Knowledge representation method for cross-media knowledge reasoning task

Abstract

The invention provides a knowledge representation method for a cross-media knowledge reasoning task, and the method comprises the steps: extracting the RDF triad information of a cross-media knowledgegraph, and enabling the RDF triad data of the cross-media knowledge graph to be represented as an initial low-dimensional vector; training vector representation between positive and negative triad samples by utilizing a maximum interval cost function, mining similarity (or difference) between the positive and negative triad samples, and adding the similarity (or the difference) into the maximum interval cost function, so that the capability of recognizing similar entities by model knowledge reasoning is improved. According to the method, knowledge representation and knowledge reasoning can becarried out on the cross-media knowledge graph triples constructed on the basis of the RDF, entity linking and knowledge classification are carried out by utilizing the knowledge reasoning model learned by the method, and the accuracy of link prediction and triad classification in the cross-media knowledge graph can be improved.

Description

technical field [0001] The invention relates to a knowledge representation method for cross-media knowledge reasoning tasks, which belongs to the technical fields of natural language processing, artificial intelligence, etc., and specifically relates to a knowledge reasoning method for RDF triples in cross-media knowledge graphs, including knowledge graphs link prediction and classification. Background technique [0002] With the development of the big data era, data from all walks of life has shown explosive growth. Knowledge Graph provides a powerful engine for the efficient use of these massive data resources. Now the mainstream knowledge graph has developed into the core supporting many artificial intelligence (AI) applications. Common AI applications include intelligent search, automatic question answering, recommendation systems, and decision support. Although in the past ten years, the knowledge map has made great progress, but there are still some limitations. Among...

AI Technical Summary

Problems solved by technology

[0003] In recent years, a series of knowledge representation methods have been proposed using deep learning methods, but these knowledge representation methods have the problem of low reasoning accuracy in knowledge reasoning. In addition, the current mainstream knowledge representation-based knowledge graph reasoning at home and abroad The work mainly includes TransE (Bordes A, Usunier N, Garcíadurán A, etal.Translating Embeddings for Modeling Multi-relational Data[C].International Conference on Neural Information Processing Systems.2013: 2787-2795, Translating Embedded Models Based on Translation), TransH (Wang Z, Zhang J, Feng J, et al. Knowledge Graph Embedding by Translating on Hyperplanes[C]. Twenty-eighth AAAI Conference on Artificial Intelligence.AAAI Press, 2014:1112-1119, Embedded Model Based on Hyperplanes), TransR (Lin Y, Liu Z, Sun M, et al.Learning entity and relation embeddings for knowledge graphcompletion[C].Twenty-ninth AAAI Conference on Artificial Intelligence. 2015, Embedded Model Based on Entity and Relation Space), CTransR(Lin Y, Liu Z, Sun M, et al. Learning entity and relation embeddings for knowledge graph completion [C]. Twenty-ninth AAAI Conference on Artificial Intelligence. 2015, Embedded Models Based on Clustering and Entity Relation Space) and TransD (Ji G ,He S, Xu L,et al.Knowledge Graph Embedding via DynamicMapping Matrix [C].Meeting of the Association for Computational Linguistics&the Inter national Joint Conference on Natural Language Processing.2015, an embedded model based on dynamic mapping matrix), that is, the knowledge map contains a large number of fact triples, and entities (including concepts and attribute values) are represented as nodes in the knowledge map. The connection between nodes represents the relationship, which is stored in the form of (head entity, relationship, tail entity) (expressed as (h, r, t)), and the acquired knowledge is displayed in a network structure. For each triple (h, r, t), the translation model regards the relation r as a translation operation from the head entity h to the tail entity t, and the TransE model represents the entities and relations in the knowledge map as low-dimensional embedded vectors, and each The relationship is regarded as a transfer in the embedded space. For the triplet (h, r, t) established in the knowledge graph, the sum of the low-dimensional embedded head entity vector h and the relationship vector r based on the representation is close to the tail entity vector The value of t, that is, h+r≈t, otherwise far away, the scoring function used by the model is Use the second-order norm of the vector to calculate the distance; the TransE model is suitable for dealing with one-to-one relationships, and cannot handle one-to-many, many-to-one and many-to-many relationships well; the TransH model overcomes one-to-many and many-to-one And the disadvantages of many-to-many relationships, the TransH model regards the relationship as a transfer on a specific relationship hyperplane, using the normal vector w of the hyperplane r and relation transfer vector d r Representation, first map the head and tail entities to the hyperplane, and get the mapped entities Then construct h ┷ +d r ≈t ┷

Images