53 results about "Language representation" patented technology

A system and method is described for internationalization of web pages by extracting translatable content from extensible mark-up language (XML), or similar data-centric meta-language representations of web pages or data used to build web pages. The extracted translatable content is stored in a translation task repository (TTR) accessible by the web developer and the translator. The XML representation is then modified to include selection control logic to select the appropriate translations for insertion into the final web page. The translator accesses the TTR to translate the appropriate content and saves the translations back to the TTR associated with the original translatable data. The translations are obtained from the TTR as selection cases for the selection control logic of the XML representation. As the XML is converted into the web source code, the selection logic and translations are embedded therein facilitating building the web site in multiple different languages.

A system for extracting a number of data elements from one or more data sources. A text representing tables using a markdown language is extracted from spreadsheets and or other grid-based documents. The text is provided to a language model with a prompt. The prompt may be a chain-of-thoughts prompt. The prompt includes several requests and / or steps that cause the language model to extract one or more tables from the spreadsheet and output the tables using the markdown language or a different markdown language. The tables extracted from the spreadsheet are converted into table chunks and indexed for retrieval by a retrieval augmented architecture. When a prompt to extract particular information from the spreadsheet is provided, one or more relevant table chunks are identified and provided to a language model for extraction. Using the language model to separate tables improves information extraction accuracy while maintaining downstream instructions.

Techniques for real-time accent conversion are described herein. An example computing device receives an indication of a first accent and a second accent. The computing device further receives, via at least one microphone, speech content having the first accent. The computing device is configured to derive, using a first machine-learning algorithm trained with audio data including the first accent, a linguistic representation of the received speech content having the first accent. The computing device is configured to, based on the derived linguistic representation of the received speech content having the first accent, synthesize, using a second machine learning-algorithm trained with (i) audio data comprising the first accent and (ii) audio data including the second accent, audio data representative of the received speech content having the second accent. The computing device is configured to convert the synthesized audio data into a synthesized version of the received speech content having the second accent.

According to the time series prediction method based on the multi-modal contrast learning technology, original multivariable time series data are converted into structured visual representation and language representation, multi-modal representation with consistent inner performance can be constructed without depending on external natural language or real image data, and the time series prediction method is high in practicability. The deep semantic understanding capability of the model on the complex operation state of the rail transit is effectively enhanced; a multi-modal contrast learning mechanism is introduced, visual and text modal representation is aligned in a shared embedding space, positive sample consistency is maximized through InfoNCE loss, negative sample interference is suppressed, and the robustness and generalization ability of time sequence features are remarkably improved; and the importance of each variable on a prediction task is dynamically evaluated by using the aligned multi-modal representation, and key variables are automatically screened, so that the redundant information interference is reduced, and the prediction precision and the calculation efficiency of the model in a high-dimensional multi-variable scene are also improved.

The invention discloses an urban park perception value evaluation method based on multi-modal data fusion, and the method comprises the steps: extracting a multi-dimensional perception tag and an emotion intensity score from park social media text comment data through a pre-trained language model, and carrying out the aggregation of the multi-dimensional perception tag and the emotion intensity score, and generating a subjective perception result; performing element identification on the image comment data of the park social media by using the three models to obtain different scene element coverage rates, character / facility counts and language representations of visual elements; constructing a plurality of objective environment indexes to describe different environment states, extracting variable information from three model output results, POI data, remote sensing land class proportion and road network information, and constructing mapping with the corresponding objective environment indexes to quantify each objective environment index; and aligning and outputting the subjective perception result and the objective environment index in the park and period dimensions. According to the method, automatic quantitative evaluation of subjective and objective characteristics of urban parks can be realized on the national scale, and standardized technical support is provided for urban renewal and ecological management.

A Cybersecurity Event Handling Processor (CEHP) and method for processing security alerts includes: a File System containing a Universal Target Schema (UTS) of target language representations (UTS JSONs); a Normalizer running Feature Extraction and Word Embeddings algorithms; a Tree Converter; and a Transformer running linguistic and structural matching algorithms. The CEHP: (a) captures threat events in one or more native formats generated by cybersecurity tools; (b) runs Feature Extraction and Word Embeddings algorithms for tokenization and categorization of the captured events to create normalized events; (c) converts the normalized events into trees and then translates the trees into event representations in JSON (or XML) format (Event JSONs); and (d) runs nearest neighbor and / or linguistic and structural matching algorithms to compare the Event JSONs to the UTS JSONs to generate output JSONs (Translation JSONs) from the UTS corresponding to the captured events.

A small sample based general image counting method, comprising: performing feature extraction on a first image to obtain first features; performing attention calculation based on the first features, a general language representation and a general visual representation to obtain first example features of a first example target, wherein the general language representation is used to describe categories of different objects, the general visual representation is used to describe visual information of different objects, the first example target is an object related to an example frame in the first image, and the first example features include a first language representation and a first visual representation, the first language representation is used to describe a category of the first example target, and the first visual representation is used to describe visual information of the first example target; performing matching on the first example features and the first features to obtain a correlation feature map; and obtaining a first counting result related to the first example target based on the correlation feature map. The method can greatly improve the generalization of the counting algorithm.

A program translation system may generate a first program source internal representation expressed using a source language representation of a source programming language and a first program target internal representation that is expressed using a target language representation of a target programming language. The system may interpret the first program source and target internal representations with first program data to generate a first program source and target results. The system may, when the source and target results are inconsistent, modify the target language representation. The system may, when the source and target results are consistent, generate a second program source internal representation of a second program. The system may use the second program source internal representation to generate a second program target internal representation of the second program. The system may use the second program target internal representation to generate a second program in the target programming language.

The application discloses a kind of urban park perception value evaluation methods based on multi-modal data fusion, utilize pre-trained language model to park social media text comment data extraction multidimensional perception label and sentiment intensity score, and subjective perception result is aggregated and generated;Park social media image comment data is identified using three kinds of models, obtain different scene element coverage, figure / facility count, and language representation of visual elements;A plurality of objective environment indexes are constructed to describe different environmental conditions, variable information is extracted from the output results of the three models, POI data, remote sensing land class proportion and road network information, and a corresponding objective environment index is constructed to map to quantify each objective environment index;Subjective perception result and objective environment index are aligned in park and period dimension and output.The present application can realize the automatic quantification evaluation of subjective and objective characteristics of urban park at national scale, and provide standardized technical support for urban renewal and ecological management.

The application discloses a commodity semantic representation vector generation method and device, and relates to the technical field of electronic commerce. A specific embodiment of the method comprises the following steps: performing word piece covering processing on each commodity title in a commodity title set respectively, obtaining a word piece index vector, a sample length vector and a text segment vector corresponding to each commodity title according to the commodity title after the word piece covering processing, and inputting the vectors into a pre-training model to obtain a first vector representation corresponding to each commodity title and a covering prediction result; then, adjusting parameters of the pre-training model to generate a semantic representation vector extraction model; and extracting vectors from the commodity title set by using the semantic representation vector extraction model to obtain a commodity semantic representation vector. The embodiment is more suitable for commodity language representation vector extraction, and can obtain a more accurate semantic representation vector, improves the semantic representation effect, and improves the use effect of the commodity semantic representation vector.

The present disclosure relates to multilingual speech recognition models for speech processing systems and applications. In various examples, described herein are multilingual speech processing models for speech processing systems and applications. The systems and methods described herein can use an end-to-end model that is capable of performing both ASR processing and translation processing to generate text presented in various languages. For instance, a user can provide at least audio data representing speech and an indication of a target language for translating the speech. The model can then generate one or more audio representations associated with the speech and one or more language representations associated with the target language. Additionally, the model can combine the audio representations with the language representations (such as by performing stitching, masking, fusing, adding, etc.) to generate one or more combined representations. The model can then process the combined representations to generate text that corresponds to the speech and is presented in the target language.

The application discloses a learner multi-modal sentiment analysis method and device based on text main driving, extracts multi-modal data with student related sentiment information embedded in an online classroom, uses a language representation pre-training model BERT to perform feature extraction on text modal data, uses an LSTM to perform feature extraction on audio and visual modal data, uses a cross-modal attention mechanism to fuse multi-modal information, and outputs a final result of multi-modal feature fusion, and performs sentiment analysis according to the final fusion result. The contrast learning technology is used to promote single-modal feature coding quality, maintain task related modal data uniqueness, ensure that multi-modal fusion results sufficiently learn unique sentiment information of various modal data generated in a classroom, improve learner participation in an online learning environment, and thus promote teaching quality.

The invention discloses a large model reasoning enhancement method combining knowledge graph embedding and gating residual connection, and relates to the technical field of artificial intelligence and natural language process.The method comprises the steps that initial knowledge embedding is obtained through knowledge graph embedding learning, and a knowledge unit representation library is constructed through semantic space mapping processing; identifying a knowledge demand based on the text processing training data to obtain a query triple, retrieving a candidate knowledge set in the knowledge graph and constructing an enhanced training sample; inputting the enhanced sample into a large language model of an integrated knowledge gating residual connection module, and performing dynamic gating fusion processing according to the current layer hidden state and the knowledge unit representation to obtain an enhanced hidden state after knowledge injection; and generating a prediction result based on the enhanced hidden state, determining a joint loss function, and optimizing model parameters to obtain a reasoning enhanced large language model. According to the method, deep alignment and dynamic regulation and control of knowledge semantics and language representation are realized, and model reasoning accuracy and fact consistency are improved.

The application discloses a novel word-level contrastive learning framework for sign language translation and a sign language translation system, relates to computer vision and sign linguistics, and provides a novel word-level contrastive learning framework ConSLT, which comprises a video input module, a visual extraction module, a sign language coding module, a sentence embedding module, a sign language decoding module, a contrastive learning module, a loss calculation module and an output module. The method comprises the following steps: 1) selecting a sign language corpus for modeling; 2) extracting sign language visual features; 3) performing end-to-end sign language video conversion; 4) calculating sentence embedding in a training stage; 5) constructing positive example pairs and negative example pairs; 6) calculating a sign language translation model loss; and 7) outputting a sign language translation result. From the perspective of natural language processing, the application explores contrastive learning of sign language translation, directly utilizes data itself as supervision information, and can learn good sign language representation under a low-resource condition, so that the sign language translation system is more accurate and fluent. The ConSLT framework is not limited by a model and is suitable for different models.

The method and system for migrating web application firewall (WAF) rules across different WAF providers is presented. The method includes parsing a plurality of WAF rules from a plurality of WAF providers, wherein the plurality of WAF rules is expressed in varying provider-specific formats; enriching a source WAF rule of a source WAF provider with organizational context; generating, using a trained cross-provider semantic similarity model, a provider-agnostic language representation of the source WAF rule based on the provider-specific format of the source WAF rule and the organizational context; constructing a capability model of a target WAF provider; generating, using the trained cross-provider semantic similarity model, a target WAF rule for deployment in the target WAF provider based on the provider-agnostic language representation of the source WAF rule and compatible with the capability model; and coordinating a staged deployment of the target WAF rule in the target WAF provider.

A method is provided for transforming a high-level language representation of a tensor computation graph into a low level language. The method includes assigning a tensor shape and a loop primitive. The method also includes generating, from the tensor computation graph and the assigned loop primitives, an initial loop structure. The method further includes positioning the layers of the tensor computation graph within a nested loop structure to provide a final loop structure, collapsing loops in the final loop structure, and mapping the collapsed loops to hardware components configured to execute the collapsed loops. The method can be applied to artificial intelligence (AI) and machine learning (ML) use cases for improved optimization of neural networks including compilation optimization for improving performance of simulations such as medical simulations, healthcare simulations, weather simulations, and / or simulations related to other complex systems, which can also support decision making.

The embodiment of the invention provides an entity question and answer generation method and device, equipment and a storage medium. The method comprises the following steps: analyzing a to-be-processed knowledge file to obtain a processable text; according to the processable text, performing semantic segmentation by adopting a bidirectional language representation model to obtain one or more to-be-extracted text blocks; adopting a trained entity question and answer generation model to obtain a named entity corresponding to each to-be-extracted text block, and obtaining an entity question and answer pair corresponding to each named entity; the method comprises the following steps: acquiring a training named entity and a training question-answer pair for training an entity question-answer generation model through a cue word project and a language large model; and according to the question statement, screening out an entity question and answer pair matched with the question statement, and outputting an answer statement in the entity question and answer pair. The training named entities and the training question and answer pairs are generated through the cue word engineering and the language large model and are used for training the entity question and answer generation model, and the richness and the retrieval recall rate of a knowledge base are improved.

This invention belongs to the field of computer vision and pattern recognition technology, and discloses a time retrieval method and related apparatus for learning multi-view auxiliary representations, aiming to solve the technical problem of unreliable time retrieval results in existing multimodal retrieval methods. The technical solution of this invention includes: constructing a multi-view auxiliary representation set based on text features; injecting target-specific visual attributes from the video into the auxiliary representation while maintaining semantic consistency to obtain an auxiliary representation that fuses visual attributes; achieving cross-modal association between the auxiliary representation and video features through multi-view semantic alignment to obtain enhanced visual features; and completing target time location through a detection head optimized by a loss function. The technical solution disclosed in this invention, through a multi-view auxiliary representation framework, actively injects visual evidence into the language representation, reduces the risk of overfitting sparse text cues, establishes semantic alignment of visual perception, and significantly improves the robustness and localization accuracy of time retrieval.

The invention discloses a data analysis method and system based on artificial intelligence, and relates to the technical field of artificial intelligence data process.The method comprises the steps that an input text is split, and semantic representation is output through a multi-scale pyramid; generating semantic anchor points through a dynamic time warping mechanism, aligning input semantic representations, constructing a domain knowledge structure, and converting the domain knowledge structure into a semantic prior tensor; inputting the semantic priori tensor into a priori injection channel of a text encoder, and encoding the semantic representation to form an encoded representation; constructing a semantic graph according to the coded representation and the semantic representation, inputting the semantic graph into a graph neural network, and outputting a graph enhanced representation; according to the coding representation and the graph enhancement representation, constructing a joint representation, training an analysis model, and outputting a cross-language representation; clustering the cross-language representation to generate an adversarial sample, and constructing a semantic causal path diagram with an original sample; the problems of semantic segmentation, alignment deviation and weak cross-language migration are solved.

A universal target tracking enhancement method based on implicit language guidance comprises the following steps: step 1) obtaining template data and a search image, and converting the template data and the search image into a processable visual embedding vector through a patch embedding layer; 2) generating implicit language representation; 3) inputting a frozen CLIP text encoder to extract high-level semantic language features; 4) performing multi-scale extraction on the visual features after patch embedding to obtain hierarchical visual features from a shallow layer to a deep layer; 5) performing cross-modal fusion on the semantic language features of each level and the visual features of the corresponding level to generate enhanced visual features; according to the method, on the premise that additional explicit text input is not needed, the robustness and accuracy of target tracking in a complex scene are remarkably improved, and meanwhile low calculation overhead and good universality are guaranteed.

A system for extracting a number of data elements from one or more data sources. A text representing tables using a markdown language is extracted from spreadsheets and or other grid-based documents. The text is provided to a language model with a prompt. The prompt may be a chain-of-thoughts prompt. The prompt includes several requests and / or steps that cause the language model to extract one or more tables from the spreadsheet and output the tables using the markdown language or a different markdown language. The tables extracted from the spreadsheet are converted into table chunks and indexed for retrieval by a retrieval augmented architecture. When a prompt to extract particular information from the spreadsheet is provided, one or more relevant table chunks are identified and provided to a language model for extraction. Using the language model to separate tables improves information extraction accuracy while maintaining downstream instructions.

The invention provides a personalized recommendation method and system based on a language modeling and elastic reasoning collaborative architecture, and belongs to the technical field of recommendation. The system comprises a data module, a language representation construction module, an elastic sub-model generation module, a dynamic routing training module and a self-adaptive recommendation module. In the implementation process, the language representation construction module formats user behaviors and article information into a natural language sequence through a predefined personalized prompt template, unified coding is conducted through a large language model, and user and article language representations with rich semantics are generated. The elastic sub-model generation module carries out multi-scale structure segmentation on a feed-forward layer and multi-head attention in a unified Transform architecture, and a series of nested sub-models with gradually increased calculation complexity and consistent semantics are constructed. In the model training process, the dynamic routing training module randomly activates sub-models of different scales for forward and reverse propagation in each training step, introduces a task-aware routing mechanism, and dynamically selects the optimal sub-model according to the characteristics of the current recommendation task in the reasoning stage. And the self-adaptive recommendation module loads the trained elastic language recommendation model, and calls the sub-model selected by the dynamic route to generate a recommendation result according to the context information and the task type of the target user. Through wide experimental verification, the prediction performance superior to that of a full-size model is realized on various recommendation tasks, and the problem of performance fluctuation of a large language model in a recommendation scene due to mismatching of a model scale and task requirements is effectively solved.

This disclosure relates to a sentiment classification method and apparatus, belonging to the field of natural language processing technology. The method includes: extracting comment text data containing character text and emoticon text; inputting the comment text data into a preset language model to obtain a first feature vector of the character text and a second feature vector of the emoticon text; the preset language model being a language representation model pre-trained using unlabeled first sample comment text data and fine-tuned using labeled second sample comment text data; performing sentiment enhancement processing on the first feature vector according to a preset sentiment dictionary to obtain a third feature vector of the character text; and inputting a concatenated feature vector of the third and second feature vectors into a fully connected layer in the preset language model to obtain the sentiment classification result corresponding to the comment text data. This method reduces reliance on labeled data during sentiment classification and performs sentiment prediction from multiple feature dimensions, improving prediction accuracy.

The invention provides a sequence recommendation method and system based on double-view collaborative fusion, and belongs to the technical field of recommendation. The system comprises a data module, a language representation construction module, a hierarchical semantic discretization module, a representation enhanced recommendation modeling module and a Top-K recommendation module. In implementation, the data module collects behavior information of interaction between a user and an article in the intelligent terminal and metadata of the article through a network protocol and stores the behavior information and the metadata in the database. And the language representation construction module formats the user behavior information and the article information retrieved from the database into text prompts by using the user prompt template and the article prompt template, and encodes the text prompts by using a large language model to obtain user language representation and article language representation with rich semantics. The hierarchical semantic discretization module utilizes language characterization to train a user semantic ID generation model and an article semantic ID generation model to generate a user semantic ID and an article semantic ID that can capture potential classification information. A representation enhanced recommendation modeling module utilizes data retrieved from a database to train a sequence recommendation model based on dual-view collaborative fusion. And the Top-K recommendation module recommends interested articles to the target user by loading the trained recommendation model. Wide experiments prove that the method can fully capture the potential mode of user interest evolving along with time, so that articles which are possibly interacted with the user in the future can be better recommended to the user.

A computer system identifies that a first markup language portion extracted from a markup language document of a website corresponds to a first actionable element, where the first markup language portion is a variable length representation. In response to the identification that the first markup language portion corresponds to the first actionable element, the computer system creates a first code representation corresponding to the first markup language portion using a recurrent neural network (RNN) encoder. The computer system identifies first additional information corresponding to one or more predefined targets. The computer system creates a final fixed length markup language representation that includes the first code representation and the first additional information. The computer system inputs the final fixed length markup language representation into a model.

The application belongs to the technical field of sign language recognition, and discloses a sign language recognition method and system based on LSTM and NLP, which contains two stages of model training and sign language recognition. In the model training stage, a timing key point sequence sample is obtained by using a sign language word unit data set and a gesture detection model, an LSTM-SL model is trained, a sign language representation four tuple is obtained by using sign language common corpus word segmentation, and a W2V-SL model is trained. The sign language recognition stage includes converting a sign language video to be recognized or a real-time sign language picture into a sign language word unit prediction sequence by using the LSTM-SL, outputting a sign language sentence conforming to a language specification by using a sliding window mechanism and the W2V-SL model, and proposing a sign language recognition system based on the method. The application simplifies the collection difficulty of timing characteristics of sign language recognition, effectively enhances the output effect of natural sentences, and improves the robustness and accuracy of the sign language recognition system.

The invention relates to a Chinese text error correction method based on model fusion and scene self-adaption, belongs to the field of electric digital data processing, and solves the problems that a single model is limited in language representation capability and insufficient in generalization capability, and a traditional model cascading mode has information redundancy and static fusion, so that the error correction efficiency is low. Meanwhile, the problem that the error correction performance is limited due to the fact that a general error correction model is not optimized for a specific scene is solved. Comprising the following steps: inputting a to-be-corrected text into a trained text scene recognition model to obtain a corresponding scene category; based on the scene category, loading trained three-channel mutual error correction model parameters corresponding to the scene; inputting a to-be-corrected text into the three-channel mutual error correction model loaded with the parameters to generate a first error correction result, a second error correction result and a third error correction result; wherein the three-path mutual error correction model comprises a first mutual error correction path, a second mutual error correction path and a third mutual error correction path which are parallel; and voting word by word based on the first, second and third error correction results to obtain an error-corrected text corresponding to the to-be-corrected text. And scene-based self-adaptive efficient error correction is realized.

The invention provides an intelligent question answering system and a question answering method thereof, and relates to the technical field of artificial intelligence, the method comprises the following steps: receiving a question statement input by a user, and converting the question statement into a question word vector; inputting the question word vectors into the trained language representation model to obtain a plurality of answers; wherein the language representation model is obtained by training question and answer data after entities are defined in the knowledge graph; screening out qualified answers from the answers based on the knowledge graph; wherein the knowledge graph comprises question and answer data after the entity is defined; and converting the qualified answers into answer word vectors, calculating the similarity between each answer word vector and the question word vector, and outputting the answer corresponding to the answer word vector when the similarity is the highest as the optimal answer. Based on the knowledge graph retrieval information and the answering efficiency of the question and answer system, the answers are verified through the knowledge graph, the optimal answer is further screened out, and the answering accuracy is improved.

The present application relates to a kind of Chinese text error correction method based on model fusion and scene self-adaption, belong to electric digital data processing field, solve the limited language representation ability of single model and the insufficient generalization ability, traditional model cascade mode exists information redundancy and static fusion, while general error correction model is not optimized for specific scene, leading to error correction performance is limited.The problem of including the input of text to be corrected text training good text scene identification model, obtains corresponding scene category;Based on scene category, load the corresponding training good three-path mutual error correction model parameters of this scene;The text to be corrected text is input into the three-path mutual error correction model after loading parameters, to generate first, second and third error correction results;Wherein, three-path mutual error correction model includes parallel first, second and third mutual error correction path;Based on first, second and third error correction results, word-by-word voting is carried out, and the text after error correction corresponding to the text to be corrected text is obtained.The realization of high-efficiency error correction of scene self-adaption.

The invention belongs to a document generation method, and provides a document analysis and enhancement generation method and a related device aiming at the technical problem that factual errors or information missing easily occur in generated contents when table contents are designed by an existing document retrieval enhancement generation technology. The method comprises the following steps: analyzing a document image through a layout detection model, identifying to obtain different functional regions of the document image, then constructing a document structure tree for representing a spatial and hierarchical relationship between the functional regions, respectively performing fine-grained semantic analysis on each functional region by using a visual language model, inputting semantic analysis into a large language model, and constructing a document structure tree for representing the spatial and hierarchical relationship between the functional regions; the corresponding natural language description is obtained, then the natural language description is coded into dense vectors, and all the dense vectors are constructed into a vector index database; according to the method, fine-grained layout detection, multi-modal structure understanding and structured language representation are introduced into a retrieval enhancement generation process, and an end-to-end analytic framework is formed.