942 results about "Text entry" patented technology

The invention discloses a large-model real-time voice text intention recognition method and system based on text prediction, and the method comprises the steps: obtaining the real-time voice data of a user, carrying out the real-time voice recognition processing through a streaming voice recognition interface, and obtaining a part of transcriptional text; inputting the partial transcription text into a mask language model for text prediction, and generating a plurality of high-credibility complete sentence candidates; based on the complete sentence candidates, the complete sentence candidates are input into a large language model in parallel for intention recognition, a corresponding intention result is obtained, and a mapping relation between the candidate sentences and the intention recognition result is established; and obtaining a sentence completely expressed by the user, calculating the similarity between the complete actual sentence and a plurality of high-credibility complete sentence candidates through a multi-level text similarity algorithm, selecting the candidate sentence with the highest similarity score, and directly obtaining a corresponding final intention recognition result based on the mapping relationship. The objective of the invention is to solve the technical problem of high response delay of an existing voice intention recognition system.

The invention relates to the technical field of multi-modal large language model training, in particular to a prompt construction method and system for a multi-modal large language model, computer equipment and a medium. The method comprises the following steps: extracting a key frame set from an input video stream; executing a motion reconstruction process on the video stream to generate motion track information; and performing visualization processing on the motion track information to generate a track visualization graph. Performing space-time correlation coding on the key frame set and the motion track information to generate an enhanced key frame; a multi-modal prompt is constructed in a mode of integrating visual input and text input, and the multi-modal prompt is input into a preset multi-modal large language model for spatial reasoning. Through the mode, the technical problem that an existing prompting method is difficult to give consideration to the spatial reasoning precision and the calculation efficiency is solved, efficient and accurate spatial reasoning of the multi-modal large language model is achieved, and the calculation efficiency, the reasoning precision and the environmental adaptability of the model are improved.

The invention discloses an Agent-based education evaluation method and system, and relates to the technical field of artificial intelligence and education evaluation, and the method comprises the steps: obtaining historical evaluation data and knowledge graph labeling data of students, constructing a student knowledge state matrix, calculating a mastering probability value and a forgetting attenuation coefficient of each knowledge point, and generating an evaluation question recommendation sequence; inputting the question text and the student answering text into a multi-Agent collaborative evaluation model to generate a multi-dimensional scoring result; dynamically updating the knowledge point mastering probability value of the student in the continuous answering process, calculating the mastering degree change rate of each knowledge point node, constructing an adaptive difficulty adjustment function, and generating a question difficulty parameter and a knowledge point coverage range parameter of the next round of evaluation; and taking the question difficulty parameter and the knowledge point coverage range parameter as constraint conditions, screening a candidate question set meeting the conditions from a question bank, and generating a self-adaptive evaluation path and a capability diagnosis report. The invention also discloses a method.

The invention discloses a knowledge destruction attack method and device based on an RAG system and a medium, and relates to the technical field of internet security, and the method comprises the steps: inputting a target question and an error answer into the RAG system, and generating an initial confrontation text; performing multiple rounds of iterative optimization processing on the initial adversarial text to obtain a target adversarial text; inputting the target adversarial text into a knowledge base corresponding to the RAG system; the RAG system responds to a question demand input by a user, and retrieves and outputs a question answer corresponding to the question demand from the knowledge base; and inputting the question demand and the question answer into a large language model, so that the large language model outputs a wrong answer corresponding to the target question. The method and the device are used for solving the problems of low output result precision, poor attack effectiveness and low concealment when knowledge destruction attack is carried out based on an RAG system in the prior art, and the precision of the output result is improved under the condition that the knowledge destruction attack is effectively carried out with high concealment.

The invention relates to a KV cache optimization method and device, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: calculating a key vector and a value vector corresponding to each element in a text input sequence input into a large language model; through a multi-head potential attention mechanism, performing low-rank joint compression on the key vector and the value vector to obtain a potential vector, and storing the potential vector in a KV cache space; based on a scaling law, determining an optimal compression dimension, regenerating an adaptive potential vector and updating a KV cache space; for the same text input sequence, generating corresponding query vectors, and grouping the query vectors according to a preset grouping rule; calculating a semantic association weight between each group and the correspondingly called potential vector, and taking the semantic association weight as a group attention calculation result; in the reasoning process, potential vectors and grouping attention calculation results are calculated to calculate attention weights. By adopting the method, the storage requirement of the KV cache can be further reduced.

Exemplary system and methods for processing a natural language query in an edge computing system are disclosed. A processor of the computing system receives a natural language textual input as a query from a user interface and receives one or more containers of documentation over a communication channel. The processor generates a query embedding vector from the textual input. The processor extracts text from the received container and generates text chunks of specified length from the extracted data. Text embeddings are generated from the text chunks and stored in memory for a specified period. The query embeddings are compared with the text embeddings to determine relevant context information. The processor passes the relevant context information and the query through a trained neural network to generate a response. The response generated by the trained neural network is formatted and output to a user interface.

The invention relates to the technical field of image processing, in particular to a directed target detection method and device, electronic equipment and a computer storage medium, and the method comprises the steps: obtaining a target detection task containing a target detection image and a corresponding target annotation text, inputting the target detection image into a preset image classification model, and outputting a target image feature; using the feature pyramid network to construct a target multi-scale feature map based on the features, inputting the target annotation text into a preset language processing model, outputting target text features, performing feature fusion on the target multi-scale feature map and the target text features to obtain target fusion features, and inputting the target fusion features into a preset target detection model to obtain a target detection result. And outputting the position and the category of the target in the target detection task by the model. According to the invention, the accuracy and effectiveness of target detection are improved by fusing the image and text features.

The invention provides a false news detection method and device, electronic equipment, a product and a storage medium, and relates to the technical field of information processing. The method comprises the following steps: determining a similarity result between a news text and each piece of false news in a preset false news set; determining whether the news text is a to-be-detected text or not based on a comparison result of the similarity result with the maximum similarity in the similarity results and a preset similarity threshold value; if the news text is the to-be-detected text, inputting the news text into a false news detection model to obtain a false news detection result output by the false news detection model; if the news text is not the to-be-detected text, determining the news text as a false news text; wherein the false news detection model is obtained by training based on the sample news text and the false news detection result label corresponding to the sample news text. According to the method, occupation of system resources can be reduced, and the efficiency of false news detection is remarkably improved while high detection accuracy is kept.

The invention discloses a text risk identification method and device fusing a fine-grained knowledge graph and deep learning, and the method comprises the steps: carrying out the combined entity identification of a standardized text set, and extracting evaluation dimensions and viewpoint contents to form an entity pair set; performing sentiment polarity analysis on the entity pair set to obtain sentiment tendency, and constructing a triple (evaluation dimension, sentiment tendency and viewpoint content); constructing a fine-grained emotional knowledge graph based on the triple, and optimizing the fine-grained emotional knowledge graph by adopting structure perception and joint loss; based on the to-be-recognized text, retrieving from the atlas to construct an enhanced text, and inputting a fusion atlas perception reasoning model comprising a BERT coding layer, an atlas attention module, a cross-layer residual attention mechanism and a KAN reasoning layer to obtain a text risk recognition classification result. According to the method, an emotion structure modeling and semantic reasoning mechanism is introduced, the accuracy, interpretability and robustness of text risk recognition are improved, and the method is suitable for multi-field text analysis tasks under complex contexts.

A method for generating photorealistic 4D scenes from text inputs is disclosed. The method utilizes a text-to-video diffusion model to generate a reference video and a freeze-time video. A canonical 3D representation is reconstructed using deformable 3D Gaussian Splats (D-3DGS) based on the freeze-time video. Temporal deformations are learned to capture dynamic interactions in the reference video. The method employs a novel Score Distillation Sampling strategy combining multi-view and temporal aspects to enhance consistency and robustness. The resulting 4D scenes feature multiple objects interacting with detailed background environments, viewable from different angles and times. The method enables flexible camera control and integration with augmented and virtual reality applications. Some examples include features such as image-to-4D generation.

The invention relates to the technical field of computers, in particular to a multi-modal sample data generation method and device, electronic equipment and a storage medium, and the method comprises the following steps: for a target webpage, obtaining a webpage screenshot and element information of UI elements in the webpage; generating annotation information based on the element information of the UI element; taking the annotation information and a preset task cue word as text input, taking the webpage screenshot as picture input, and inputting the text input and the picture input into a large language model to obtain description information which is output by the large language model and is used for describing the webpage screenshot; and taking the webpage screenshot and the description information as sample data for training the multi-modal large language model. According to the embodiment of the invention, the generation efficiency of the sample data of the multi-modal large language model can be improved.

Approaches presented herein provide for the use of reinforcement learning to fine-tune a generative model, such as a motion diffusion model, for a specific objective, such as to generate representations of human motion corresponding to provided text input. A discriminator can be used to guide the training of the generative model. In at least one embodiment, the discriminator can compare the input text and generated motion representation (or embeddings of each) to determine an alignment value or match score, for example, which can then be used to adjust the network parameters or weights of the generative model to improve the alignment between input text and generated motion.

The invention relates to the technical field of natural language processing, in particular to a retrieval system and method based on retrieval enhancement generation, and the method comprises the steps: obtaining a user query text through a user interaction interface, and carrying out the embedding processing of the user query text to obtain a query vector; inputting the query vector into a vector knowledge base, performing approximate nearest neighbor search on the query vector and a storage vector in the vector knowledge base, and inputting the storage vector meeting a preset semantic distance into a candidate text block sequence; analyzing the semantic association strength between each text block in the candidate text block sequence and the query vector; performing descending sorting on each text block in the candidate text block sequence according to the similarity score vector; and inputting the enhanced prompt text into the large language model for reasoning analysis to obtain a retrieval result. According to the method, the text semantic similarity and the context correlation can be considered at the same time in the retrieval process, and the accuracy of the retrieval result is improved through a dynamic weighting mechanism.

The invention provides a content security auditing method and device, electronic equipment and a storage medium, and belongs to the technical field of artificial intelligence, and the method comprises the steps: if it is determined that a scene category of an input text belongs to an exemption security auditing scene set, sending the input text as a question and answer prompt word to a question and answer large model for answering; otherwise, when it is determined that the initial risk level of the input text is a low risk, inputting the input text into the security auditing model to determine the final risk level of the input text; and if the final risk level is low risk, sending the input text as a question and answer prompt word to the question and answer large model for answering. According to the method, the differentiated auditing processing capability of strict interception of high-risk content, intelligent rewriting of medium-risk content and rapid release of low-risk content based on the scene and the user portrait is realized in a high-concurrency scene, the balance between auditing efficiency and safety is effectively balanced, and the safety of the user is improved through double-layer auditing of a safety word library rule and a safety auditing model. And the auditing accuracy and flexibility are improved.

Systems, methods and non-transitory computer readable media for editing user interfaces from free text and / or sketches are provided. Digital data associated with a design of a user interface may be received. A presentation of the user interface based on the design may be caused. A selection of a portion of the user interface may be received. A textual input in a natural language and / or a sketch may be received. The textual input and / or the sketch may be analyzed to determine at least one change to a portion of the design of the user interface corresponding to the selected portion of the user interface. The determined at least one change may be implemented to generate a modified design of the user interface. A second presentation of the user interface based on the modified design may be caused.

In implementations of systems for generating captions, a processing device implements a caption generation service to receive an input for caption generation that includes a text input indicating example language or content for the caption and an action input indicating a desired action. The processing device receives the text input via a user interface. The caption generation service generates a textual prompt for a machine-learning model based on the action input and text input. The machine-learning model uses the textual prompt to generate the caption in a specified structural format. The processing device then causes the generated caption to be presented to a user via the user interface.

The invention discloses an AI file query method based on human-computer interaction, which comprises the following steps: S1, collecting user input, and generating a query text; s2, intention recognition and slot extraction are performed on the query text, a slot set is generated, and standardization processing is performed; s3, utilizing the improved CoSent model to generate a query vector and a document vector, and storing the query vector and the document vector in a vector index; s4, retrieving the candidate document based on the query vector in the vector index, calculating semantic similarity and obtaining a comprehensive score in combination with a matching result; s5, carrying out reordering on the candidate documents; s6, generating a permission set based on the user identity, and executing permission filtering and field masking; and S7, inputting the document fragments passing the permission filtering and the query text into a text generation model, outputting a query result, and recording a log and user feedback. According to the invention, the accuracy and safety of file query are obviously improved.

Disclosed herein are system, method, and computer program product aspects for textual data extraction, generation, and evaluation. Text is input into a first fine-tuned large language model (LLM) to generate an atom (e.g., a textual phrase in a particular category). The atom is input into a second LLM that has been fine-tuned for structured output corresponding to the particular category of information. A logical structure is generated based on a structured output of the second LLM, wherein the logical structure represents the textual phrase of the atom and contains a contextual attribute associated with the textual phrase. The embodiment then stores the logical structure into a knowledge graph as a modifier node having a time-variant attribute (e.g., a timestamp associated with the textual phrase and / or the contextual attribute).

The invention discloses a cross-platform business system integration method based on a knowledge graph. The method comprises the following steps: S1, reading data from a business platform database and generating a field text, an entity identifier and a relation record; s2, inputting the field text into an ALBERT model, and generating a field semantic vector through a field word vector gating layer; s3, forming a triple according to the entity identifier and the relation record, and writing the triple into a knowledge graph; s4, inputting the triple into a TransH model, and generating an entity structure embedding vector through dynamic rotation updating of a relation vector and a hyperplane normal vector; s5, generating an entity embedding vector according to the field semantic vector and the entity structure embedding vector; s6, calculating a plurality of distances based on the entity embedding vectors to obtain the similarity of the cross-platform candidate entity pairs; and S7, performing repeated entity judgment on the similarity, and updating the knowledge graph. According to the invention, accurate alignment of cross-platform service entities is realized, the data fusion efficiency and consistency are improved, and the method is suitable for a multi-source service integration scene.

The invention provides a medical magnetic resonance image reconstruction method and system, and belongs to the technical field of image reconstruction, and the method comprises the steps: obtaining sampling coding and under-sampling k space data; processing the under-sampled k space data through inverse Fourier transform, and converting the k space data into an image domain to obtain under-sampled image data with blurring or artifacts; carrying out preliminary restoration on the undersampled image data based on sampling coding, further applying data consistency operation on the image to obtain a rough image, and taking the rough image as an image condition vector after underspace coding; acquiring a text serving as a cue word, inputting the text into a text encoder, and encoding the text into a high-dimensional semantic embedding vector by the text encoder; and inputting the high-dimensional semantic embedding vector and the image condition vector into a reconstruction model based on the correction flow to predict a full-sampling MRI image, and generating a final reconstructed MRI image.

According to one embodiment, an edge device includes a communication interface connectable to an edge server, a storage unit for storing a dialogue LLM trained to interact with a user for setting a function of the edge device and a database storing reference information related to functions of the edge device. A controller unit receives request text input indicating a user's desired process to be provided using the edge device, then inputs the request text input to a generative AI that is based on the dialogue LLM and can access the database to identify a function of the edge device that can provide the user's desired process. The controller unit outputs confirmation question text to confirm the identified function corresponds to the user's desired process, and then sets the edge device to execute the identified function if confirmed.

The invention provides a multi-mode industrial defect detection method and device, equipment and a storage medium, and relates to the technical field of industrial anomaly detection. The method comprises the steps of performing two-dimensional contour registration on an RGB image and a CAD model of a to-be-detected part, and generating an alignment sample information file; filling a five-segment thinking chain template including positioning, evidence, working procedures, root causes and suggestions to obtain a thinking chain example set file; dividing the thinking chain example set file into a plurality of domain subsets according to part types, each domain subset comprising at least two examples, respectively extracting a text input identification sequence and an image pixel tensor, combining the text input identification sequence and the image pixel tensor to form a data set, and performing supervision fine tuning on the large model by using the data set to obtain a part model; and after obtaining the large model after supervised fine tuning, carrying out reinforced fine tuning by utilizing a reference-query comparison sample pair to obtain a multi-modal industrial defect detection model, and realizing detection of a to-be-detected part. According to the method, the micro defect positioning precision and the cross-production-line generalization ability can be remarkably improved.

The invention belongs to the technical field of virtual reality, and particularly relates to a virtual-real fusion scene automatic construction method based on AIGC script generation, which comprises the following steps: step 1, inputting a script text input by a user into a large language model for structured analysis, generating a corresponding three-dimensional geometric grid, a surface texture map and an initial placement position according to the updated feature vector of each node; 2, collecting an eye image sequence of a user through an eye movement tracking module, and executing boundary fusion processing on rendering results of all rendering areas to output adaptive rendering frames; and step 3, performing collision detection and iterative position correction on each three-dimensional geometric grid, and performing assembly and serialization export on each corrected three-dimensional geometric grid. According to the invention, end-to-end automatic construction from the script text to the interactive virtual-real fusion scene is realized.

Systems and methods are described for generating, using the first trained machine learning model and based on text input, a single-layer image comprising a plurality of objects; generating a plurality of masks associated with the plurality of objects; determining a plurality of attributes associated with the plurality of objects; generating, using a second trained machine learning model, a plurality of textual descriptions respectively corresponding to the plurality of objects; inputting the plurality of textual descriptions, and the plurality of attributes, to the first trained machine learning model; generating, using the first trained machine learning model, a plurality of images respectively corresponding to the plurality of textual descriptions; and generating the multi-layer image by combining the plurality of images and by using the plurality of masks, wherein the plurality of images respectively correspond to a plurality of layers of the multi-layer image.

The invention discloses a table extraction method and device, equipment, a medium and a product. The method comprises the steps of obtaining an image corresponding to a table in a to-be-processed document, a first prompt text and a second prompt text; inputting an image corresponding to a table in the to-be-processed document and the first prompt text into a first multi-modal large model to obtain a comma separation value csv text of the table in the to-be-processed document; and inputting the csv text of the table in the to-be-processed document, the image corresponding to the table in the to-be-processed document and the second prompt text into a second multi-modal large model to obtain a hypertext markup language (HTML) text of the table in the to-be-processed document, through the technical scheme of the invention, the efficiency and accuracy of table extraction can be improved.

The invention discloses a semantic constraint enhanced image generation method based on a pre-training diffusion model, and the method comprises the steps: inputting a text Text into a CLIP text encoder of the pre-training diffusion model to obtain text semantic embedding, inputting the text semantic embedding and a noise signal into the pre-training diffusion model to obtain a latent space feature representation, and generating a latent space feature representation; according to the latent space feature representation, the attention weight matrix and the position set of the attribute-object phrases, obtaining a salient region map corresponding to each attribute-object phrase in the attribute-object phrase set; according to the salient region map corresponding to each attribute-object phrase, obtaining a regionalized image corresponding to each attribute-object phrase; updating the noise signal according to the gradient of the noise signal; and gradually denoising the updated noise signal to obtain a final noise signal, and inputting the final noise signal into the pre-training diffusion model to generate a target image. According to the invention, the accuracy and controllability of multi-semantic target image generation are improved.

Systems, methods and non-transitory computer readable media for hyper personalization of user interfaces are provided. In some examples, a textual input associated with a desire of a particular individual to affect a design of a user interface is received. The textual input and digital data associated with a first individual are used to generate a first version of the design. First digital signals are transmitted to a first computing device associated with the first individual to cause the first computing device to present the user interface based on the first version of the design. The textual input and digital data associated with a second individual are used to generate a second version of the design. Second digital signals are transmitted to a second computing device associated with the second individual to cause the second computing device to present the user interface based on the second version of the design.

The invention relates to the technical field of computer vision, and discloses a visual language processing method and device based on multi-encoder fusion, and the method comprises the steps: obtaining a to-be-processed visual input image, and determining a corresponding visual coding strategy; based on the visual coding strategy, distributing the visual input image to at least two visual encoders, and generating initial visual features corresponding to the visual encoders; performing consistency processing and fusion processing on all the initial visual features in sequence to generate fusion features; performing feature conversion on the fusion feature to obtain a target visual feature matched with the visual language model input layer; and inputting the target visual features and the text input data into a visual language model for visual text association processing, and generating a visual language task result. According to the method, the multi-source visual encoder is fused, the visual features of different granularities in the image are comprehensively extracted and fused, and the comprehensive perception and understanding ability of the visual language model to the image content can be improved.

The invention provides a text-based monitoring information identification method and device and electronic equipment, and relates to the technical field of intelligent security, and the method comprises the steps: obtaining an input text input by a user; the semantics of the input text is related to the monitoring information; according to the input text, a plurality of text corpora in the text corpus and a plurality of image corpora in one-to-one correspondence with the plurality of text corpora, obtaining the similarity between the input text and each text corpora; when at least one similarity smaller than a set threshold exists, obtaining a negative sample of the monitoring information based on the first text corpus corresponding to the similarity; calculating the similarity between the monitoring image of the current frame in the monitoring video and the input text, and the similarity between the monitoring image and each obtained negative sample; and under the condition that the similarity between the monitoring image and each negative sample is smaller than the similarity between the monitoring image and the input text, determining that the monitoring image is an image with monitoring information. According to the invention, the accuracy of the alarm classification identification result of the open alarm classification function can be improved.

The invention discloses an intelligent management method and system for an electric power operation ticket based on deep learning, and the method comprises the following steps: collecting a historical compliance operation ticket text and an electric power industry specification text, and carrying out the text preprocessing operation; based on the preprocessed text, respectively constructing training samples for operation step prediction and operation sequence prediction; inputting the training sample into the constructed operation order prediction model, calculating step prediction loss, prediction sequence loss, sequence consistency score and prediction violation score, weighting the calculation step prediction loss, the prediction sequence loss, the sequence consistency score and the prediction violation score to serve as a total loss function, and training the model; inputting an operation task and an operation ticket text into the trained model, and calculating a semantic distance; and if not, obtaining the prediction probability and the sequence prediction probability of the operation task and the operation ticket text at the lexical element position, and positioning a wrong lexical element. According to the method, through joint operation step prediction, operation sequence prediction, sequence consistency evaluation and violation detection, high-precision prediction and error lexical element positioning of the electric power operation order are realized.