47results about How to "Semantic rich" patented technology

The invention discloses a Chinese text key information extraction method based on a pre-trained language model, which comprises the following specific steps of: (1) classifying key information to be extracted, easily concluding information categories forming rules, and extracting by using a regular matching method; and (2) extracting the named entities by using a sequence labeling model. (3) constructing the sequence labeling model by adopting a method of finely adjusting a pre-training language model, wherein firstly, a large-scale unlabeled text corpus is used for learning to obtain the pre-training language model, and word boundary features are introduced in a pre-training stage; (4) replacing the data content matched by using the rule with the corresponding rule template label so as tocomplete fusion of rule matching and the deep network; and (5) performing fine adjustment on the pre-trained language model according to the marked training data, and migrating the pre-trained language model to the sequence marking task of the named entity. According to the method, text context semantic features can be effectively extracted, and each information type can be effectively identifiedin a complex information type scene.

The invention discloses methods for labeling and searching advanced semantics of images based on network hot topics and a device. The method for labeling the advanced semantics of the images comprises the step of: searching images similar to the entity semantics of images to be labeled and accompanying texts by using an entity semantic work of the images to be labeled based on a search engine of text key words, then extracting topics from the accompanying texts, establishing association relationship among topics, among images, and between images and topics, based on the association relationship, classifying the images with similar topics and visual features into a type, and classifying similar topics corresponding to the images with the similar visual features into a type; and selecting image types which are the most similar to the images to be labeled in the visual features of the images, and taking the corresponding topics as hot topics. According to the invention, by means of the process, the advanced semantic labeling the images is realized, and in addition, the obtained advanced semantics can be used to accurately describe the images to be labeled through denoising.

The invention discloses a knowledge graph representation learning method fusing entity description, hierarchical types and text relation information. The knowledge graph representation learning methodcomprises the following steps: firstly, preprocessing a knowledge base and a corpus, and extracting structured information (triple), entity description information, hierarchical type information andtext relation information; secondly, learning the representation based on the structured information by using a translation-based model TransE; learning a representation based on the entity description information by using the CNN; constructing a mapping matrix of an entity hierarchical type by using WHE, learning a representation based on hierarchical type information, learning a representation of sentences by using a position-based PCNN, allocating a weight to each sentence by using a sentence level attention mechanism, and learning a representation based on text relation information; and then constructing a total energy function and a total loss function of the model according to the representation based on the four kinds of information. The knowledge graph representation learning method carries out learning on the model parameters in combination with a loss function, and finally obtains optimized entities and relation representations.

The invention provides an image change detection method based on a depth-separable convolution network, which is used for solving the technical problem of low detection accuracy existing in the conventional image change detection methods. The implementation steps include: constructing a training sample set, a verifying sample set and a test sample set; using a variant U-Net of a full convolution network as a basic network to establish a depth-separable convolution network; constructing a loss function of the trained depth-separable convolution network; training, testing and verifying the depth-separable convolution network; and using the verified finally-trained depth-separable convolution network to perform test to obtain a change detection result graph. The invention has rich image feature semantics and structure information extracted by the depth-separable convolution network, has strong image expression ability and discrimination, can improve the change detection accuracy, and canbe used in the technical fields of land cover detection, disaster assessment, video monitoring and the like.

The invention discloses a sub-pixel characteristic point detection-based image matching method. The method comprises the following steps of: performing down-sampling on an image to be detected until the smaller one of the length and the width of the image to be detected is less than 8 pixels, acquiring and performing geometric progression variance Gaussian filtering on each down-sampled image to obtain continuous and progressive Gaussian blur images of which the maximum value of the variance is 2, establishing and acquiring the Gaussian pyramid of the image to be detected, evaluating a second derivative Ixx in x direction, a second derivative Iyy in y direction and a second derivative Ixy in xy direction; calculating the Hessian matrix of the Gaussian pyramid of the image to be detected and all local maximum value points of a Harris function value; and fitting a three-dimensional quadratic function with each local maximum value point on the Harris function value and 26 neighborhood points around each local maximum value point, updating the position of an extreme point of the quadratic function by using the floating number position of the extreme point, performing sub-pixel accuracy positioning modification on the position of the extreme point and outputting all characteristic points.

The invention relates to a method for determining an area where a traffic target is located based on a traffic video data image. The method is technologically characterized by including the steps of recording a focus point produced when the driver observes a piece of real traffic video, training a traffic target detection template, conducting convolution through the traffic target detection template and the video, extracting out features, based on an object, of the video, extracting motion features of the video through a space-time direction filter, using pixels of an eye focus position as a positive sample, using the pixels of positions except the eye focus position as negative samples, and training a classifier of a support vector machine through a machine learning method according to labels of the samples and features of the samples. Therefore, for any given traffic driving video, object level features and the motion features are inputted into the trained classifier of the support vector machine after being extracted, and the area where the traffic target is located in the traffic driving video can be forecasted.

A method and system for tagging original data generated by things in the Internet of Things (IoT). The method includes: performing relativity detection on acquired Web messages to obtain Web messages related to various events; acquiring address information contained in the related Web messages; determining adjacent things for various events based on the acquired address information; and tagging original data generated by the adjacent things determined using at least partial content of the related Web messages as metadata. The system includes modules for carrying out the above method.

The invention belongs to the technical field of Chinese medical named entity identification, and provides a Chinese medical named entity identification method and system, a storage medium and equipment. The Chinese medical named entity recognition method comprises the steps of obtaining clinical text data; converting the clinical text data into a character embedding representation, a medical concept embedding feature vector and a cross-language Chinese embedding representation of a medical text respectively, and splicing the character embedding representation, the medical concept embedding feature vector and the cross-language Chinese embedding representation to obtain a multivariate data fusion feature vector; inputting the multivariate data fusion feature vector into a named entity recognition model based on multiple graphs, and recognizing a Chinese medical named entity type, wherein the named entity recognition model based on multiple graphs comprises a multi-graph network and an LSTM-CRF model, the multi-graph network is used for receiving a text graph formed by taking a multivariate data fusion feature vector as a node, outputting a final state of the node and transmitting the final state to the LSTM-CRF model, and the LSTM-CRF model outputs a recognition result. The Chinese medical named entity identification accuracy is improved.

The invention provides a network representation learning method based on multi-order proximity similarity. Compared with the traditional network representation based on structure analysis which only considers the relation between the first-order proximity similarity and the second-order proximity similarity, the invention focuses on the modeling of high-order adjacent similarity between nodes, thecalculation methods of different kinds of indirect neighbor similarity are designed respectively, especially considering that the information will fade with the distance in the process of network propagation, therefore, the method of the invention can predict the different neighboring nodes of the current node, more accurately find the neighboring nodes with the largest degree of correlation withthe target node, so that the expression vector with richer semantics and higher reliability and authenticity can be obtained.

The invention belongs to the machine learning technical field and relates to a multi-mark classification method, a multi-mark classification device, a medium and a computing device. According to the embodiments of the present invention, after the original positive example set and original negative example set of marks are obtained; by means of classification alignment, specific attributes and operation for inserting specific attributes of related marks are determined, so that the specific attributes can be adopted to indicate correlations between the marks, and therefore, the data and semantics of the marks can be enriched; and therefore, multi-mark classification is more accurate than a method only adopting a single mark in the prior art. For example, deserts are correlated to camels, a picture which is dominated by a camel and contains a small part of a desert is classified as a desert picture; and if a picture contains a lake at dusk, and if the reflection of a sinking sun is in the lake, the picture would be classified as a lake picture in the prior art, while, the reflection of the sun in the lake is correlated to the sinking sun, the picture will be classified as a dusk scenery picture by the method of the invention.

The invention discloses a domain-adaptive foggy day image target detection method and device, and belongs to the technical field of target detection, and the method comprises the following steps: carrying out preprocessing on an obtained target detection data set; carrying out model multi-scale performance reconstruction on the backbone network; training the reconstructed backbone network by using the preprocessed target detection data set to obtain a target detection model; building a domain classifier for the target detection model; training the target detection model with the domain classifier being built by adopting the foggy day image and the preprocessed target detection data set to obtain a domain-adaptive detection model; and performing target detection on the foggy day image to be detected by using the domain-adaptive detection model. The method and device have the advantages of being high in detection precision, high in real-time performance and applicability, low in omission ratio and the like, and the performance of the detection model in a foggy day scene is improved.

The invention discloses an image feature extraction method based on weighted depth features, and the method comprises the following steps: selecting an existing network model for the pre-training of an image data set, and then removing a network prediction layer in the network model, and obtaining a final network model; inputting the to-be-extracted image into the final network model for forward calculation, and extracting convolution layers in front of all pooling layers in the final network model as a depth feature map of the image; calculating a feature aggregation vector for each convolution layer, and performing extraction and zero-mean normalization processing on each feature aggregation vector to obtain a depth feature vector; carrying out dimensionality reduction on all the depth feature vectors, and distributing the corresponding weight parameters to the depth feature vectors after dimensionality reduction to serve as indexes of each convolution layer; and fusing the depth feature vectors with the weight parameters to obtain depth image features. The method is high in universality, good in obtained feature robustness, high in expression capability, rich in semantic information and easy and convenient to calculate.

The invention belongs to the field of natural language processing, and particularly relates to a BERT and external knowledge-based generative automatic abstracting method, which comprises the following steps of: acquiring document data, and acquiring keywords corresponding to the document data through a TextRank module; inputting the document data into the BERT model for encoding to obtain encoded document information; external knowledge is retrieved from an external knowledge module through keywords, and the external knowledge and document information are fused through a gating mechanism; the fused information is input into a Transform model to be decoded, and an abstract is generated; according to the method, the BERT model is used for encoding the document data to capture more context information and internal information, the encoding quality is improved, the keyword is used for obtaining external knowledge to be fused with the document information, the Transform model is used for enriching the semantics of the generated abstract, the smoothness and integrity of the generated abstract are improved, and the high-quality abstract is generated.

The invention provides a method for generating a Chinese word vector with multi-submodule information. The method is characterized in that a six-submodule information and attention mechanism fusion method which comprises words, characters, radicals, components, fonts and pinyin is adopted, an improved Chinese character embedding representation form is learned and fused into word embedding with proper weight, and a high-precision word vector is generated. Compared with the prior art, the method has the advantages that the appropriate weight is allocated to each part of sub-module information according to the attention mechanism, so that the weight of the sub-module with less semantic meanings is reduced, the weight of the sub-module is improved, the modules with richer semantic meanings areprovided, Chinese word embedding is improved, and considerable performance improvement is realized.

The invention discloses a solution method for a cognitive description program based on generation-verification. When one cognitive description program is input: firstly, a predicate dependency graph of the cognitive description program is firstly constructed based on morphology and grammar correctness analysis and safety check, block-shaped topological sorting of program blocks is generated, then a forward reasoning technology is used to sequentially generate instantiation program blocks according to the topological sorting, and finally the instantiation cognitive description program is generated; secondly, a constant-true constant-false character set of the instantiation cognitive description program is obtained and used for deleting redundancy rules and redundancy characters so as to simplify the program; then, dependency relationship of subjective characters is confirmed according to the predicate dependency graph, and possible solution of the program is generated in a heuristic mode according to the dependency relationship of subjective characters; and after that, subjective character values are verified to be correct or not, and whether the possible solution is of the cognitive description program or not is confirmed. The whole solution method adopts a backtracking process to generate all of possible solutions and verify obtained all of solutions.

The invention discloses a human body intention recognition method. The method comprises the steps of collecting a feature signal of a current human body in real time; generating multi-source data features corresponding to the current human body and fixation point coordinates selected by eyes based on the feature signals; recognizing the multi-source data features and fixation point coordinates selected by eyes, and generating voice texts corresponding to the multi-source data features and scene image description texts corresponding to the fixation point coordinates; performing entity extraction on the voice text and the scene image description text to generate entity fragments corresponding to the voice text and the scene image description text; processing the entity segment by adopting aco-exponential resolution algorithm to generate a target object; and generating a human body intention recognition result based on the voice text, the scene image description text and the target object. Therefore, by adopting the embodiment of the invention, the recognition result is obtained after the mouth-eye cooperative interaction information of the specific scene is processed, so that the accuracy of recognizing the human body intention by the machine is improved.

The invention discloses a heterogeneous graph neural network representation method based on meta-paths, which comprises the following steps of: 1, determining a plurality of meta-paths of a target node, sampling different types of neighbor nodes of the target node in each meta-path, and grouping according to the types; 2, performing feature extraction, node initial heterogeneous content coding and feature aggregation on the neighbor nodes obtained in the step 1 to obtain heterogeneous neighbor information; 3, respectively aggregating heterogeneous neighbor information of neighbor nodes generated in each meta path, and obtaining corresponding embedded representation; 4, combining and optimizing the embedded representation in each meta path based on the attention mechanism again, and generating a final embedded representation of the target node.

The invention discloses a social recommendation method and system based on a graph convolutional network, and belongs to the field of machine learning and social recommendation. The method mainly comprises the steps of data set construction, neighborhood aggregation, entity representation generation, model training and prediction and the like. According to the invention, two types of information of articles and social friends interacting with a user side are considered at the user side, and two types of information of users and collaborative similar articles interacting with an article side are considered at the article side, so that a user-article interaction network, a user social network and an article collaborative similar network are modeled in a unified network; and the embedding is improved by explicitly modeling the high-order connectivity among the user-article, the user social network and the article collaborative similar network, so that the deep potential interest preference of the user on the article can be captured, the semantic-rich user / article representation is finally generated, and the purpose of improving the recommendation accuracy is achieved.

The invention provides a protocol reverse engineering system based on reinforcement learning and a working method thereof. The protocol reverse engineering system comprises a decoding environment module and a decoding intelligent agent module, wherein the decoding environment module receives the state of the previous moment and the current action output by the decoding intelligent agent module, and inputs a new state to the decoding intelligent agent module. Compared with the prior art, the method has the advantages that the protocol structure tree is used for representing the protocol structure, and the method is high in expression capacity, rich in semantics and capable of supporting a complex structure; and 2) a protocol reverse engineering method based on deep reinforcement learning isinnovatively provided, after a large amount of data is trained, the application range is wide, and reverse engineering can be carried out on data of different types of protocols. And 3) when the trained decoding model executes protocol reversing, any priori knowledge is not needed, manual participation is not needed, and protocol reversing can be executed in a full-automatic manner.

The invention relates to the technical field of the Internet, and particularly discloses a title generation method and device based on concept extraction, equipment and a medium. The method comprises the following steps: carrying out keyword extraction on a target text to obtain a plurality of keywords; obtaining concept tags corresponding to the keywords from a concept tag library; obtaining a corresponding title template according to the concept tags, and filling the title template according to the concept tags to obtain a spliced title; and processing the spliced title based on a pre-trained title generation model to obtain a target title. The concept tags corresponding to the keywords of the target text contain richer semantics, and the title template is filled with the concept tags to obtain the spliced title, namely, the title framework containing the key information of the target text is obtained; then the target title is generated for the spliced title according to the title generation model, and finally the smooth and accurate target title which is easy to read and rich in semantics is obtained.

The invention discloses an intelligent customer service system based on an ontology, comprising an interactive terminal, a service logic module and a data management platform. The service logic moduleinputs questions and receives answers through the interactive terminal. The service logic module intelligently processes the questions input by the interactive terminal to obtain knowledge for answering the questions. The service logic module comprises an intelligent response engine and an ontology model, the intelligent response engine interacts with the interactive terminal, The data managementplatform realizes the management of the knowledge in the ontology model, the user management and the permission setting. The invention can intelligently 'understand' the problems of users, and adoptsa knowledge network to establish a knowledge base, which is convenient for maintenance.

The invention discloses a meta-path-based heterogeneous network similar node query method. The method comprises the steps: 1, generating a path greedy tree, expanding the greedy tree according to theinput source node and the short text description, and carrying out the semantic matching of the short text in the greedy tree expansion process; 2, determining a meta-path sequence; traversing a greedy tree to obtain an edge type sequence, determining a node type sequence according to the edge type sequence, traversing the generated greedy tree, and separating a path connected with an input node pair from the greedy tree; 3, calculating the importance of the meta-path: defining a computational formula of meta-path importance according to factors influencing the meta-path importance, and calculating the importance of the meta-path by means of the number of instance nodes in the greedy leaf nodes; 4, generating a query instance in combination with the plurality of meta-paths; wherein the instance node pairs conforming to the meta-path semantics have high similarity in the meta-path semantics; therefore, in order to obtain a query result instance, only a node pair with high similarity insemantics of each meta-path needs to be found.

The invention discloses a remote supervision relation extraction method based on PGAT and FTATT, and relates to the technical field of remote supervision relation extraction. The method comprises thesteps of obtaining an NYT data set, and dividing sentences containing the same entity pair in the data set into a packet; acquiring word vector representation of sentences in each package; extractingsequence features of sentences based on Bi-LSTM; extracting syntactic structure features of sentences based on PGAT; allocating weights to different sentences in the packet by using FTATT; performingweighted summation on the feature vector of each sentence in the packet and the weight coefficient of the sentence to obtain a feature vector of the packet; and according to the feature vector of thepacket, performing relationship classification on the entity pairs in the packet. According to the method, syntactic structure information of sentences can be captured by using PGAT, so that extractedsentence features contain rich information in the aspects of semantics and grammar, an attention mechanism is finely adjusted by adopting FTATT, noise data is dynamically discarded as much as possible, and the accuracy of relationship extraction is improved.

The invention discloses a steel product surface defect detection method based on model compression, and belongs to the technical field of steel surface defect detection, and the method comprises the following steps: S1, preprocessing a hot rolled strip steel surface defect image sample; s2, building a hot rolled strip steel surface defect detection network based on an anchor-free target detection framework; s3, carrying out model compression on the defect detection network, and training a model by adopting a linear strategy and a training sample; and S4, testing the defect detection network on the independent verification set and outputting a detection speed and a detection result. According to the invention, the steel surface defect detection network for identifying and positioning is firstly established, and then redundant parameters in the defect detection network are removed, so that the high storage space of the model and the consumption of computing resources are reduced, and the defect detection network is more stably and efficiently landed on a hardware platform.

The invention provides a method for predicting the referenced number of a paper by utilizing review opinions based on deep learning. The method comprises the following steps: training the review opinions of the paper by utilizing a deep component and a width component; the deep component comprising an abstract-comment matching mechanism and a cross comment matching mechanism and being used for learning deep features of review opinions; firstly, extracting comments related to an abstract through the abstract-comment matching mechanism, and removing information irrelevant to the quoted number ofa prediction paper; then, the cross comment matching mechanism capturing the consistency and diversity among different review opinions so as to describe the interaction among a plurality of reviewers; meanwhile, integrating width features through the width assembly; and finally, predicting the referenced number of the paper by using the combination of the depth component and the width component.According to the method, the semantic information in the review opinions is deeply described, semantic representation is enriched, and prediction of the reference number of the paper is more accurateby mining the text information of the review opinions.

The invention discloses a Bi-LSTM-CRF model-based content-related advertisement putting method and system, belongs to the technical field of advertisement putting, and is used for solving the problem that a deep learning-based named entity recognition model for a small-scale data set is difficult to obtain characteristics automatically, so that the model is difficult to obtain a good recognition effect, and the problem that advertisement recommendation cannot be accurately put is further solved. According to the technical key points, the method comprises the following steps: inputting a training data set into a Bi-LSTM-CRF model for training, and obtaining an optimal prediction model; inputting to-be-predicted data into the optimal prediction model to obtain predicted commodity words; matching related advertisements according to the commodity words, and obtaining advertisement information with the highest matching degree; and putting the advertisement carrying the advertisement information. According to the method, the features of the commodity words are combined on the basis of the Bi-LSTM-CRF algorithm, the data are enhanced in a feature engineering mode, the data are made to have richer semantics, a system suitable for document commodity word extraction is constructed and used for content-related advertisement recommendation, and the precise advertisement putting effect is improved.

The invention discloses a method for expressing RUCM by OML, which comprises the following steps: 1, determining elements to be converted and relationships according to relationships among RUCM modeling elements; 2, formulating a mapping rule between the RUCM and the OML in the step (1) according to the OML modeling elements, and forming an oml vocabulary of the RUCM; and 3, reading the text instance of the RUCM, calling the vocabulary in the step 2, and outputting the corresponding OML text instance. The method is used for meeting the requirement for knowledge sharing in the prior art. Automatic conversion from RUCM instances to OML instances is achieved, and the efficiency of a demand stage in industrial design is improved; manpower participation in the conversion process is omitted, andthe data processing efficiency is improved.