307 results about "Categorization" patented technology

A method and system for automated support sub-topic classification using large language models (LLMs). The method includes collecting user transcripts and LLM-generated summaries, performing unsupervised learning to identify common themes, generating sub-topics, and creating labeled datasets. A supervised learning model is trained to categorize user transcripts into the identified sub-topics. The system performs categorization of new user queries and generates appropriate action responses.

Techniques for training and using machine learning models for sensitivity detection. A method for sensitivity detection training includes fine-tuning a language model by iteratively applying the language model to prompts and adjusting weights of the language model. The prompts indicate classifications for a set of first resources and characteristics of an entity. The fine-tuned language model is queried with respect to classifications of a set of second resources. The fine-tuned language model is queried using prompts indicating the second classifications and data indicating characteristics of an entity, where outputs of the language model include a sensitivity for each of the second classifications. Training data including the second classifications is labeled based on the sensitivities output by the language model. A sensitivity detection machine learning model is trained using the labeled training data set such that the trained sensitivity detection machine learning model is configured to output sensitivities for resource classifications.

The application discloses a resource consumption abnormal object identification method and related device, which is applied to the field of artificial intelligence. By obtaining unlabeled samples and positive samples, determining target feature dimensions, inputting the unlabeled samples and the positive samples into a semi-supervised learning framework to iteratively train a classification model, obtaining multiple identification feature values corresponding to the unlabeled samples output by the classification model in the iterative training process, and determining resource consumption abnormal objects, the resource consumption abnormal object identification process under a small amount of positive samples is realized. In the training process, the positive samples in the unlabeled samples are continuously mined heuristically for supplementation and added to the next round of iteration, effectively solving the sample imbalance problem in the identification scene and improving the accuracy of resource consumption abnormal object identification.

The application discloses a target detection model evaluation system and method based on light weight, relates to the technical field of target detection, and ensures data specification by acquiring a data set and converting the data set into a YOLO format, divides sample types by matching scores, and dynamically adjusts weights by using a segmented function to optimize classification; subsequently, an SPPF-Attention unit is constructed by combining a 1*1 point-by-point convolution, a depth-by-depth convolution and an MHSA module, and a light weight model is built; samples are expanded by Mosaic data enhancement, and the training process is controlled by using a segmented function; finally, the performance of the model is evaluated from multiple dimensions of parameter quantity, calculation quantity, mAP value and inference time consumption, the organic unification of model light weight, high-precision detection and high-efficiency inference is realized, the training efficiency and generalization capability are improved, and the multiple requirements of actual application are met.

Methods for training and using an artificial intelligence (AI) model to identify a waste classification for a waste object. The method to train the AI model includes generating training data and providing the training data to train the AI model on (i) a set of training inputs and (ii) a set of target outputs. A first training input includes first data representing first images of a first solid waste. A second training input includes second data representing second images of a second solid waste. A first target output identifies a first waste label corresponding to landfill waste. A second target output identifies a second waste label corresponding to recyclable waste.

The application discloses a smart contract vulnerability detection method based on static analysis and large language model fusion, comprising the following steps: obtaining a smart contract source code to be detected, using a static analysis tool to analyze the pretreated smart contract source code, and generating a static analysis abstract; the smart contract source code and the static analysis abstract information are spliced according to a preset prompt template, and then input into a large language model to obtain a vulnerability detection result. The application effectively integrates the structured priori knowledge of static analysis into the input features, significantly improves the recognition ability of the model for seven core Solidity vulnerabilities, can accurately adapt to the classification task while maintaining the efficiency, and effectively prevents the false negatives, false positives and model security risks.

The present application relates to a kind of generalization adversarial sample group generation method based on easily confused class feature injection, belong to the technical field of adversarial attack of computer vision field.It introduces the class disturbance update momentum in the direction of easily confused class feature vector in the process of generating class-oriented generalization disturbance sample group, the class disturbance update momentum makes the adjacent easily confused class of adversarial example faster update, to realize the generalization attack to class.The present application only integrates the features of easily confused class in the same class image, forms disturbance template, realizes class-oriented generalization disturbance sample group;Variable step size iterative attack means is used to explore easily confused class in the process of neural network classification, realize easily confused feature injection, while introducing class momentum parameter in the process of iterative attack, improve attack efficiency and alleviate the over disturbance shortcomings of generalization disturbance.

To provide a device and method that can provide an environment in which participants in a "space" for thinking and discussing can consider things from a wide variety of perspectives. [Solution] According to this embodiment, the map classifies fields related to society in the first direction, classifies the level of each field in the second direction (social domain in the Arctic direction), and classifies words related to individuals in the third direction (personal domain in the Antarctic direction). Multiple nodes with semantic content are placed in the social domain and the personal domain according to the words (attribute information) at their placement location.

A concept tagging model is trained based at least in part on a result of a large language model used to analyze training data. Text content of a website is received. The concept tagging model is applied to the text content to identify one or more concept terms in the text content. An embedding model is applied to the identified one or more concept terms to determine one or more embeddings. Using the one or more embeddings, a closest matching granular category in a taxonomy of categories is identified for the website.

The application relates to the technical field of artificial intelligence, and discloses a multi-dimensional cognitive state analysis method and device, electronic equipment and a storage medium. The method comprises the following steps: constructing a cognitive state classification system and a domain sample data set; determining target dialogue text; learning hierarchical structure features of the target dialogue text on a hyperbolic space based on the domain sample data set, and determining cognitive representations of cognitive state dimensions presented by the target dialogue text in the hyperbolic space; the cognitive state dimensions at least comprise four dimensions of emotion, thinking mode, position and intention; performing semantic training on the cognitive representations based on a preset large language model, and reasoning out a cognitive state image satisfying cognitive analysis text constraints and cognitive hierarchical structure constraints; the hierarchical overlap problem of the cognitive state is solved by using exponentially increasing space capacity, and geometric structure knowledge in the hyperbolic space is injected into the large language model, so that the joint modeling and reasoning capability of the model for the multi-dimensional cognitive state in a complex social scene is significantly improved.

This invention discloses a lightweight few-shot incremental learning method for power line defects. The method includes: first, constructing an inspection image dataset containing basic and incremental categories, and designing a lightweight feature extraction network; training the network using the basic category data to generate prototype vectors for each category, forming an initial prototype classification model; in the incremental stage, constructing incremental prototypes for a small number of samples of the new category, and introducing a prototype transfer mechanism based on inter-class tension, adaptively adjusting the new prototype based on the geometric similarity and discrimination conflict strength between the new and existing prototypes to alleviate category conflicts; simultaneously, introducing an adaptive feature fusion mechanism that perceives discrimination uncertainty, dynamically fusing the discrimination results of the main and auxiliary branches to improve recognition stability. This invention, employing the above-mentioned lightweight few-shot incremental learning method for power line defects, can achieve efficient and stable power line defect recognition in scenarios with few samples and continuously expanding categories.

The application discloses a personalized pet nutrition and health care product recommendation method and system, and relates to the field of artificial intelligence. The specific steps comprise the following steps: collecting pet related information and labels from multiple source systems; performing normalization and nonlinear interaction feature construction on numerical features, and performing adaptive embedding and category cross on category type features to obtain enhanced original vectors; screening features by using label association information, and performing clustering fusion to obtain a fusion feature matrix; adopting an improved feature cross network, introducing low-rank decomposition and pet health knowledge prior to construct a classification model; constructing a loss function introducing a rule constraint regular term, training the model and updating parameters; and recommending nutrition and health care products to pets by using the trained model. The application can accurately capture the complex relationship between features, and significantly improve the accuracy of model prediction.

This invention belongs to the technical field of natural language processing and dialogue systems. It discloses a multi-target dialogue recommendation method based on hierarchical prompt optimization, including hierarchical target label modeling, learnable prompt embedding initialization, graph attention enhancement mechanism, multi-task learning framework, and response generation model. Hierarchical target label modeling encodes the hierarchical relationships between target labels into a graph structure. Learnable prompt embedding initialization integrates label semantics and deep hierarchical information into the encoder input space. The graph attention enhancement mechanism enhances the label vector structure perception ability through information transmission and aggregation between label prompts. The multi-task learning framework jointly optimizes masked language modeling and hierarchical multi-label classification. The response generation model uses the predicted target as a control signal to guide generation. This method effectively models the hierarchical dependency relationship between target types and target entities, enhances the structure perception ability of label vectors, maintains the model's language understanding ability, and improves the accuracy of target prediction.

Systems, methods, and apparatuses are disclosed for detection and prevention of adversarial attacks against large language models. Techniques may include receiving an input associated with a target large language model, analyzing the input with a pre-trained classification algorithm to determine a first deconstruction process to be applied to the input, and modifying the input with a first deconstruction model using the determined first deconstruction process. Techniques may also include determining a score of a likelihood of the input being adversarial based on an output of the first deconstruction model and by applying a classification model and updating at least one of the first deconstruction model or the classification model based on the score.

This disclosure provides systems, methods, and devices for automatic product classification using deep learning and generative artificial intelligence (AI) models for a harmonized system (HS) product classification. A method includes generating embeddings based on an input dataset. The method includes applying a deep learning model to the embeddings to produce a prediction set including classifications corresponding to the embeddings. The method includes converting the classifications to a first set of similarity metrics. The method includes determining a second set of similarity metrics based on the embeddings using a semantic similarity model. The method includes generating a third set of similarity metrics based on an output of the semantic similarity model. The method includes outputting a ranked set of predictions corresponding to the input dataset based on the first set of similarity metrics, the second set of similarity metrics, and the third set of similarity metrics.

According to the present disclosure there is provided systems and methods for categorising an item. The system comprises processing means configured to: obtain a first image comprising an item; extract a portion of the first image, the portion comprising the item; determine, using the extracted portion of the first image, one or more characteristic features of the item using a first machine learning model; determine a colour value associated with the extracted portion of the first image; and categorise, using a second machine learning model, the item according to one of a plurality of predetermined item categorisations based on at least the determined one or more characteristic features and the colour value, wherein the second machine learning model is different to the first machine learning model.

Certain aspects of the disclosure provide techniques for categorization by a device. An example method includes receiving input information regarding a plurality of classification targets and a plurality of categories for classification of the plurality of classification targets; generating a plurality of embeddings for the input information using a first model, the plurality of embeddings including: a set of first embeddings associated with the plurality of classification targets, and a set of second embeddings associated with the plurality of categories; determining that a similarity score for the set of first embeddings and the set of second embeddings fails to satisfy a threshold; generating, based on the similarity score failing to satisfy the threshold and using a graph neural network (GNN), a classification of the plurality of classification targets in accordance with the plurality of categories; and outputting information regarding the classification.

The application provides a training method and system of an emotion classification model for multi-modal physiological signals, relates to the technical field of cross between artificial intelligence and biomedical signal processing, and the method comprises the following steps: collecting multi-modal physiological signals of a plurality of sample users; processing the multi-modal physiological signals of the plurality of sample users to obtain multi-modal physiological vectors of the plurality of sample users; using the multi-modal physiological vectors of the plurality of sample users to perform autoregressive pre-training on a large language model; using the multi-modal physiological vectors of the plurality of sample users and text vectors corresponding to emotion classification prompt texts to fine-tune the large language model subjected to the autoregressive pre-training, and obtaining an emotion classification model. In the process of emotion prediction, the problems of multi-modal signal mode loss, inconsistent sampling rates and non-uniform channel numbers can be effectively overcome, and the accuracy of emotion prediction is improved.

The application relates to the technical field of retrieval enhancement generation evaluation, and provides a retrieval enhancement evaluation method and device based on multi-round dialogue context perception, which comprises the following steps: a first large language model is constructed; the first large language model is used for question rewriting according to a question and context; historical multi-round question and answer data of real users is extracted; a multi-round evaluation set is generated based on the historical multi-round question and answer data, a knowledge base document, a second large language model and a RAG system; the second large language model is used for summarizing a business classification, a question type and an interaction paradigm of each multi-round question and answer; the first large language model is used for converting the multi-round evaluation set into a single-round question set; and the RAG system is evaluated by using evaluation indexes, the multi-round evaluation set and the single-round question set. The application realizes automatic generation of the multi-round evaluation set, reduces the manual labeling cost, reduces the construction difficulty of the multi-round evaluation set, and realizes evaluation of the multi-round dialogue RAG task.

The application relates to the technical field of artificial intelligence, in particular to a training method and device for a fault diagnosis model of a variable-configuration aircraft. The method comprises the following steps: obtaining an initial data set, wherein the initial data set comprises initial flight sequence data, a fault type and initial configuration parameters; constructing a training data set according to the initial data set, wherein the fault type and an initial symbol sequence are labeled as labeled information of the training data set; inputting the training data set into an initial fault diagnosis model to obtain predicted information output by the initial fault diagnosis model; determining a target loss value according to the predicted information and the labeled information; and updating model parameters of the initial fault diagnosis model based on the target loss value to obtain a target fault diagnosis model. The training method for the fault diagnosis model of the variable-configuration aircraft can realize the collaborative optimization of the feature extraction capability of the model on the flight sequence data, the classification capability of the model on the fault categories and the generation capability of the model on the diagnosis text.

This invention relates to the field of data analysis technology, and in particular to a query analysis method, apparatus, device, and storage medium based on a large language model. The method provides comprehensive and effective data support for enterprise decision-making. It achieves semantic fusion of multi-source heterogeneous data through a unified knowledge graph, breaking the limitations of traditional data silos and providing a comprehensive and accurate business knowledge foundation for the large language model, reducing model illusions from the source. It achieves differentiated analysis processing through task type classification, automatically scheduling different capability modules of the large language model according to user needs, balancing the efficiency of basic queries with the accuracy of in-depth analysis. It obtains the full business context through entity links, limiting the reasoning process of the large language model to the enterprise's real business scenarios, significantly improving the business relevance of the analysis results. It achieves in-depth diagnostic analysis through causal attribution, fully leveraging the logical reasoning capabilities of the large language model and overcoming the limitations of descriptive analysis.

The application provides a large language model reasoning rectification method and device, the method comprises the following steps: obtaining reasoning behavior information and corresponding hidden layer activation state information of a first reasoning step; classifying the reasoning behavior information, and calculating a deviation score of a word element deviating from a potential error behavior category by using a benchmark activation prototype and the hidden layer activation state information; extracting a candidate rectification control vector from a pre-set rectification control vector library in combination with the first reasoning behavior category information, and calculating a dynamic rectification vector by using the candidate rectification control vector and the real-time deviation score; finally, modifying the hidden layer activation state information by using the dynamic rectification vector, and inputting the modified hidden layer activation state information into the large language model. The method and device can block the spread of the error reasoning path without retraining the model, reduce the reasoning cost, and improve the model reasoning accuracy.

The invention relates to the technical field of artificial intelligence and human-computer interaction, and discloses a rehabilitation training adjustment method and system based on multi-modal feature fusion and a storage medium, and the method comprises the steps: obtaining voice, face and physiological signals, constructing a multi-modal data stream, extracting features, and outputting a speech error classification identifier and a physiological signal quality index; splitting the facial action net displacement into healthy and affected sides, correcting affected side features based on healthy side features, and constructing visual representation; fusing each feature to output a comprehensive state vector; updating a dynamic baseline in a task gap, mapping a state vector, and screening to obtain an effective action set; calculating a composite reward and storing the composite reward in an experience playback pool to update the strategy network; and based on the action set, re-weighting the large model output probability, and generating a target interaction corpus. According to the method, the state sensing precision is improved by correcting the deviation of the affected side through the uninjured side, and a closed loop of self-adaptive adjustment of the rehabilitation difficulty and safe interaction text generation is realized by combining the dynamic baseline and penalty mechanism optimization reinforcement learning.

This invention provides a training method for a psychological crisis early warning model and a psychological crisis early warning method, relating to the field of natural language processing technology. Through the psychological crisis early warning model, implicit expression data undergoes style transformation to obtain style transformation results. These results are then used to classify psychological crises, yielding target crisis classification results. The psychological crisis early warning model is then trained using the style transformation results, the target crisis classification results, and the explicit expression data corresponding to the implicit expression data. This training enables the target psychological crisis early warning model to possess strong style transformation and psychological crisis classification capabilities, thereby reducing the difficulty of detecting implicit psychological crisis data and improving the accuracy of detection.

The application discloses a class-incremental continual learning method based on classification layer unification and improved EWC, a neural network is trained through a training set to obtain a recognition model, and the recognition model is used for recognizing the category of things in a picture, the training comprises old task training and at least one new task training, and in the new task retraining process, the model parameters of the old task are taken as a base point to limit the parameter offset. The class-incremental continual learning method based on classification layer unification and improved EWC solves the new category bias problem and improves the average accuracy of the model.

Hallucinations in LLMs pose significant challenges and it is hard to accurately identify hallucinated predictions of LLMs in the absence of references. The present disclosure utilizes sample responses from LLMs and classifies them using a model trained on Natural Language Inference (NLI) scores. The plurality of NLI scores include an entailment, a neutrality and a plurality of contradiction scores. Post computing NLI scores, an average NLI score is computed based on the plurality of NLI scores. Further, a plurality of individual response predictions are obtained by classifying the plurality of responses based on the NLI scores using a hallucination classifier. Simultaneously, overall response predictions are obtained and a confidence score is computed for the overall prediction of hallucination classifier Finally, an optimal response is predicted based on the confidence score associated with the overall prediction of the hallucination classifier and the plurality of NLI scores.

The application provides an industrial defect sample classification method and system based on prototype learning, comprising: acquiring an industrial defect image; based on a pre-constructed industrial open set defect detection dataset and the industrial defect image, generating an unknown class feature sample set through Monte Carlo sampling; based on the unknown class feature sample set, generating a reliable pseudo label through an adaptive pseudo label generation mechanism; based on the reliable pseudo label, performing joint prototype learning on known class samples and the unknown class feature sample set in the industrial open set defect detection dataset to construct known class feature prototypes and unknown class feature prototypes; and classifying the industrial defect image according to the known class feature prototypes and the unknown class feature prototypes; and the application performs joint prototype learning on the known class samples and the unknown class feature sample set in the industrial open set defect detection dataset based on the reliable pseudo label, can realize collaborative learning of the known class and the unknown class feature, and adapt to complex and diverse defect classification scenarios under the industrial open set.

The present application relates to the technical field of artificial intelligence, and provides a chemical reaction classification model training method and device based on negative example enhancement, which comprises the following steps: obtaining initialization chemical reaction negative example samples by inversely rewriting first chemical reaction positive example samples according to an initial negative example template library, training a to-be-trained model by using the first chemical reaction positive example samples and the initialization chemical reaction negative example samples, obtaining an intermediate state model; obtaining first prediction error samples on the first chemical reaction positive example samples and second prediction error samples on second chemical reaction positive example samples by using the intermediate state model; updating the initial negative example template library to obtain a target negative example template library according to the prediction error samples; generating target chemical reaction negative example samples according to the target negative example template library, and continuing to train the model according to the target chemical reaction negative example samples and the first chemical reaction positive example samples until a preset condition is met, so as to realize the cooperative promotion of template updating and model training, and improve the prediction accuracy and generalization ability of the trained chemical reaction classification model.