170 results about "Text corpus" patented technology

A text sequence recommendation method and system based on a large language model is disclosed, belonging to the technical field of recommendation algorithms. The method includes: a data preprocessing stage, a large language model pre-training stage, a sequence model fine-tuning stage and a matching stage. According to this disclosure, a large language model is introduced into a text sequence recommendation task, so that text can be better modeled by utilizing rich pre-training corpus of the large language model; meanwhile, sequence modeling is performed on the text, the capability of sequence recommendations modeling in a large model is activated, an ID-based recommendation paradigm in a traditional recommendation algorithm is eliminated, and recommendation task learning processing is better performed in a cold start scenario and a knowledge transfer scenario; and finally, a recommendation result is finally optimized by a sequence model.

Disclosed in some examples are methods, systems, devices, and machine-readable mediums which create various data visualizations from a corpus of raw data collected from one or more sources. For example, natural language text documents of a corpus may be classified based upon the topic of the data. The documents may be labeled with the phase during which the documents were collected or observed. A visualization may then be generated which shows a correlation between the phase and the topics observed. For example, a number of times a particular topic appeared in a particular phase. The visualization may be two-dimensional, three-dimensional, or the like.

The present disclosure relates to the technical field of text processing, and provides a text data enhancement method and device. The method comprises: obtaining a text corpus set, wherein the text corpus set comprises a plurality of text corpora; converting each text corpus into a text vector by using a diffusion process of a text diffusion model, and obtaining a first target noise vector corresponding to each text corpus by adding noise to each text vector for a plurality of times in succession; predicting a plurality of noises added in the diffusion process by using an inverse diffusion process of the text diffusion model, and removing the plurality of predicted noises in turn by using the first target noise vector corresponding to each text corpus, to obtain a restored text vector corresponding to each text corpus; and converting the restored text vector corresponding to each text corpus into a text, to obtain a first data enhancement text corresponding to each text corpus. By using the above technical means, the problem that the text obtained by a traditional data enhancement method deviates from the original text is solved.

The present disclosure provides a domain dictionary constructing method and apparatus, the method including: segmenting a domain corpus sample to obtain a first character segment set, where the first character segment set includes at least one first character segment; calculating an inter-character correlation index of the at least one first character segment; determining a second character segment set according to the correlation index, where the second character segment set includes a first character segment of which the correlation index is greater than or equal to a preset threshold; determining a third character segment set according to the second character segment set, where the third character segment set includes the second character segment set and second character segments of the domain corpus; constructing a domain dictionary according to the third character segment set.

This application provides a company name matching method, apparatus, device, medium, and program product. The method includes: constructing a structured corpus covering all components of a company name and automatically generating diverse training samples, enabling a company name matching model to generalize and learn the deep semantic combination rules of company names. When obtaining a first company name and a second company name to be matched, the first company name and the second company name are respectively input into the trained company name matching model to obtain corresponding first feature vectors and second feature vectors. Then, the distance between the first feature vector and the second feature vector is calculated, and the distance is compared with a preset threshold to determine whether the first company name and the second company name refer to the same entity. This method achieves the effect of improving the accuracy and generalization ability of company name matching.

Techniques are provided for a large language models (LLM) to reduce risks associated with potential introduction of factual inaccuracies and / or hallucinations in responses to user queries to the LLM, for example when responding to user queries pertaining to advanced scientific fields. One or more documents from a document corpus are identified to generate a response to a user query, and relevant information points for (e.g., citations to) the one or more documents are associated therewith to demonstrate accuracy of the response. Validation techniques may be provided to further ensure accuracy of the response to the user query.

The application discloses a financial risk control report generation method, device, equipment, storage medium and product, relates to the technical field of artificial intelligence, and the method comprises the following steps: generating a preference data set through a supervised fine-tuning technology, a preset risk control report corpus and a preset large language model; performing iterative preference training on the preset large language model according to the preference data set; and generating a financial risk control report through a large language model after training and an intelligent agent-based retrieval enhancement generation technology. The application uses the risk control report corpus of a user to generate a preference data set through supervised fine-tuning, a corpus and a large language model, thereby reducing the dependence on a large amount of labeled data; the model is trained through the preference data set, a database query mode is provided to improve the report generation efficiency of the model, thereby solving the problem that a large number of high-quality samples are required for training in the prior art, and only a small number of samples and iteration times are required to align the output of the large language model with the expected behavior of the user.

The application discloses a kind of achievement transformation report intelligent generation method and system, comprising: constructing achievement transformation corpus;Based on corpus, use five-element prompt word framework to extract entity and relationship, check, implicit completion and weight update are carried out in combination with achievement transformation value evaluation model, and optimized knowledge graph is constructed;Knowledge cluster is divided to graph, and dual-mode index network of paragraph vector index and structured knowledge index is constructed;After receiving user input, core entity set is obtained by semantic coding and entity recognition;With the set as center, structured knowledge enhancement retrieval is executed, and evidence fragment is screened by fusing semantic similarity and knowledge correlation degree score;Finally, sort constructs input context, calls natural language generation model to generate structured report, simultaneously extracts reasoning path and is associated with evidence source and outputs.The application solves the problems of poor precision in manual writing, poor adaptation of general model and insufficient semantic mining, and improves the scientificity, pertinence and traceability of the report.

The application provides a kind of industrial vertical field corpus construction data cleaning method, belong to industrial big data technical field, the application is through collecting multi-modal industrial data and recording original time stamp, unified time reference is generated using dynamic time warping algorithm for time alignment, cross-modal fusion feature vector is generated by multi-modal attention mechanism learning feature interaction weight, key data section is identified based on event-driven trigger rule and is associated as event data package, quality evaluation and abnormal calibration are carried out through step response consistency detection, multi-modal knowledge graph is constructed and cross-modal alignment model based on dual optimization constraint satisfaction framework is used for feature completion, finally, data quality is ensured through adaptive verification process, solve the technical problems that multi-modal industrial data is difficult to ensure alignment accuracy and original information integrity in time dimension alignment and semantic level fusion process.

The application provides a large model-based backdoor attack method, system, storage medium and equipment. A tokenizer and a known corpus of an attacked model are obtained, a set of candidate words is integrated by selecting adverbs in the known corpus; each adverb in the set of candidate words is converted into a word unit sequence composed of several basic word units; a tail word unit and a preset specific word unit are combined into a trigger combination, and the frequency of the trigger combination in the original training set is counted; when the frequency of the trigger combination is less than a first threshold value and the number of word units ending with the tail word unit is not less than a second threshold value, the corresponding adverb is added to a trigger substructure set; a data training set is constructed, the data training set includes a dirty training set, the input of the sample in the dirty training set contains the trigger combination, and the output of the sample in the dirty training set is an attack result; and the attacked model is trained by using the data training set. The problem of insufficient concealment of the trigger used for attack in the large model is solved.

The application discloses a log analysis method, device and equipment and a readable storage medium. The method comprises the following steps: collecting logs; extracting feature words from each log to obtain feature words in each log; corresponding feature words in each log form a feature word set of each log; according to the feature word set of each log and a seed set containing multiple seeds, determining a corresponding seed of each log, and classifying each log into a class where the corresponding seed of each log is located. The technical scheme disclosed by the application extracts feature words from logs to extract key information from the logs, classifies the collected logs according to the feature word set of each log, obtains key information and important information such as the category of the logs from the collected logs, so that the operation and maintenance personnel can query and use the logs faster and better, and the result obtained by analyzing the logs can also be used as a corpus for machine learning related to automatic operation and maintenance.

Interactive layer neural networks for search, retrieval, and ranking. A language system includes a controller. The controller can be configured to receive a query and a document, tokenize the query into a sequence of query tokens and tokenize the document into a sequence of document tokens, generate a token pair matrix for each query and document token, retrieve, for each entry in the token pair matrix, a pre-computed similarity score produced by a neural conditional translation probability network, wherein the neural network has been trained in a ranking task using a corpus of paired queries and corresponding relevant documents, aggregate, via a sum product of each similarity score for a respective query, a ranking score for each document relative to each query; and output the document and the associated ranking score for the document.

The application discloses a power grid real-time scheduling decision method and device based on large model cooperative reinforcement learning, and relates to the field of power grid scheduling. The method comprises the following steps: obtaining input text information, analyzing the input text information based on a preset large language model to generate a real-time semantic feature vector; the large language model is fine-tuned based on a preset corpus; obtaining a power grid real-time state vector, fusing the power grid real-time state vector and the real-time semantic feature vector to obtain a fused state vector; inputting the fused state vector into a preset reinforcement learning intelligent agent to obtain scheduling action information; converting the scheduling action information and the power grid real-time state vector into a to-be-processed sequence based on a preset text conversion mapping table, and inputting the to-be-processed sequence into the large language model to generate a scheduling action explanation text in a natural language form.

A system for influencing outputs of a large language model includes at least one memory, at least one processor, a perplexity optimizer, and a corpus handler. The processor executes instructions stored in the memory to operate the optimizer and handler. The perplexity optimizer generates multiple candidate supporting texts based on a target concept. It computes a perplexity metric for each candidate within a context derived from a local corpus, using token likelihoods from a reference language model, and selects a supporting text based on the metric. The corpus handler identifies online editable corpora based on their likelihood of inclusion in language model training data. It then collects contextual information from these corpora to create the local corpus and inserts the selected supporting text into at least one of the identified online corpora.

The application discloses a kind of thinking chain enhancement method and device for RAG system, the method includes: after obtaining input question text set, relevant document set associated with question text set is obtained from corpus as enhancement information;Multiple candidate thinking chains and corresponding answer texts are generated in parallel;Each candidate thinking chain is automatically scored by multiple dimension evaluation algorithm, and comprehensive score is generated based on multidimensional score and is sorted;When determining that initial optimal thinking chain is less than preset threshold, the optimal thinking chain that does not reach threshold is progressively refined and optimized, and improved version thinking chain is obtained;Preference pair sample is constructed, and LoRA fine-tuning is carried out on model based on preference pair sample by preference optimization algorithm;The model after fine-tuning is loaded to execute input question text reasoning task.The scheme can enable model to directly generate high-quality thinking chain in subsequent reasoning, and can improve the accuracy of reasoning answer.

The application discloses a kind of multi-source knowledge fusion's administrative penalty intelligent question and answer model construction method and system, it is related to artificial intelligence and market supervision technical field.The present technique is dispersed in view of the existing technology administrative penalty knowledge, question and answer accuracy is poor, update lag and lack of explainability problem, first acquisition and pre-process multi-source market supervision data construction field corpus;Subsequently based on domain ontology guide and large language model extraction, construct and realize semantic uniformity and entity alignment administrative penalty knowledge graph;Further design by search channel and generation channel collaborative dual-channel intelligent question and answer architecture, based on knowledge graph generation traceable natural language answer;Finally, combined with including automatic incremental update, regulation time limit management and user feedback optimization knowledge evolution mechanism, realize the continuous improvement of knowledge.The application realizes the uniform modeling and intelligent question and answer of multi-source supervision knowledge, significantly improves the accuracy, explainability and real-time update capability of administrative penalty intelligent question and answer.

The application discloses a structured and unstructured fact knowledge fusion corpus construction method and device. Structured fact knowledge is selected from a pre-constructed structured knowledge base and is converted into a fact retrieval query. A plurality of unstructured fact knowledge text candidates are retrieved from a pre-constructed unstructured text base according to the fact retrieval query. Entities in the unstructured fact knowledge text candidates are obtained by using entity recognition and entity linking. The fact knowledge text candidates are matched with the structured fact knowledge in the knowledge base. The fact relevance of the fact knowledge text candidates is judged based on the matching result. Text candidates with strong relevance and the structured fact knowledge matched therewith are reserved as a piece of structured and unstructured fact knowledge matched corpus and are saved into a corpus. The above process is repeatedly performed, and finally a high-quality corpus containing a plurality of matched corpora is formed.

The present application belongs to the field of software vulnerability detection, and is a source code vulnerability detection method based on composite program representation. The method comprises the following steps: first, three kinds of source code intermediate representations are extracted from the source code, and corresponding paths are extracted; then, extra marks are removed from the extracted paths, and path sequences are constructed, all path sequences are spliced to construct a training corpus; then, a code embedding model is trained using a doc2vec model according to the constructed corpus; then, a vector representation form of the source code is obtained using the trained embedding model, and a training set, a test set and a validation set are divided; then, a metric learning model is improved by combining a regularization loss and a four-tuple loss function, the model is trained using the training set to obtain a vulnerability detection model; finally, whether the source code contains a vulnerability is detected using the trained vulnerability detection model. The present application solves the problem that the existing method is based on a single specific representation, ignores the complementary relationship between different representations, and is insufficient in representing the semantic and syntactic information of the code by using a composite source code representation method.

This invention relates to the field of machine learning technology, and discloses a method and system for training a large-scale medical language model based on knowledge graph enhancement. The method includes: acquiring a medical corpus and knowledge graph; generating entity sequences through fine-grained semantic parsing; extracting multi-hop subgraphs through multi-dimensional association path reasoning; dynamically injecting text into the graph based on dynamic weights to generate enhanced training samples; parsing and encoding into a standardized instruction set; and iteratively optimizing to obtain the target model. This invention improves the training efficiency and professional reasoning capabilities of a large-scale medical language model through adaptive knowledge fusion and standardized instruction encoding.

This invention provides a method, device, and storage medium for generating and utilizing a multimodal discourse tree (MMDT). [Solution] The method generates an Extended Discourse Tree (EDT) from a text corpus (e.g., from a Discourse Tree (DT) or Communication DT (CDT)), links data records (e.g., records containing numerical data) to the Extended Discourse Tree, and generates a Multimedia Discourse Tree (MMDT). The MMDT links any suitable text / records from heterogeneous sources. For example, entities identified from the basic discourse units of the EDT are matched with entities in the data records. The method also identifies causal links between EDTs and / or between data records, identifies rhetorical relationships for each entity match / causal link match, and merges the data records with the EDT to generate the MMDT.

This system provides an efficient management system for the RAG corpus, which is used to prompt generative AI models to produce improved responses. [Solution] The system and method include receiving a query from a user, determining a first text portion from a plurality of stored text portions that is semantically similar to the query, determining a first score associated with each of the first text portions, and generating a first prompt based on the first score, wherein the first prompt includes the query and the first text portions; sending the first prompt to a text generation model; receiving a response to the first prompt from the text generation model; presenting the response and the first text portions; receiving an evaluation of one of the presented first text portions from the user; and updating a first score associated with one of the first text portions based on the evaluation.

This invention discloses an intelligent customer service semantic analysis system for e-commerce, belonging to the field of e-commerce technology. It includes a text input module, a text preprocessing module, a multi-module collaborative semantic parsing module, an e-commerce scenario dynamic association module, a personalized reply generation module, a closed-loop self-learning optimization module, and a text output and log recording module. These modules achieve bidirectional interaction and collaborative work through a data bus, forming a closed-loop architecture. Through a semantic association fusion unit, it achieves collaborative calibration of entities, intents, logic, and emotions. The accuracy of complex sentence parsing is improved by 25%-30% compared to existing technologies, the entity extraction accuracy reaches over 98.5%, and the intent recognition F1 score reaches over 97.8%. Through dynamic dictionary updates and a closed-loop self-learning mechanism, it adapts to new product categories and new business terms in real time, reducing the entity false negative rate to below 1.2%, eliminating the need for frequent manual corpus maintenance. Through multi-source data collection and collaborative iteration, the system's response time to new scenarios is shortened from weekly to daily.

Systems and methods for providing user interfaces to converse with a corpus of electronic documents via a large language model are disclosed. Exemplary implementations may: present a user interface configured to obtain entry of user input from a user to select one or more documents to be provided as input to a large language model for an individual conversation; responsive to selection of the individual conversation, provide an individual query as a prompt to the large language model; obtain and present an individual reply from the large language model; determine an individual document from the one or more documents that is relevant to the individual reply; present the individual document in a particular portion of the user interface; and / or perform other steps.

The application provides a kind of enterprise carbon footprint intelligent accounting and tracing method based on large language model, it is related to carbon accounting and tracing technical field, including: based on carbon accounting field corpus parameter adjustment and instruction fine-tuning of general large language model obtain carbon emission large language model, based on carbon accounting rule construction carbon emission accounting knowledge graph;Determine the carbon emission accounting boundary of the enterprise to be accounted;The initial carbon emission data set of the enterprise to be accounted is standardized to obtain a standardized data set;Adjust the carbon emission knowledge graph to obtain the enterprise accounting knowledge graph of the enterprise to be accounted;According to the preset carbon emission accounting model and the standardized data set, determine the carbon emission accounting result, and based on the enterprise accounting knowledge graph and the carbon emission accounting result, determine the carbon footprint accounting result and the carbon emission tracing chain of the enterprise to be accounted.The intelligentization, automation and precision of enterprise carbon footprint accounting and tracing are realized.

This invention proposes a cybersecurity knowledge question-answering method based on a large language model, comprising the following steps: S1, collecting cybersecurity data to construct a cybersecurity knowledge corpus and preprocessing it to obtain a standardized cybersecurity knowledge corpus; S2, using a cybersecurity-specific Prompt template to fine-tune the pre-trained large language model to generate a cybersecurity large language model; S3, receiving user cybersecurity question-answering requests, and after word segmentation, entity recognition, and intent parsing, obtaining standardized question-answering requests, further inputting them into the cybersecurity large language model, which calls the standardized cybersecurity knowledge corpus, and generates preliminary question-answering results through semantic understanding and reasoning; S4, calling the standardized cybersecurity knowledge corpus to perform consistency verification on the preliminary question-answering results, if consistent, using it as the final result, otherwise corrected by the cybersecurity large language model to obtain the final result.

The invention relates to the technical field of multi-modal corpus construction, in particular to a multi-modal corpus construction method and system. Comprising a multi-source heterogeneous data acquisition module, a multi-modal preprocessing and analysis module, a cross-modal automatic alignment and association module, a translation feature labeling and management module, a structured storage and index module, an AI intelligent retrieval and teaching application module, a semantic optimization multi-modal adaptation module and a user display terminal. The multi-source heterogeneous data acquisition module comprises a data source interface unit and a metadata extraction unit; a multi-modal corpus system deeply integrated with a translation theory is constructed to directly serve the core demand of a translation subject, and a triple screen array mode is adopted to cooperate with display, so that display content is reasonably distributed, a user can check and operate in multiple dimensions, the use experience and the working efficiency are greatly improved, and the user experience and the working efficiency are improved. And various requirements of translation teaching and the like are met.

This invention discloses a method, apparatus, computer device, and storage medium for data augmentation in multi-turn dialogue. The method includes: acquiring text corpus to create a knowledge point tree structure; traversing the nodes in the tree structure to create a knowledge singly linked list path; performing question-and-answer pairing based on the nodes in the singly linked list path to obtain a first question-and-answer pair; acquiring seed tasks and filling them into the singly linked list path; generating answers for the singly linked list path using a generative language model; then pairing the nodes and answers in the singly linked list path to obtain a second question-and-answer pair; concatenating the first and second question-and-answer pairs into a third question-and-answer pair and converting it into a question-and-answer pair vector to construct a dataset for augmenting and training the generative language model to obtain a dialogue prediction model for dialogue prediction. This invention solves the problem of low accuracy in dialogue prediction generated by generative language models in multi-turn dialogue tasks by obtaining a dialogue prediction model through data augmentation training and performing dialogue prediction.

The application discloses a bilingual question and answer method based on knowledge retrieval and generation collaborative training, and belongs to the technical field of intelligent question and answer, and comprises the following steps: obtaining a corpus g0 and a question and answer library; mapping sentences in the g0 to the same semantic space to form a set A; performing regular hierarchical clustering on the set A to generate a semantic prototype vector of each clustering cluster; generating a knowledge graph g1 and a causal diagram g2 to form a retrieval source set G; marking a sample set and a main retrieval source for a question q of the question and answer library; constructing a knowledge retrieval enhancement model and training the knowledge retrieval enhancement model into a bilingual question and answer model, which is used in intelligent question and answer. The application introduces hierarchical semantic aggregation, cross-source adaptive gating and causal biasing retrieval mechanism, solves existing problems such as scattered knowledge, language alignment and self-learning update, provides a new technical path for low-resource cross-language question and answer, and greatly improves the quality and application value of a cross-language knowledge enhancement question and answer system.

The application discloses a dynamic intelligent agent retrieval tree-based time-sensitive news retrieval method and system, an electronic device and a storage medium, and belongs to the technical field of information retrieval and artificial intelligence. The technical scheme of the application comprises the following steps: constructing a retrieval tree with a user query as a root node, the tree being expanded in semantics through multi-agent cooperation; obtaining a current news corpus and constructing a representative evaluation agent; based on the evaluation agent, dynamically selecting an optimal sub-tree from the retrieval tree that meets a preset condition; and using the optimal sub-tree to search the full news corpus to obtain a target news document. The application decouples high-cost semantic expansion and high-concurrency online retrieval, reduces query delay and computing cost, and dynamically selects an optimal sub-tree to adapt to changes in news content in real time, thereby improving the timeliness and relevance of the retrieval result.

The application discloses a text processing method and device, related equipment and computer program product. A data set composed of three different data of a target language is used to train a model in a LoRA fine-tuning manner to obtain three low-rank adaptive models corresponding to the three data sets respectively. The first data set uses real voice-text pair data, the second data set has the same text as the first data set and the voice is a synthesized voice, and the third data set includes collected high-resource text corpus and the voice is a synthesized voice. A task arithmetic merging strategy is used to calculate the sum of the first and third low-rank adaptive models and the difference of the second low-rank adaptive model to obtain a merged low-rank adaptive model. The voice recognition model of the target language is composed of the merged low-rank adaptive model and a pre-trained voice recognition model. The model is optimized by fully utilizing the synthesized data without affecting the real data effect, and the voice recognition effect of the low-resource language is improved.