31 results about "Relational encoding" patented technology

The invention relates to a dimensionality reduction expression method of a mapping knowledge domain on the basis of sub-graph division. The dimensionality reduction expression method comprises the following steps: carrying out sub-graph division on the mapping knowledge domain; carrying out CP tensor decomposition on divided sub-graphs to obtain an entity coding vector and a relationship coding vector, and outputting as a dimensionality reduction expression result.

The invention provides a product modeling method based on a behavior flow spatial sequence coloring Petri network. The method at least comprises the following steps of: I, encoding parts related to a product and connecting relations among the parts; II, establishing a semantic network of a behavior flow spatial sequence in the product on the basis of the encoding; III, establishing a behavior flow spatial connecting relation predicate matrix on the basis of the establishment of the semantic network for expressing the connecting relations among product structural units, wherein rows and columns are structural unit codes related to the product parts, and data in the matrix are connecting relation codes; and IV, converting the behavior flow spatial connecting relation predicate matrix into a coloring Petri network model of a spatial sequence. A more intelligent behavior flow product modeling method is adopted, and display expression of flow is performed through the establishment of the semantic network of the behavior flow spatial sequence and Petri network modeling, so that a transmitting sequence of a spatial structure state in a behavior flow modeling process is obtained.

The invention relates to a named entity identification method based on syntax dependency. In named entity identification, only when entity boundaries and types identified by a model are matched with boundaries and types of labeled entities, the entity is calculated to be a true positive example (TP). In most test samples, false positive examples (FP) and false negative examples (FN) are often caused by incorrect boundary judgment of entities, that is, boundary identification is much more difficult than type identification. According to the method, a self-attention mechanism is used for weakening the relation between entities and words outside the entities, and the relation between the words inside the entities is enhanced. Specifically, a self-attention mechanism is added after a bi-directional long-short term memory (Bi-LSTM) network, dependency relationships among words in a syntactic dependency tree are encoded into context information, and finally entity boundaries are judged jointly according to global features provided by the Bi-LSTM network and local features provided by the syntactic dependency tree. According to the method, the accuracy of named entity identification is improved.

Trajectory prediction may receiving a LiDAR image sequence including a set of LiDAR images and generating a LiDAR map, generating an interaction encoder result by feeding the LiDAR image sequence through an interaction encoder, generating a feature extractor result by feeding the LiDAR map through a feature extractor, generating a relation encoder result by feeding a past trajectory of a detected obstacle from the LiDAR image sequence, the interaction encoder result, and the feature extractor result through a relation encoder, generating an intention estimation result by feeding the relation encoder result through an intention estimator, generating a conditional generative model result by feeding the past trajectory of the detected obstacle, the intention estimation result, and a probability map through a conditional generative model encoder, and generating a trajectory prediction by feeding the relation encoder result, the past trajectory of the detected obstacle, and the conditional generative model result through a trajectory predictor.

The invention relates to the field of computer vision, in particular to a video event description method based on multi-mode heterogeneous feature fusion. According to the method, an I3D convolutional network is used for editing a video and extracting dynamic visual features, and a VGGish model is used for extracting audio rhythm features; performing semantic representation on the scene object information, generating a scene map, obtaining an entity code, an attribute code and a relation code, and performing map embedding on the feature vector through map convolution; performing triple multi-mode cycle fusion on the three extracted features; adaptive multi-modal data balance enables dynamic vision and audio rhythm features to be mutually matched, and ensures reasonable event extraction; and the description reconstruction decoder is used for detecting the video event by utilizing a description reconstruction algorithm and generating description of the video scene event according to the pre-training language dictionary. The problem that meaningless descriptions such as co-occurrence descriptions are frequently generated in a video description method is effectively solved, and the relation of scene events is effectively mined by effectively utilizing multi-modal information.

The invention discloses an image description generation method based on an external triple and an abstract relation. The method comprises the following steps: firstly, extracting a triple in an image description text, constructing an external relation library, and carrying out feature coding on the triple; and clustering the triples of which the text similarity is higher than a threshold value into one class. Meanwhile, the model performs target detection on the image to obtain a target visual feature set and a target category set; and querying a triple of which the target is similar to the target category in an external relation library according to the text similarity. The model predicts the target, the attribute and the relation of the image by using the visual features of the target to generate a scene graph; and fusing visual features and text features by using a convolutional neural network, and carrying out feature coding on targets, attributes and relationships. And finally, fusing target, attribute and relation coding features of the scene graph and coding features of the similarity relation and the abstract relation, and inputting the fused features into a double-layer LSTM sequence generation model to obtain final image description. According to the invention, the expression of the model generation description is richer.

The invention relates to a bidirectional coding target encoder construction method and device. The method comprises the steps of obtaining a training text; encoding the training text into a first sequence according to the encoding relationship stored in the dictionary, wherein the encoding sequence of the first sequence is consistent with the text sequence of the training text; covering each element in the first sequence in sequence according to a character sequence of the training text to obtain a plurality of second sequences; rearranging the elements in each second sequence according to a sequence opposite to the current arrangement sequence to obtain a plurality of third sequences; and inputting the second sequence and the third sequence into a self-encoding language model, and outputting the model as a target encoder. When the target encoder is constructed, feature extraction training is performed by adopting forward encoding input and reverse encoding input, so that the feature representation capability of the encoder is improved through forward and reverse bidirectional encoding in a real spatial sense, and the technical problem that the translation accuracy at the end of asentence is continuously reduced is solved.

Trajectory prediction may receiving a LiDAR image sequence including a set of LiDAR images and generating a LiDAR map, generating an interaction encoder result by feeding the LiDAR image sequence through an interaction encoder, generating a feature extractor result by feeding the LiDAR map through a feature extractor, generating a relation encoder result by feeding a past trajectory of a detected obstacle from the LiDAR image sequence, the interaction encoder result, and the feature extractor result through a relation encoder, generating an intention estimation result by feeding the relation encoder result through an intention estimator, generating a conditional generative model result by feeding the past trajectory of the detected obstacle, the intention estimation result, and a probability map through a conditional generative model encoder, and generating a trajectory prediction by feeding the relation encoder result, the past trajectory of the detected obstacle, and the conditional generative model result through a trajectory predictor.

The invention discloses an implicit discourse relation identification method based on a TransS-driven mutual excitation neural network. The method comprises the following steps: (1) constructing an embedding layer of argument and discourse relation; (2) carrying out expression learning of chapter arguments; (3) constructing attention mechanism enhanced representation learning; (4) a sentence translation embedding module (Translating Sentence Embedding, TransS); (5) constructing a chapter relationship identification module; and (6) constructing a mutual excitation mechanism. According to the method, firstly, an argument pair-relation embedding layer is used for obtaining an embedding vector of an argument pair and a relation, then distributed representation of the argument pair and the relation is modeled through an argument pair-relation encoding layer and introduction of an attention mechanism, and finally representation parameters are optimized and the relation recognition performance is improved through mutual guidance between TransS and a relation recognition module.

The invention discloses a social media popularity prediction method and device based on a visual semantic relationship, wherein the method comprises the steps of extracting paired objects and predicate relationships between the paired objects from an image of a post through a pre-trained scene graph generator, and generating a < subject-predicate-object > relationship; encoding the relationship into semantic features by using a word vector model; encoding other data of the post to obtain text features, numerical features and additional user features, and connecting the text features, the numerical features and the additional user features with semantic features; aiming at the missing problem of users of partial posts of the test set in the training set, training two Catboost models respectively by using connected multi-modal features, and performing linear combination output to obtain a preliminary popularity score; and carrying out fine adjustment on the preliminary popularity score aiming at the post content by utilizing training set data, and balancing a model prediction error brought by user information, so as to obtain a final popularity score. The device comprises a processor and a memory. According to the invention, the accuracy of popularity prediction is improved.

The invention discloses a model training method and device and a method and device for realizing text processing. The method comprises the following steps: expressing a sample text as a sample dependency relationship code according to dependency relationship information; taking the feature vectors of the words contained in the sample text as nodes, and expressing the sample text as a sample matrix; training the sample matrix and the adjacent matrix thereof by initializing a graph neural network model of the operating parameters to obtain a dependency word vector and a subordinate word vector in the sample dependency relationship code; taking the positive sample and the configured negative sample as the input of the graph neural network model, carrying out the training through employing a back propagation method, and determining an operation parameter matrix of the graph neural network model; and processing the input to-be-processed text according to the graph neural network model for determining the operation parameter matrix. According to the embodiment of the invention, the dependency relationship information is added into the model for processing the text, so that the feature representation efficiency of the text content is improved.

A control device (CD) is dedicated to the control of the transmission of coded values onto a digital interface (I) connecting a baseband device (BBD) and a baseband interface device (BAI), comprising at least a modulator (M) feeding a gain controller (GC), of a wireless communication equipiment. The control device (CD) comprises a storing means (MM1) for storing a coding table establishing a correspondence between symbols for the baseband interface device (BAI) and coding values to transmit to this radiofrequency device through the digital interface (I). Tie coding table comprises a first group of symbols comprising data words for feeding the modulator (M) and a second group of symbols comprising command words for controlling the operation of the modulator (M) and / or the gain controller (GC). The control device (CD) also comprises a control means (CRM) arranged, when it receives a symbol from the baseband device (BBD), to determine in the storing means (MM 1) the coded value corresponding to this symbol in order it could be transmitted to the baseband interface device (BAI) through the digital interface (I).

A knowledge graph path reachability analysis method based on an attention mechanism includes the following steps: 1, constructing a target triple from a knowledge base, and obtaining all path relations between a head entity h and a tail entity t in the triple; 2, carrying out relation coding; 3, performing entity type coding; 4, repeating the step 2 and the step 3 to calculate a global path mode formed by combining all path modes, calculating an energy function of a triad formed by a head entity h, a direct relation r and a tail entity t, calculating the probability whether the direct relation r can connect the head entity and the tail entity, multiplying the energy function by the probability whether the energy function can be linked, and judging whether the triple is true or not. According to the method, the utilization rate of entities and relationships is improved, the accuracy of a probability calculation result is improved through an attention mechanism, the accuracy of vectors represented by triples is improved, and the accuracy of a result for predicting whether the entities can be mutually connected or not is improved.

The invention provides an attribute fusion interactive recommendation model construction method and system based on enhanced graph convolution, and the method comprises the steps: constructing a relation encoder based on enhanced graph convolution, extracting high-order features between user items from user item interaction behaviors, and obtaining dynamic feature representations, containing high-order cooperative signals, of users and items at all moments; and constructing a comment encoder fusing attribute information, and performing corresponding processing and fusion on the comment text and the corresponding attribute to obtain accurate feature representation of the comment. The analysis and research show that the attribute activation method provided by the PGIR can well solve the problem that the negative user recommendation is inaccurate, so that the model performance is improved, and the robustness and interpretability of the recommendation are improved at the same time.

The invention discloses a video space-time action positioning method based on a progressive attention hypergraph. The method comprises the following steps: firstly, sampling a given original video to obtain a frame sequence, and obtaining a target context feature and a video spatio-temporal feature map by using a convolutional neural network; target context features and a time-space relation matrix are obtained through a time-space relation encoder; generating a long-term target first-order feature by using a progressive variable-length window method module; meanwhile, target short-term high-order features are obtained through a hypergraph module with shared attribute constraints and a diffusion mechanism; and finally, outputting spatial positions and action categories of all targets at different moments by using a target action regression module. According to the method, the window size can be adaptively adjusted according to the original duration of the action, the target first-order feature consistent with the original duration of the action can be obtained, the potential relation of the target can be captured through the hypergraph module, the target interaction relation can be effectively utilized, and the accuracy of video space-time action positioning is improved.

The invention discloses a data follow-up relationship analysis method, which is applied to an analysis server in a data analysis network. The data analysis network also includes a plurality of data acquisition devices, and all data acquisition devices in the data analysis network are bound in pairs. And generate a unique binding relationship code, obtain the binding relationship code corresponding to the data to be analyzed according to the data type and space following conditions of the data to be analyzed, group the data to be analyzed according to the binding relationship code corresponding to the data to be analyzed, and judge Whether the data to be analyzed in the group corresponding to each binding relationship code has time following, and output the judgment result of the data to be analyzed, which reduces the amount of calculation when analyzing the data following relationship, and improves the efficiency of data analysis.

The invention discloses an image paragraph description method based on relation coding and a hierarchical attention mechanism. A method model is composed of a relation coding module and a hierarchical attention decoding module. The relation encoding module captures and encodes space relation information and semantic relation information through two encoders, wherein the prior knowledge of the semantic relation is learned by training a supervised semantic classifier when the semantic relation is encoded. The hierarchical attention of the hierarchical attention decoding module uses hierarchical attention with a relational gate and a visual gate to dynamically fuse relation information and object region features, the relational gate is used for switching between spatial relation information and semantic relation information, and the visual gate is used for determining whether to embed and use visual information; the model fuses visual information in a paragraph generation process by adopting a strategy of a spatial and semantic relationship from a coarse-grained region to a fine-grained region. A large number of experiments on a Steiner paragraph description data set show that the method is obviously superior to an existing method in multiple evaluation indexes in the field.

The invention relates to a layout analysis method and device, computer equipment and a storage medium. The method comprises the following steps: acquiring a spatial conversion feature, a position feature and an image feature of each text line in a target text image; according to the space conversion feature and the position feature of each text line, obtaining a relation coding feature of each text line, and splicing the relation coding feature of each text line and the image feature of each text line to obtain a fusion feature of each text line; and decoding the fusion feature of each text line to obtain the layout information type of each text line. The content of the text image is understood by combining the spatial conversion features and the image features instead of being based on one feature, so that the analysis result is relatively accurate. Besides, each text line is analyzed in combination with the image feature and the spatial conversion feature during analysis, and the other feature can be used as an analysis basis under the condition that one feature is wrongly recognized, so that the error-tolerant rate of layout analysis is improved.

The invention discloses a small sample learning method and device based on sample pair relationship propagation. According to the method, explicit modeling and propagation are performed on the relationship between the sample pairs of the support set query set, so that a relationship code with better discrimination can be obtained. By introducing the pseudo-relationship nodes, the feature information of the query set sample can be effectively reserved. Moreover, the invention further provides an effective transduction learning strategy, and the relation information between the query set samplescan be better mined, so that a more accurate classification result is obtained. Compared with the prior art, the method has the advantages that potential information contained in the sample pair of the support set query set in each task is better mined, and higher accuracy and better generalization ability are achieved when a brand-new task is processed.

The invention discloses a face image clustering method and device for link prediction based on a self-attention mechanism, and the method comprises the steps: 1, selecting samples, supposing that the total number of the samples is N, and carrying out the feature extraction of the selected samples through a face recognition model; the method comprises the steps of 1, inputting a candidate enhancement feature set of an ith sample into a feature enhancement coding module based on context information for enhancement, 2, inputting the candidate enhancement feature set of the ith sample into a relation coding module with self-attention to obtain all possible link sets of the ith sample, and 4, combining the link sets of all the samples through a union-check set algorithm to obtain a candidate enhancement feature set of the ith sample. And a final clustering result is obtained. According to the method, a clustering task can be converted into a classification task through link prediction, and the accuracy of a clustering result can be improved; the effect of enhancing original node features is achieved by extracting and combining context information of part of neighbor nodes, and the negative influence of samples with low distinction degree is reduced.

The invention relates to a program source code coding method based on a code attribute tensor. The program source code coding method comprises the steps of generating a code attribute graph for a program source code; creating a symbol table, a node table and a code attribute tensor, and initializing; encoding the operation relationship between the data type of the AST node and the AST node into a code attribute tensor; encoding the father-child relationship between the AST nodes into a code attribute tensor; and encoding the adjacency relationship between the CFG nodes into the code attribute tensor, and outputting a final code attribute tensor. The invention provides a code attribute tensor concept and an algorithm for encoding a program source code into a code attribute tensor in order to solve the problem that an existing program source code encoding mode cannot fully embody program semantic characteristics. According to the method, under the condition that semantic information is prevented from being lost, program source codes are encoded into tensor-form data to serve as input of a machine learning model, and support is provided for subsequent program static analysis work.

The invention provides a human-computer interaction method, device and equipment, and a readable storage medium. The method is applied to electronic equipment, and the electronic equipment comprises a measuring device and electronic skin. The method comprises the following steps: reading an electric information value on the electronic skin through a measuring device; matching a coding command corresponding to the electric information numerical value from a pre-obtained coding table, and storing a mapping relation between a numerical value interval corresponding to the electric information numerical value and the coding command in the coding table, wherein the coding table corresponds to controlled equipment connected with the electronic equipment; and sending the coding command to the controlled equipment to enable the controlled equipment to execute the coding command. In the implementation process, the corresponding coding command is determined through the electric information numerical value on the electronic skin and the preset coding table and sent to the controlled device, and the controlled device achieves the corresponding function based on the received different coding commands. Since different coding tables represent different controlled devices, the compatibility of the electronic device can be improved by storing a plurality of different coding tables.

The invention provides a product modeling method based on a behavior flow spatial sequence coloring Petri network. The method at least comprises the following steps of: I, encoding parts related to a product and connecting relations among the parts; II, establishing a semantic network of a behavior flow spatial sequence in the product on the basis of the encoding; III, establishing a behavior flow spatial connecting relation predicate matrix on the basis of the establishment of the semantic network for expressing the connecting relations among product structural units, wherein rows and columns are structural unit codes related to the product parts, and data in the matrix are connecting relation codes; and IV, converting the behavior flow spatial connecting relation predicate matrix into a coloring Petri network model of a spatial sequence. A more intelligent behavior flow product modeling method is adopted, and display expression of flow is performed through the establishment of the semantic network of the behavior flow spatial sequence and Petri network modeling, so that a transmitting sequence of a spatial structure state in a behavior flow modeling process is obtained.

The invention discloses an attention mechanism relationship comparison network model method, and constructs an attention relationship comparison network model for small-sample learning under a small amount of labeled sample data. Based on the relational network architecture, the model is divided into feature encoding, feature combination, and relational encoding. The function of the feature encoding module is to extract image feature information. The feature combination part is to combine the extracted query image feature information with the training image features of each group. The information is recombined to form a new combined feature map. The relationship encoding module performs nonlinear metric learning of the network. By introducing the attention mechanism and spectral normalization method into the end-to-end deep convolutional neural network model, the model has higher classification accuracy in the case of small sample learning. , and improve the stability of the final training results of the model, and improve the image classification accuracy of the existing model in small-sample learning.

The invention discloses an image paragraph description method based on relational coding and hierarchical attention mechanism. The method model is composed of a relational coding module and a hierarchical attention decoding module. The relational encoding module captures the encoded spatial relational information and semantic relational information through two encoders, where the prior knowledge of the semantical relation is learned by training a supervised semantic classifier during semantic relational encoding. The hierarchical attention of the hierarchical attention decoding module uses hierarchical attention with relational gates and visual gates to dynamically fuse relational information and object region features. The relational gates are used to switch between spatial relational information and semantic relational information. To decide whether to use visual information for embedding, the model adopts a strategy from coarse-grained regions to fine-grained spatial and semantic relations to fuse visual information during paragraph generation. Extensive experiments on the Stanford paragraph description dataset show that the method of the present invention is significantly better than the existing methods in multiple evaluation indicators in the field.

The invention relates to a knowledge graph representation learning method based on entity and relation coding in a neural network; the method comprises the following steps: step 1, constructing a target triple from a knowledge base, and obtaining all path relations between a head entity and a tail entity in the triple; step 2, carrying out relation coding; step 3, performing entity type coding; step 4, obtaining type context vectors of the head entity and the tail entity in the step 3, and inputting the type context vectors into the LSTM in sequence; step 5, forming path modes vrho (p) and vrho (r), and calculating the cosine similarity of the two path modes; and step 6, training a representation learning method. According to the method, the semantic information of the entities and the relationships is expressed, so that the entities, the relationships and the complex semantic association between the entities and the relationships are efficiently calculated.

The invention discloses a case-based reasoning intelligent photographing system and method. The system comprises an image obtaining module, an RAM module, an analysis processing module and a case basemodule. The image obtaining module obtains a target preview and caches the target preview in the RAM module. The analysis processing module calls the target preview and carries out target main characteristic coding and target mapping relationship coding. The case base module calculates similarity between the target main characteristic coding and source case main characteristic coding, calculatesthe similarity between the target mapping relationship coding and source case mapping relationship coding and selects a photographing angle with the highest similarity for intelligent photographing composition. According to the system and the method, related characteristic information of the target image is obtained intelligently through utilization of an algorithm in the analysis processing module; on the base of case reasoning, through utilization of the algorithm in the case base module, the optimum photographing angle is intelligently recommended to a user; and the defect that in the priorart, a mobile phone photographing system cannot guide the user to adjust the photographing angle is solved.