and the depth, timeliness and situation adaptation capacity of scientific and technological knowledge extraction are remarkably enhanced.

The invention relates to the technical field of data processing, in particular to an intelligent planning data processing method and system based on BIM and GIS, and the method comprises the steps: obtaining a plurality of planning element data sets, a to-be-processed BIM element data set and a to-be-processed GIS element data set, calculating the semantic similarity weights of attribute names in the BIM and GIS element data sets, and calculating the semantic similarity weights of the attribute names in the BIM and GIS element data sets; according to the BIM and GIS data fusion method, the BIM and GIS data are classified, attribute names are clustered, one-way semantic similarity is calculated, then weighted fusion is carried out to obtain comprehensive semantic similarity of element class pairs, optimal matching element class pairs are screened, finally, the attribute names of the element class pairs are fused to generate a fusion planning data set, the fusion planning data set is fed back to a central system, and efficient fusion and optimization processing of the BIM and GIS data are achieved. According to the method, accurate matching and attribute fusion of BIM and GIS element classes are realized by quantifying attribute semantic differences and accurately calculating element class specificity and bidirectional semantic similarity, the problems of heterogeneous data evaluation distortion, wrong matching and fusion deviation are effectively solved, and reliable support is provided for planning data accurate application.

A method and apparatus for Chinese medical named entity recognition based on multi-level adaptive semantic enhancement, the method comprising: (1) representing Chinese text as T={C1、C2、···、C N}, construct character C i Features, including character features, boundary features, radical features, and pinyin features; (2) Character-level C is generated through the ERNIE-Health pre-trained model. i The features are transformed into vector representations, including character feature vectors e. c Boundary eigenvector e b , radical feature vector e r Pinyin feature vector e p (3) Input the four character-level features into the character-level adaptive semantic enhancement module. Use convolutional layers to compress the character features, perform nonlinear transformation through gating mechanism and ReLU activation function, dynamically adjust semantic weights, and use multilayer perceptron for decompression to obtain enhanced character-level features; (4) Input the enhanced character-level features into the sentence-level adaptive semantic enhancement module, and adaptively learn the contribution of different characters in the sentence through compression and decompression mechanism; (5) Input the enhanced features after multi-level adaptive semantic enhancement module into BiLSTM-CRF module for label prediction. This invention can better capture semantic differences in context, solve the limitations of existing methods in feature weight allocation, and improve the overall performance of CNER task.

The invention discloses a layout-guided document image adaptive enhancement method, and mainly solves the problems of poor readability, loss of detail structures and poor matching performance of repaired texts due to neglecting of internal structural semantic differences of document images in the prior art. Comprising the following steps: 1) inputting a low-resolution LR document image, and obtaining a semantic segmentation mask graph M based on a deep learning detection segmentation technology; 2) performing feature extraction and up-sampling operation on the LR image by using an existing generator to obtain an intermediate feature; 3) constructing an L-SPADE module as a conditional feature modulator, generating a spatial adaptive modulation parameter graph by using M through network learning, and performing affine transformation on a normalized feature graph; and 4) training the image restoration generator embedded with the L-SPADE module, and finally obtaining a restored image.According to the method, the document image restoration precision can be improved, the restoration effect is effectively aligned with a downstream task, and the overall efficiency of document digital processing is remarkably improved.

The embodiment of the application discloses a code rate control method, device and electronic equipment, relates to the technical field of video coding, and can stabilize code rate and improve the accuracy of code rate control. The method comprises the following steps: acquiring two continuous images, wherein the two images comprise a target frame to be encoded and a previous frame of the target frame; calculating a semantic difference degree between the target frame and the previous frame, wherein the semantic difference degree is the distance between feature vectors of original images of the target frame and the previous frame, and the feature vectors are extracted through a convolutional neural network; and based on the semantic difference degree, a code rate control strategy is selected to encode the target frame. The application is suitable for code rate control processing in a video coding scene.

The application discloses a kind of image adversarial sample detection methods based on sample semantic difference, it is related to network security technical field, comprising the following steps: training condition generation model, original sample is classified and reconfiguration, generates the reconfiguration sample corresponding to original sample, original sample and reconfiguration sample are as positive sample pair, the rest sample combination is as negative sample pair, the semantic extractor is trained by contrast learning, the semantic feature of original sample and reconfiguration sample is extracted in combination with semantic extractor and the similarity feature value of semantic feature is calculated, the similarity feature value is as input feature, training isolation forest model, then the semantic feature of the to-be-tested sample is extracted according to the same method after reconfiguration, the similarity feature value is substituted into isolation forest model detection and obtains the adversarial sample in the to-be-tested sample;The adversarial sample detection method of the application can generate image according to label, does not rely on the network structure and intermediate layer output of target classifier, can be effectively applied to black box scene, significantly expand its application range.

This invention discloses a video semantic-driven resource allocation method in the Internet of Vehicles (IoV), belonging to the IoV field. First, an IoV system model for video semantic communication between the vehicle and edge server is constructed. Then, experimental data is statistically analyzed and combined with regression methods to construct a video semantic-driven resource allocation guidance model. Next, the detection accuracy per unit vehicle is calculated using the video bitrate, and maximizing the detection accuracy per unit vehicle is used as the optimization objective to construct a video semantic-driven resource allocation optimization model in the IoV. The model is simplified by analyzing the monotonicity of the objective function. Based on the video semantic-driven resource allocation optimization model, a reinforcement Q-learning algorithm model is constructed. The resource allocation optimization model is trained and solved by constructing a state space, action space, and environmental feedback, and executing actions, observing states, and obtaining rewards. This invention considers the semantic differences of different videos and the non-steady-state channel conditions of the IoV, optimizes the spectrum allocation of the video semantic task from the vehicle to the edge server, and maximizes the average target detection accuracy.

This invention relates to the field of semantic processing technology, specifically to a large-scale adaptive training method and system based on task semantic awareness. The method includes the following steps: acquiring text structure information, extracting batch semantic vectors to determine consistency, tracking semantic differences across steps to form a trajectory, merging stage boundaries, locating abnormal nodes and verifying loss gradients, and mapping parameter tuning instructions into an adaptive training control structure. In this invention, through structured semantic expression and a real-time feedback mechanism, the training process can be dynamically adjusted to adapt to changes in task semantics, avoiding semantic shift and inconsistency issues. Precise labeling of semantic states and monitoring of change trajectories enable the model to adjust training strategies in a timely manner at different semantic stages, thereby improving training efficiency and stability. Accurate location of key nodes allows training to focus on training steps with significant impact, avoiding ineffective training. Through semantic-driven parameter adjustment and training control, the model's adaptability to complex tasks and resource utilization are improved.

The invention discloses a text-image pedestrian re-identification method oriented to semantic fuzziness and excessive decision confidence, and belongs to the technical field of pedestrian re-identification. The method mainly comprises the steps that based on a distributed feature alignment module, image and text modal features are coded into Gaussian distribution, a feature center and an intra-class semantic change range are jointly expressed through a mean value and a variance, and the modeling ability of cross-modal semantic difference and visual diversity is improved; the uncertainty penalty alignment module based on evidence deep learning explicitly models the uncertainty in the image-text matching process by introducing Dirichlet distribution and subjective logic theories, and applies adaptive penalty to a high-uncertainty matching relationship, so that the excessive confidence of the model on a mismatching result is inhibited. According to the method, a probability distribution-based feature expression mode and an uncertainty perception training mechanism are adopted, so that the matching accuracy and generalization ability of a cross-modal retrieval system in view angle change, semantic fuzziness and background interference scenes are remarkably improved.

The invention relates to an inference method and device, computer equipment, a storage medium and a computer program product. The method comprises the steps of obtaining query objects, performing semantic transformation processing on the query objects to further obtain a plurality of query objects, and inputting the query objects into different reasoning models to obtain a plurality of reasoning results output by the reasoning models for the query objects; for any two reasoning results, determining a semantic difference between the two reasoning results, and determining a logical relationship between the two reasoning results; based on a target knowledge base corresponding to the query object, determining the fact confidence of each reasoning result; and determining a target reasoning result from the reasoning results based on the semantic difference of the reasoning results, the logic relationship and the fact confidence. The method can improve the reliability of the reasoning result.