59 results about "Semantic learning" patented technology

The invention discloses an image retrieval ranking method based on deep learning. The method includes the following steps: extracting low-level features of images in a query object and a training database, conducting high-level semantic learning and feature fusing through a deep network to obtain high-level features of multiple query object-image pairs and initially-determined parameters of the deep network, respectively conducting linear regression on the high-level features of the query object-image pairs to obtain ranking grades of the query object-image pairs, obtaining a ranking list of images in a training data set related to the query object, comparing the ranking list with a real ranking list of the images in the training data set to obtain pairing loss values of the image pairs related to the query object, adjusting the initially-determined parameters of the deep network to obtain the final parameters of the deep network, calculating the low-level features of a new query object, obtaining a deep network corresponding to the new query object, and conducting searching in the training data set to obtain an image list related to the deep network.

The invention provides a recommendation method oriented to labels in software question and answer community based on deep learning for the first time. The method comprises the steps of step 1, preprocessing; step 2, word meaning expression learning; step 3, phrase and sentence semantic learning; step 4, semantic fusion. The phase before preprocessing is a training stage, in which a dictionary is constructed and the structure and the weight of a model are stored. The phase after the pre-processing is a deployment stage, in which a new problem is converted into an index sequence, and the storedmodel is loaded to predict and recommend the most likely k labels, with k being a positive integer. Problem resource management and maintaining an existing label system in a software question and answer community can be provided. An automatic labeling service is provided for the users. The method aims to overcome the defects of existing label recommendation technology, and accuracy and efficiencyof the method can be improved, so that the label recommendation can be more practical in a software question and answer community.

The invention relates to the technical field of video mining and analysis of social media and specifically provides a dynamic video recommendation device so that problems of depicting the dynamic interest of the user and carrying out video recommendation accurately can be solved. The dynamic video recommendation device comprises a video recommendation model that is capable of obtaining a recommended video at a current time according to video information of one or more historical times. To be specific, the video recommendation model consists of a semantic learning unit, a user behavior analysisunit, a recurrent neural network, a user relevance mining unit, and a model training unit. With the units, video semantic information, user interest information, and correlation degrees of differentuser interests are dug out fully and video recommendation is carried out based on the information, so that dynamic description of video watching preferences of the users is realized, the video recommendation accuracy is improved, and the accuracy of on-line dynamic video recommendation of the internet network is enhanced.

The invention discloses a method and device for analyzing correlation among objects in a social medial network, which are widely applied to many important related fields. The method comprises the following steps of: 1, extracting low-level features of the objects; 2, performing high-level semantic learning on the extracted low-level features through a deep network to obtain high-level features of the objects; and 3, obtaining the correlation among the objects according to the high-level features of the objects. In the method, high-level abstract features are learnt out of content information of the objects at first, and then correlation modeling is performed on the basis of the high-level features. A concept of implicit features is introduced in the method, and under a framework of a generative model, the implicit features generate the content information (high-level features) of the objects and the implicit features of the objects interact to obtain correlation information among the objects the maximum probability. The method adopts a Monte Carlo approximate inference algorithm to deduce parameters and hidden variables of a model.

The invention discloses a cross-modal retrieval method and device based on deep adversarial discrete hash learning, and the method comprises the steps: forming an image network through three full connection layers, forming a text network through four full connection layers, and obtaining a final hash value through the activation of a tanh function. Through adversarial training of the two networks,it can be ensured that feature expressions of the two modes tend to be consistent; semantic similarity is guaranteed through a cosine triple loss function to be weighted, meanwhile, relevancy ordering of similar samples is guaranteed, discreteness of hash values is maintained in the optimization process, quantization errors are reduced, and finally two robust hash functions for keeping semantic similarity, reducing heterogeneous gaps and being small in accumulative error are obtained. According to the method, the hash function is learned by utilizing adversarial training and keeping semanticsimilarity and relevancy ranking, so that the retrieval precision is improved, and the obtained hash learning method has stronger semantic learning capability.

The invention discloses an automatic speech recognition method and an automatic speech recognition system based on artificial intelligence. The system mainly comprises four modules of a speech preprocessing module, a speech feature extraction module, a speech training recognition module and a text correction module. According to the system, the speech training recognition module is adopted to learn the speech features and speech corresponding character codes, firstly, convolution learning of frequency spectrum characteristics is carried out through a feature learning layer, then the semantic information among the frequency spectrum features is learned through a semantic learning layer, finally, the comprehensively learned information is decoded through an output layer, and a correspondingtext is output. Therefore, a Chinese character mapping table is directly used for encoding and decoding labels, phoneme encoding and decoding of the text are not needed, then the text is decoded intothe text, and the training process is simplified.

The invention discloses a rumor detection method combining a self-attention mechanism and a generative adversarial network. The rumor detection method comprises the steps of collecting rumor text datato form a rumor data set; based on a self-attention mechanism, constructing a generative adversarial network generator comprising a self-attention layer; constructing a discriminator network, and respectively carrying out rumor detection and classification on the original rumor text and the text decoded by the generator; training the generative adversarial network, and adjusting model parametersof a generator and model parameters of a discriminator; and extracting a discriminator network of the generative adversarial network, and performing rumor detection on the to-be-detected text. Compared with an existing rumor detection method, the rumor detection method is higher in detection precision and better in robustness; a self-attention layer is adopted in the generator, key features are constructed through semantic learning of rumor samples, text examples rich in expression features are generated to simulate information loss and confusion in the rumor propagation process, and the semantic feature recognition capacity of the discriminator is enhanced through adversarial training.

The invention discloses a method for constructing a universal embedding framework of a multi-semantic heterogeneous graph, which comprises the following steps of: 1, constructing a neighborhood exploration strategy alpha-exploration, and smoothly splicing two exploration strategies, namely, DFS and BFS, so as to adapt to different heterogeneous network structures; 2, based on alpha-exploration, constructing an HNSE model, wherein the HNSE model comprises an alpha-exploration neighborhood exploration layer, a multi-semantic learning layer and a node classification layer; and learning low-dimensional embedding of the nodes while heterogeneous information and semantic information of the nodes are reserved; 3, realizing a multi-layer HNSE model in a residual form, and connecting a full-connection output layer behind the multi-layer HNSE model; and 4, constructing three expansion strategies of the HNSE. According to the method, each vertex of the multi-semantic heterogeneous graph is embedded by aggregating adjacent/meta-path neighbor nodes of different types, and a node aggregation sampling strategy combining meta-path neighbors and direct neighbors is designed for the HNSE, so that a multi-head attention mechanism in the HNSE is guided, and capture of node multi-semantic information is improved by utilizing meta-paths.

The invention relates to a cross-domain emotion classification system based on attention mechanism fusion. The system comprises a comment text preprocessing module used for obtaining vector forms of texts in a source domain and a target domain; a text semantic learning module which is used for learning a semantic dependency relationship between words; an attention mechanism fusion module which isused for fusing different attention modes to obtain comprehensive weights of words for text classification; a hierarchical attention module which is used for calculating attention weights of the textfrom a word level and a sentence level respectively and judging weights of words for sentence representation and sentences for document representation; and an emotion category output module which is used for obtaining a final emotion classification result by utilizing the classification function. According to the method, the potential universal features of the target domain and the source domain can be automatically extracted, the features are abstracted and combined, and finally the emotion category of the text of the target domain is recognized.

The invention discloses a text extraction method and system based on fusion pre-training and a medium. The method comprises the steps of obtaining a to-be-extracted text; performing pre-training coding on the to-be-extracted text through a pre-training model to obtain a corresponding character vector; selecting at least part of the character vectors to perform semantic extraction on adjacent texts, and splicing to obtain semantic feature vectors; performing feature selection and fusion on the semantic feature vectors to obtain effective word feature vectors; and carrying out shunt decoding on the effective word feature vectors to respectively obtain a word segmentation result and an entity recognition result. Coding is performed based on the pre-training model framework to obtain the character vector, and at least part of the character vector is fused to perform semantic extraction of the adjacent text to learn semantic information of the text, so that the semantic learning ability is enhanced, the finally obtained word segmentation result can effectively avoid the problem of fuzzy boundary, and the accuracy of text extraction is improved.

The invention provides a latent semantic learning method for multi-view multi-label data, which comprises the following steps of: reading multi-view multi-label data as the input of a preprocessor, then carrying out data preprocessing, inputting a preprocessing result into a trainer, and learning to obtain a latent semantic subspace and a trained model; when the label which is not labeled with the multi-view data is predicted, preprocessing the unlabeled multi-view data, and then inputting the unlabeled multi-view data into the trained model for prediction to obtain a multi-label category. The invention provides a learning and multi-label classification method fusing multi-view angle information. Input features and output multi-labels are mapped into a common potential semantic subspace. According to the method, the problem of dimensionality disasters caused by high-dimensional sparse data can be solved, and meanwhile, multi-view information is fused to classify multi-label data.

The invention provides a Transformer-based code programming language classification method, which comprises the following steps: (1) collecting question and answer posts from Stack Overflow as a data set, and carrying out data preprocessing on data in the original data set; (2) carrying out word embedding on the data subjected to word segmentation by using the BPE to convert words into vectors; (3) on the basis of the constructed data set, performing fine tuning on the RoBERTa model, inputting a generated word vector into the RoBERTa model, performing code semantic learning through a double-layer Transform encoder, and generating a semantic representation vector Xsematic; and (4) mapping the semantic vector Xsemination to a programming language category label through a linear layer, and obtaining a corresponding programming language through a Softmax algorithm. The method has the beneficial effects that the code type can be quickly identified according to the code snippets so as to play a role in assisting developers to quickly find the solution on the question and answer website.

The invention discloses a cross-version deep defect prediction method capable of relieving a class overlapping problem. The method comprises 1, a deep semantic learning-oriented overall framework in cross-version software defect prediction; 2, a semantic feature learning model based on a convolutional neural network; 3, a hybrid nearest neighbor cleaning strategy for deep semantic learning. According to the method, a hybrid nearest neighbor cleaning strategy is adopted to relieve the class overlapping problem existing in semantic features learned by deep learning. Specifically, for an abstractsyntax tree corresponding to a source code, a convolutional neural network is adopted to learn deep semantic features, and then a hybrid nearest neighbor cleaning strategy is adopted to perform resampling and data cleaning on a labeled data set. By adopting the hybrid nearest neighbor cleaning strategy, the class imbalance problem and the class overlapping problem can be processed, and the statistical analysis result of the data shows that the strategy can improve the performance of software defect prediction based on deep semantic learning.

The invention discloses a many-to-many speaker conversion method based on an improved STARGAN and an x vector. The method comprises a training stage and a transforming stage; a speech conversion system is achieved via the combination of the improved STARGAN and the x vector; the method is the further improvement of STARGAN in a speech conversion application; a two-step antagonistic loss is provided and can effectively solve the problem of excessive smoothness caused by cyclic consistence loss expressed by L1; furthermore, a generator adopts a 2-1-2D CNN network, a semantic learning capacity and a speech spectrum synthesis capacity of the model can be well enhanced, and the problem of relatively poor speech similarity and naturalness after conversion in the STARGAN is overcome. Meanwhile, the x vector has better representation performance for short speech, the personality feature of the speaker can be fully represented, and a high quality many-to-many speech conversion method under thecondition of non-parallel text is achieved.

The invention provides a Chinese query expansion method based on pattern mining and word vector similarity calculation, which. The method comprises the following steps: firstly, retrieving a Chinese document set through user query to obtain an initial retrieval document, and performing word vector semantic learning training on the initial retrieval document set to obtain a word vector set comprising query word items and non-query word items; then mining extension words for the pseudo-correlation feedback document set by adopting a Copulas-function-based associated extension word mining method,and establishing an associated extension word set; and performing cosine similarity operation of two vectors in the word vector set to obtain a word embedding extension word set and a word vector association extension word set, finally fusing the word embedding extension word set and the word vector association extension word set to obtain a final extension word, combining the final extension words with the original query to form a new query, and retrieving the document set again to realize query extension. According to the method, association mode mining and word vector learning are fused, high-quality extension words can be mined, the information retrieval performance is improved, and the method has good application value and popularization prospects.

Embodiments of the specification disclose a risk address identification method and apparatus, and an electronic device. The risk address identification method includes: acquiring an address word sequence corresponding to an input address; determining an address word in the address word sequence, the determined address word matching a risk word corresponding to a risk address; generating an observation sequence corresponding to the address word sequence according to the determined address word; processing the observation sequence using a hidden Markov model obtained based on semantics learning before and after address words, to obtain a decision vector, wherein the decision vector represents probabilities of the risk address being matched by address words contained in the address word sequence; and identifying whether the input address is a risk address by making a classification decision on the decision vector.

The invention discloses a knowledge graph completion method based on multi-semantic learning. The method comprises the steps of learning entities e1 and r to a plurality of hidden semantic representations through a plurality of conversion matrixes; in the step of embedding and capturing the plurality of hidden semantics of the entity and the relationship in the previous knowledge graph, obtaining a plurality of feature embedding for the same entity or relationship; optimizing embedding of the entity and the relationship by utilizing a deep residual attention network; introducing denoising network optimization entity embedding and relationship embedding; then, simply describing the structure of the denoising network; and introducing a multi-step fusion process to fully fuse entities and relationships; according to the knowledge graph completion method provided by the invention, the problem of a large amount of noise caused by introduction of a plurality of hidden semantics can be effectively reduced. And meanwhile, the denoising network and the multi-step fusion network can fully fuse entities and relationships to obtain the most consistent prediction result.

The embodiment of the invention provides an image processing method and device based on artificial intelligence and a storage medium, and the method comprises the steps: a training sample set, which comprises a first image sample set with a class label and a second image sample set composed of triples, is acquired; semantic learning training and metric learning training are conducted on an original model through the training sample set, the original model comprises a first branch network and a second branch network, and the first branch network and the second branch network comprise shared network parameters; an image feature extraction model is determined according to the trained original model, and the image feature extraction model is used for extracting the feature vectors of the images, so that the model has the semantic extraction capability while realizing metric learning, and the accuracy of image retrieval can be improved based on the image features extracted by the model.

The invention discloses a multi-scale feature fused multi-supervision face in-vivo detection method, which comprises the following steps of: acquiring an image data set, and preprocessing the data set; gradient texture features are extracted through central difference convolution, and codes are fused; extracting multi-scale discriminant features through a group receptive field branch, and splicing and fusing the multi-scale discriminant features with a gradient texture branch; fusing the two features, inputting the fused features into a residual structure, carrying out deep semantic learning and coding, and inputting a result into a depth map generator and a mask generator to obtain a feature map; a depth map is used for supervision, and a binary mask is used for auxiliary supervision; and fusing output results of the depth map generator and the mask generator, calculating a prediction score, and realizing end-to-end living body detection. According to the invention, the performance and generalization ability of the network can be improved, and the method has the advantages of small parameter quantity and end-to-end detection; compared with an existing mainstream living body detection algorithm, the method is higher in detection precision and better in robustness.

The invention relates to an implicit discourse relationship classification method. The method comprises the following steps: for first discourse information and second discourse information to be classified, based on an implicit discourse relationship classification model, obtaining a semantic interaction graph structure of semantic information in the first discourse information and the second discourse information; for the semantic interaction graph structure, obtaining semantic features corresponding to the first discourse information and the second discourse information; projecting the obtained semantic features into a hyperbolic space by means of index mapping, and obtaining a classification result of the first discourse information and the second discourse information, wherein the classification model based on the implicit discourse relationship is a model which is pre-established and trained and comprises a context presentation layer, a semantic learning layer, a convolution layer, an aggregation layer and a prediction layer; and enabling the context presentation layer and the semantic learning layer to execute a process of obtaining a semantic interaction graph structure, enabling the semantic learning layer, the convolution layer and the aggregation layer to execute a process of obtaining semantic features, and enabling the prediction layer to execute a process of obtaining a classification result.

The invention provides a training method of a natural language processing model, and relates to the field of artificial intelligence, in particular to the field of natural language processing. According to the specific implementation scheme, the following steps are included: carrying out semantic learning aiming at multiple tasks on an input text, obtaining semantic features aiming at the multiple tasks, wherein the multiple tasks comprise multiple branch tasks; performing feature learning for each branch task based on the semantic features to obtain a first output result for each branch task; calculating loss for each branch task according to the first output result for each branch task; and adjusting parameters of the natural language processing model according to the loss of each branch task. The invention further provides a natural language processing model training device, a natural language processing method and device, electronic equipment and a storage medium.

The invention discloses a body movement and language factor matching method and device for a virtual image, and the method comprises the steps: virtual image limb movement generation: presetting a user-defined movement which comprises the position of the virtual image in a map and a body movement path, and enabling the user-defined movement to generate corresponding movement two-dimensional data; virtual image and language matching interaction: performing semantic learning on the action two-dimensional data, performing virtual image and language matching interaction, and generating motion control information; and virtual image skeleton driving: transmitting the motion control information to a bottom layer driver of the virtual image, wherein the bottom layer driver controls a skeleton driving action of the virtual image according to the motion control information. According to the method and device, the semantic and action matching of the virtual image is realized, so that the emotional expression, expression interaction, limb actions and the like in the communication process are close to a real person to the maximum extent, and the language and the action are synchronous and consistent in expression.