51 results about "Generic knowledge" patented technology

The invention discloses an intelligent policy question-answering method and system based on retrieval enhancement generation and a medium. The method comprises the following steps: constructing a universal knowledge base and a plurality of mutually independent domain knowledge bases; in response to user input, the following steps are executed: performing routing analysis on the user input through a first large language model to generate a structured routing decision; retrieving the general knowledge base to obtain related general knowledge text segments and vectorization expressions thereof; according to the routing decision, a domain knowledge base corresponding to the at least one policy domain identifier is retrieved in parallel, and related policy text fragments corresponding to the at least one domain knowledge base and vectorization expressions of the related policy text fragments are obtained; obtaining an initial answer set based on retrieval results of the general knowledge base and the domain knowledge base; and performing intelligent fusion processing on the initial answer set through a second large language model, and generating and outputting response content. According to the method, the problem of knowledge updating lag is effectively solved, and the accuracy, timeliness and cross-domain specialty of policy questions and answers are improved.

The invention discloses a data scarcity industrial vertical field-oriented knowledge-guided retrieval enhancement generation method, belongs to the technical field of natural language processing and industrial intelligence crossing, and can improve the retrieval accuracy and generation reliability of a large language model in an industrial scene. According to the method, a'sparse vector + dense vector 'mixed knowledge base is constructed, and general knowledge and industrial field texts are fused; after the model generates an initial text, judging whether external retrieval is needed or not through multi-dimensional evaluation token confidence; extracting attention weights for the low-confidence tokens, and screening key tokens to generate a retrieval query; dynamically adjusting the weight of a retriever based on a BGE-M3 model, and optimizing a retrieval result through reordering; the retrieval knowledge is converted into a context with an index, and a prompt template is constructed to generate a correction value iteration calibration text; and finally, optimizing the text format, and generating a structured response meeting industrial requirements. The method solves the problems of lack of professional knowledge of large language models in the industrial field, shallow retrieval and generation fusion, lack of knowledge calibration mechanisms and the like.

The present invention generally relates to the field of robotics, and, more particularly, to a method and system for task anticipation by integrating large language models and classical planning. Conventional methods for task anticipating use data-driven deep network architectures and Large Language Models (LLMs) for task estimation but they do so at the level of high-level tasks and require a large number of training examples. Thus, embodiments of present disclosure provide a method and system for task anticipation by integrating large language models and classical planning. The disclosed method and system leverages the generic knowledge of LLMs through a small number of prompts to perform high-level task anticipation, using the anticipated tasks as joint goals in a classical planning system to compute a sequence of finer granularity actions that jointly achieve these goals.

The invention provides a method and a device for evaluating a space moving track, and relates to the technical field of artificial intelligence. The invention discloses a spatial movement track assessment method, which comprises the following steps of: generating track-risk image data according to a spatial movement track and a corresponding risk map; inputting the trajectory-risk image data and a pre-constructed task cue word into a pre-constructed visual-language large model to obtain an evaluation result; and transmitting the evaluation result to the target end. According to the technical scheme provided by the embodiment of the invention, the space movement track and the risk map which need to be evaluated are uniformly converted into the image representation as the input of the vision-language large model, and the image representation and the task cue word are input into the vision-language large model; the general knowledge and the cross-modal reasoning ability of the vision-language large model are utilized to finish track evaluation in real time, and the expansibility and the interpretability are good.

The embodiment of the invention provides a multi-modal universal model self-evolution method and system based on incremental learning, and the method comprises the steps: receiving a current target task and multi-modal data corresponding to the current target task, and determining a corresponding trained target universal model; extracting data features of the multi-modal data to obtain multi-modal data features, and processing the multi-modal data features and the multi-modal data through a target special model to obtain a target result; historical multi-modal data features corresponding to the historical multi-modal data are obtained, and general-purpose model parameters of the trained target general-purpose model are updated; and monitoring the performance index of the trained target general-purpose model according to a preset monitoring period, and if the performance index corresponding to the current preset monitoring period does not meet the preset performance index, adjusting the trained target general-purpose model according to a preset adjustment mode. According to the method, the defect of insufficient collaboration between general knowledge and special knowledge is effectively overcome, and the collaboration effect and the updating effect of updating according to new knowledge are remarkably improved.

The embodiment of the invention discloses a training method and device of a vertical domain question and answer model, equipment and a storage medium, and belongs to the technical field of artificial intelligence. The method comprises the steps that a first data set, a second data set and a third data set are obtained, the first data set is a data set representing general knowledge, the second data set is a data set representing vertical domain knowledge, and the third data set is a data set representing question and answer tasks based on the vertical domain knowledge; training a vertical domain model based on the first data set and the second data set, wherein the vertical domain model is a model obtained by adding at least one decoding layer in a basic model; and on the basis of the trained vertical domain model, a vertical domain question and answer model is trained based on the third data set, the vertical domain question and answer model is a model obtained by adding a low-rank adjustment structure in the trained vertical domain model, and the vertical domain question and answer model is used for executing a question and answer task based on vertical domain knowledge. By adopting the scheme provided by the invention, the training quality of the vertical domain question and answer model can be improved.

The invention discloses a privacy enhancement continuous learning method based on lexical-level differential privacy and memory shaping, which is characterized by comprising a lexical-level dynamic differential privacy module and a privacy guide memory shaping module, the lexical element level dynamic differential privacy module adaptively allocates privacy budget according to semantic sensitivity of different lexical elements in a text, and introduces Gaussian noise in an embedded layer; the privacy guide memory shaping module comprises memory regularization and privacy perception forgetting; the memory regularization is used for dynamically constraining model parameters; the privacy aware forgetting is used to selectively impair memory for highly sensitive information. Compared with the prior art, the method has the advantages that higher accuracy and lower forgetting rate are maintained while privacy is guaranteed, forgetting of sensitive information and long-term maintenance of general knowledge are realized, disastrous forgetting in continuous learning is effectively relieved, and the method is applicable to privacy sensitive application scenes such as recommendation systems, medical diagnosis and financial analysis.

The invention provides a large language model staged pre-training method and system. The method comprises the following steps: training a Transform model by using a basic data set, optimizing a negative logarithm likelihood target, and adopting an AdamW optimizer and a cosine attenuation learning rate; continuing training by using the universal knowledge data set based on the first-stage parameters; and weighting the professional data in the training field by adopting an oversampling strategy. By structuring a training target and a data type, the model can efficiently learn language basis, general knowledge and professional skills in stages. Experiments show that according to the method, the training efficiency of the model in the basic stage is improved by 40%, the overall training time is shortened by 30%, and meanwhile the accuracy rate of tasks in the professional field is 15%-20% higher than that of traditional end-to-end training. Finally, model parameters are evaluated through professional ability, and both universal language understanding and domain specialty are achieved.

The invention discloses a method and device for relieving catastrophic forgetting of a language model, equipment and a storage medium, and relates to the technical field of machine learning. According to the method, linear interpolation is carried out between the pre-training model parameter set and the fine-tuning model parameter set, so that a linear interpolation point of the model can still keep a relatively low loss value when the model learns new task knowledge; then the pre-training model parameter set and the fine-tuning model parameter set are added to form a fusion model parameter set, and the linear interpolation operation only performs one-time simple arithmetical operation after fine-tuning is completed without introducing any additional trainable parameters or changing the model structure, so that when the model adapts to a new task, the model can be quickly and accurately trained. Original general knowledge and performance on old tasks can be reserved to the maximum extent, and meanwhile complexity and expenditure of a model reasoning stage are not increased.

The embodiment of the invention discloses a large model training data generation method, and the method comprises the steps: determining the risk control capability of a target large model and the description of a knowledge field according to a data generation task requirement of a risk control business scene, so as to determine a core concept and a format requirement based on a determined task portrait, the core concept is used for obtaining matched rich knowledge from a universal knowledge base, the format requirement is used for reprocessing the obtained knowledge, and a training sample is obtained by generating a large model to train a target large model for executing risk control business. Through the automatic training sample generation process corresponding to the demand, the existing generation and labeling problem depending on manpower is avoided, the efficiency is improved, the quality difference between samples can be reduced, and the accuracy of the training model is improved.

This invention relates to the field of computer vision technology, providing a method for training a behavior recognition model, a behavior recognition method, and an apparatus. The method and apparatus utilize a frozen pre-trained image-text model (a second video feature extractor and a text encoder) as a general knowledge base to guide a finely tuned student model (a first video feature extractor) with temporal modeling capabilities. A multi-head residual projection network is used for feature-level knowledge transfer and representation alignment. The behavior recognition model trained by this method can accurately understand dynamic behaviors in videos and possesses strong zero-shot reasoning ability (i.e., generalization ability), effectively recognizing behavior categories not seen during the training phase. Furthermore, the entire model framework is clear, training is stable, and the final model exhibits excellent recognition performance for both known and unknown behavior categories under open-vocabulary settings.

The invention relates to the technical field of student intelligent questioning and answering, and discloses a student intelligent questioning and answering method and system based on a large language model and rapid retrieval, and the method comprises the steps: combining a student portrait to generate question vectors, rejecting noise points of a question vector set, obtaining hot questions, and storing the hot questions into a hot knowledge base after the hot questions are expanded by experts; generating a summary of the knowledge document by using a large language model, and adding the summary to each text block; generating a block summary for each text block, and storing the block summary and the text blocks into a universal knowledge base; inputting questions proposed by students into the large language model to generate a thinking chain; for sub-questions needing to be retrieved, firstly retrieving the hotspot knowledge base, if the sub-questions are not matched with the hotspot knowledge base, retrieving the block summary of the general knowledge base, and obtaining corresponding text blocks through mapping to obtain retrieval results; and guiding the big language model to fuse the thinking chain and the retrieval result to generate a structured answer. According to the method, the summary information is introduced after the text blocks are segmented, so that the information loss in the text segmentation process is reduced, and the reasonability of knowledge base construction is improved.

The invention discloses an AI model illusion suppression method and system based on retrieval enhancement generation and cue word engineering collaboration, and aims to solve the illusion problem of an existing large language model and the defects that a traditional scheme is high in cost, low in retrieval content utilization rate, universal knowledge conflicts, high in generation randomness and the like. The method comprises the following steps: constructing a high-precision domain vector index database, performing similarity retrieval and noise filtering on user query, generating a multi-level structured cue word containing role definition, thinking chain constraint and the like, dynamically switching a general knowledge shielding or fusion mode based on recall quality, and outputting a final result through multi-path reasoning sampling and consistency verification. According to the method, underlying model parameters do not need to be modified, the illusion rate is remarkably reduced, generated content is factual and accurate, logic coherence and traceability are achieved, deployment is flexible, cost is low, and the method is suitable for factual question and answer scenes in multiple fields such as medical treatment, industry and legal consultation.

An object of a system according to an embodiment is to provide optimal advice to a user.SOLUTION: A system according to an embodiment includes a general-purpose AI, an individual-specialized AI, and a cooperation processing unit. The AI having general-purpose intelligence has general-purpose intelligence. The individual-specific AI learns a conversation record or an action history of the user. The cooperation processing unit generates an optimal advice for the user through cooperation between a AI having general-purpose intelligence and the individual-specialized AI.SELECTED DRAWING: Figure 1

The machine learning in the neural networks module can analyze an annotation and its metadata on the annotation made by a first user on a first computing device to make an embedding regarding the annotation and then cooperate with the persistence knowledge store to store the embedding of the machine learning's understanding of the annotation and its metadata. The delivery module can proactively push a notice regarding a potentially related embedding out to a second user on a second computing device based on a threshold amount of relatedness between one or more factors of i) a first task undertaken by the first user and a second task undertaken by the second user, ii) a role of the first user and a role of the second user, and iii) a subject matter of the embedding to a subject matter of a task undertaken by the second user.

The invention discloses a project management system development method based on low code configuration, relates to the technical field of computer software, and remarkably improves the prediction adaptability and practicability of a project management system in a new project type through a low code configuration and meta-learning fused architecture. A low-code interface allows non-technical users to intuitively configure project parameters, such as a construction period or a risk index, so that a system customization threshold is reduced, and enterprises can quickly respond to innovative project requirements without depending on professional development teams; the meta-learning module uses general knowledge of a pre-training model to complete model adjustment only with a small amount of new data, and overcomes the defect of weak generalization ability of a traditional model in a data sparse scene; the few-sample adaptive mechanism is optimized through internal and external loops, so that the model is ensured to be fast converged while keeping stability, and the training time and resource consumption are reduced; the verification process in the deployment stage enhances the reliability of the system, and the dynamic consistency verification can detect the output deviation of the model in time.

The invention provides an intelligent knowledge question-answering method, device and equipment, and relates to the technical field of artificial intelligence. The method comprises the following steps: acquiring a plurality of sample sequences and a general knowledge vector set of power vertical domain knowledge, wherein the sample sequences comprise a plurality of word units and time steps; performing word unit prediction analysis on the global context and the local context corresponding to each time step to obtain an alignment loss item; the alignment loss item represents the prediction difference between the global context and the local context; based on the general knowledge vector set, the historical model parameters and the currently updated model parameters, performing calculation to obtain knowledge loss items; the knowledge loss item represents the disturbance of the model parameter change on the general knowledge; based on the sample sequence, constructing a loss function by using a prediction loss item, an alignment loss item and a knowledge loss item, and training to obtain a question and answer model; and predicting the question of the user through the question and answer model to obtain the answer of the question. The answer prediction accuracy can be improved.

The application provides a construction method of a knowledge distillation model suitable for power line fitting detection, comprising the following steps: inputting image data of power line fittings into a basic knowledge distillation model to extract multi-scale feature data; performing redundant compression processing and convolution kernel learning on the multi-scale feature data to obtain a light knowledge vector and fitting detection general knowledge; training the basic knowledge distillation model using the knowledge and teacher-student global knowledge prototypes, and performing differential distillation training on classification heads and regression heads based on objects and information effectiveness respectively in the process to obtain a preliminary knowledge distillation model; and combining a comprehensive loss function to perform end-to-end weighted optimization on a student model in the preliminary knowledge distillation model to obtain an optimized final knowledge distillation model. The application can solve the problems of high model complexity and insufficient adaptability to complex power fitting detection scenes in related art knowledge distillation-based detection methods.

The invention belongs to the field of customer service, and particularly relates to a cross-library question answering system and method based on natural language, and the system comprises a multi-terminal access layer which is used for receiving natural language questions inputted by different types of clients; the intelligent matching core layer is used for performing semantic analysis and intention recognition on an input natural language question, outputting a retrieval instruction containing a region tag and a user intention, and obtaining corresponding knowledge entries from the double-layer knowledge base layer according to the retrieval instruction; the double-layer knowledge base layer comprises a nationwide universal knowledge base and a regional localization database, the nationwide universal knowledge base is used for storing nationwide universal rules, and the regional localization database is used for storing local rules in different regions; and the result output layer is used for integrating the retrieved knowledge entries and returning a final answer to the user through the multi-terminal access layer.

The invention discloses a federal recommendation method based on dual decoupling, which relates to the field of federal learning, and comprises the following steps: performing dual additive decoupling on user embedding and article embedding, and decomposing each representation into a private component reserved locally and a universal component globally shared by a server. Private components capture personalized preferences of users and unique understanding of items, and generic components learn generic knowledge of a group of users and a set of items. When a client is trained locally, each component is updated by minimizing a composite loss function including prediction loss, a component differentiation regular term and a global component rarefaction regular term, and the weight of the regular term is dynamically adjusted. And the client only uploads the universal component update to the server for aggregation. Through a dual decoupling and dynamic regularization mechanism, on the premise of not leaking original data, personalized and common information is effectively separated and learned, the recommendation accuracy and the personalized level are remarkably improved, meanwhile, the communication cost is greatly reduced, and privacy protection is enhanced.

The invention discloses a general knowledge construction method and system based on vertical and horizontal multi-dimensional linkage and a storage medium. The method comprises the following steps: receiving a user operation instruction; acquiring data of longitudinal analysis dimensions (including a basic information layer, a logic association layer and a significance principle layer) and transverse application dimensions (including a resource attribute layer, a rule technique layer and a comprehensive scheme layer) from a domain knowledge graph database according to the instruction; an intelligent linkage mechanism is triggered through a rule engine, and the intelligent linkage mechanism comprises longitudinal internal linkage, transverse internal linkage and cross linkage (including learning-to-use, use-to-learn and reflection backtracking). The system comprises a processor, a memory and a corresponding software module. Through longitudinal and transverse multi-dimensional knowledge organization and intelligent linkage, seamless connection and deep fusion of knowledge from analysis to application, from learning to practice and from problems to principles are realized, and the efficiency and effect of knowledge acquisition, understanding and application are improved.

The embodiment of the application provides a kind of based on the collaborative training method of multi-modal general-purpose model of federal learning, by receiving data to be retrieved, the data features of data to be retrieved are extracted, and multi-modal data index is established based on data features;Based on multi-modal data index, data features are processed through attention mechanism, and fusion data features are obtained;Fusion data features are input into retrieval general-purpose model, and knowledge extraction is carried out on fusion data features by knowledge extraction layer of retrieval general-purpose model, to obtain general knowledge representation;General knowledge representation is input into retrieval special-purpose model, and general knowledge representation is processed based on retrieval special-purpose model, to obtain the retrieval result corresponding to data to be retrieved, and the general model parameters of retrieval general-purpose model and the special model parameters of retrieval special-purpose model are adjusted based on retrieval result.The method effectively solves the deficiency in the aspect of difficult to effectively fuse multi-modal heterogeneous data, and significantly improves the generalization ability of model.

The invention relates to the technical field of intelligent questioning and answering, and discloses a fire safety knowledge questioning and answering method and a machine questioning and answering system. The method comprises the following steps: performing multi-level semantic analysis on a fire safety problem input by a user to generate structured problem data; matching the data with the dynamically updated fire safety knowledge graph, and intelligently triggering dual-mode response according to a matching result; when the matching succeeds, immediately extracting answer elements; and when the matching is incomplete, starting a deep inference engine to perform intention splitting, and linking a user portrait database to obtain personalized information so as to complement the context. And fusing the general knowledge elements and the personalized information to construct an answer framework, and generating and outputting natural language answers. According to the method, the response efficiency and depth are considered through a dynamic routing mechanism, and active personalized information is utilized for complementing, so that the conversion from universal response to precise customized suggestions is realized, and the practicability and accuracy of a question answering system are improved.

The invention provides a sub-graph retrieval method and device, electronic equipment, a medium and a computer program product, and can be applied to the technical field of big data. A pre-constructed general knowledge graph is analyzed into text information, the text information comprises m paragraphs, and each paragraph comprises ni statements; matching the obtained retrieval demand with the text information, and taking a matched statement as reconstructed text information; from the reconstructed text information, a sub-graph may be constructed. According to the method and the device, the knowledge graph information is restored into the text information, and the text information has higher specialty for field and business description and is more humanized in description, so that the method and the device are more in line with language expression habits of retrieval requirements input by a user, information retrieval is performed in the restored text information, key information cannot be omitted, and user experience is improved. The retrieval result is more comprehensive and accurate, the sub-graph for a certain specific field and / or a certain specific business range can be constructed by taking the retrieved statement as the reconstructed text information, and rapid analysis of specific problems is facilitated.

The application discloses a large language model field adaptation fine-tuning method based on knowledge dimension and task dimension collaborative courses, and belongs to the technical field of artificial intelligence and natural language processing. The application comprises the following steps: firstly, a general field training set and a target field training set are constructed, and the general field training set is filtered to remove the field, so as to reduce the field overlap between the two types of corpus; secondly, the general field samples and the target field samples are mapped to a unified semantic space, a plurality of knowledge clusters are obtained through knowledge clustering, and a knowledge specificity score is calculated according to the distance between each knowledge cluster and the general field center, so as to construct a knowledge dimension course from general to professional; then, the target field training samples are annotated for task complexity, and are merged to form three types of task complexity levels, i.e. low, medium and high, so as to construct a task dimension course from shallow to deep; finally, a two-dimensional course sequence is generated by combining the knowledge dimension sorting and the task dimension sorting, an exponential mixing strategy is adopted to dynamically improve the proportion of the target field samples in the training batch, the basic large language model is fine-tuned for field adaptation, and a target field adaptation model is obtained. The application can improve the performance of the target field while effectively maintaining the original general knowledge and reasoning ability of the model, reduce the catastrophic forgetting, improve the training stability and the cross-field application ability.

The present disclosure relates to a natural language processing method and device, a model training method and device, and a storage medium. The present disclosure pre-trains a machine learning model through each triple, so that the pre-trained machine learning model can seamlessly and naturally process various natural language understanding tasks under the machine reading comprehension paradigm. In addition, because the data format used for model training in the pre-training stage is consistent with the data format used for model training in the fine-tuning stage, the pre-training target and the fine-tuning target are the same, thereby enabling seamless connection between the pre-training stage and the fine-tuning stage. After pre-training the model using a large amount of low-cost data, the pre-trained machine learning model can be calibrated using a small amount of target task data, thereby enabling the general knowledge learned in the pre-training stage to be successfully transferred to the fine-tuned model, and ensuring the accuracy of the fine-tuned model.

The application provides a code generation method and device based on an up-mixing mixed diversity expert large model, and belongs to the field of artificial intelligence. The method comprises the following steps: obtaining code description information; inputting the code description information into a trained code large model to obtain code output by the code large model; the code large model is a large model based on an up-mixing mixed diversity expert; the code large model is constructed by adding a shared expert and multiple ordinary experts on the basis of a dense model; the shared expert reuses parameters of an FFN module of the dense model; the shared expert is used for extracting general knowledge of word elements; the ordinary expert is initialized by using a random initialization method; and the ordinary expert is used for extracting special knowledge of word elements. The code large model is constructed by adding a shared expert and multiple ordinary experts on the basis of a dense model, a diversity up-mixing mechanism is adopted, and the accuracy of generated code is improved.

The embodiment of the invention provides a multi-modal general-purpose model cooperative training method based on federal learning, and the method comprises the steps: receiving to-be-retrieved data, extracting data features of the to-be-retrieved data, and building a multi-modal data index based on the data features; processing the data features through an attention mechanism based on the multi-modal data index, and obtaining fused data features; inputting the fusion data features into a retrieval general model, and performing knowledge extraction on the fusion data features through a knowledge extraction layer of the retrieval general model to obtain general knowledge representation; and inputting the general knowledge representation into the special retrieval model, processing the general knowledge representation based on the special retrieval model to obtain a retrieval result corresponding to the to-be-retrieved data, and adjusting general model parameters of the general retrieval model and special model parameters of the special retrieval model based on the retrieval result. According to the method, the defect that multi-modal heterogeneous data cannot be effectively fused is effectively overcome, and the generalization ability of the model is remarkably improved.

The invention relates to the technical field of data processing, and particularly provides a general knowledge graph construction method, system and device and a medium, and the method comprises the steps: constructing a general framework based on a modular hierarchical architecture, the general framework comprises a basic data layer, a model service layer and a knowledge service layer, and each layer is decoupled through a standardized interface; field self-adaptive fine tuning is carried out on the core extraction model in the model service layer through a field adapter, and efficient adaptation to a specific field is achieved; based on a dynamic Schema mapping mechanism, generating a knowledge extraction task according to an entity and relationship type defined by a user; calling a core extraction model to execute a task, and extracting domain knowledge from the multi-source heterogeneous data; knowledge is stored in the basic data layer, and derivative knowledge is generated by using an inference model of the model service layer, so that a complete knowledge graph is constructed. According to the method, the construction efficiency and the cross-domain adaptation capability are remarkably improved, and the deployment period is greatly shortened.

An intention recognition method is used for improving the efficiency of recognizing user intentions. In the method, aiming at task data input by a user, a plurality of concern points related to the task data are analyzed through a large language model, the large language model is promoted to identify user intentions corresponding to the task data under the different concern points respectively, and finally candidate intention results obtained under the concern points are synthesized, so that the user intentions corresponding to the task data are identified. And determining a final target intention result. According to the scheme, the roles under different attention angles are set for the large language model by analyzing the multiple attention points related to the task data, so that the large language model recognizes the user intention from different attention angles, the phantom problem of the large language model is solved, and the accuracy of the output target intention result can be ensured; moreover, the general knowledge storage and semantic comprehension capabilities of the existing large language model are utilized to the greatest extent, rules or training data do not need to be provided manually, and the efficiency of realizing intention recognition is effectively improved.