Knowledge graph completion method based on neural network

Abstract

The embodiment of the invention provides a knowledge graph completion method based on a neural network, and relates to the technical field of natural language processing. The method comprises the steps of S1, converting triples (s, r, o) in a knowledge graph K into sentences [s r o] in an embedding layer, and converting the sentences [s r o] into a k-dimensional dense vector representation form xi= [vs, vr, vo] by utilizing a vector representation technology; s2, learning long-term dependence of sentences from the input feature vector xi in a loop layer by utilizing a BiLSTM network to obtaina loop layer feature vector hRNN; s3, learning local structure information from the feature vector hRNN in a convolutional layer by using a convolutional neural network to obtain a convolutional layer feature vector hCNN; and S4, converting the convolutional layer feature vector hCNN into a score of each triple (s, r, o) in a full connection layer. According to the knowledge graph completion method, the long-term dependence and local structure information of facts in the knowledge graph are captured by using the recurrent and convolutional neural network without depending on any external data, and meanwhile, the transfer characteristics of entities and relationships are reserved, so that the learning ability is higher.

Description

technical field [0001] The present invention relates to the technical field of natural language processing, in particular to a neural network-based knowledge graph completion method. Background technique [0002] With the advent of the big data era, knowledge graphs have gradually become a current research hotspot. The knowledge graph is used to store structured facts in the real world, and its essence is a semantic network, with nodes representing entities and edges representing the relationships between entities. Knowledge graphs are widely used in many scenarios, such as semantic search, intelligent question answering, auxiliary decision-making, etc. However, the knowledge graph faces serious problems such as data sparsity and missing data. In recent years, many knowledge graph completion methods have been proposed to fill in missing facts. [0003] Traditional knowledge graph completion models are mainly divided into two categories: one is the embedding model of shall...

AI Technical Summary

Problems solved by technology

Another article "A novel embedding model forknowledge base completion based on convolutional neural network" published at the North American Annual Conference of the Association for Computational Linguistics in 2018 uses 1D convolution instead of 2D convolution, converting each fact into a 3-column matrix, and Extract the global relationship between embeddings from the same dimension, it only uses filters with a shape of 1×3, and does not take advantage of the convolution operation well

[0004] However, the above model ignores the fact that the facts in the knowledge graph come from the text, and only embeds part of the facts in the knowledge graph without contextual information, so the feature learning ability is still insufficient.

Images