Knowledge representation method based on combination of text embedding and structure embedding

Abstract

The invention discloses a knowledge representation method based on the combination of text embedding and structure embedding, and the method comprises the steps: 1, carrying out the preprocessing of an entity description text in a knowledge base, and extracting a subject term from each entity description; 2, encoding the subject term into a term vector by using fasttext, wherein each entity description is expressed as a multi-dimensional term vector; step 3, inputting the processed multi-dimensional word vectors into a bidirectional long-short memory network (A-BiLSTM) with an attention mechanism or a long-short memory network (A-LSTM) with an attention mechanism for encoding, processing the multi-dimensional word vector representing each entity into a one-dimensional vector, namely text representation, and training an existing StransE model to obtain structural representation of the entity; 4, introducing a gating mechanism, and proposing four methods related to text embedding and structure embedding combination to obtain a final entity embedding matrix; and 5, inputting the entity embedding matrix into a ConvKB knowledge graph embedding model, a TransH knowledge graph embedding model, a TransR knowledge graph embedding model, a Distmult knowledge graph embedding model and a Hole knowledge graph embedding model, and improving a knowledge completion task.

Description

technical field [0001] The invention belongs to the fields of knowledge graph, natural language processing and information retrieval, and relates to heterogeneous information fusion and knowledge graph completion in knowledge representation learning, especially a kind of knowledge combining entity description-based text embedding and structure embedding in a knowledge base display method. Background technique [0002] Divide knowledge representation embedding methods into two aspects, one aspect is the structure-based embedding model using only symbolic triples. Bordes et al. were inspired by the interesting translation invariance phenomenon in the word vector space for the relationship between lexical semantics and syntax, and proposed the TransE model, which regards the relationship in the knowledge base as a certain translation vector between entities. Compared with previous models, the TransE model has fewer parameters and lower computational complexity, but it can dire...

AI Technical Summary

Problems solved by technology

(1) In these methods, the optimal combination of structural representation and textual representation has not been well studied, and a good repre

Images