SKG-Learning: a deep learning model for sentiment knowledge graph construction in social networks

Traditional sentiment analysis methods pay little attention to the inseparable relations between evaluation words and evaluation aspects, and the relations between evaluation words and topics. There have been many studies of knowledge graph (KG), which can effectively store and manage massive amounts of information and is suitable to associate emotion words with evaluation aspects and topics. This study proposes SKG-Learning based on a deep learning model to construct a sentiment knowledge graph (SKG) for sentiment analysis. Entity and relation are the cornerstones of SKG; thus, the task of SKG-Learning is divided into named entity recognition and relation extraction. We propose a bidirectional long short-term memory model (Bi-LSTM) with background knowledge embedding and co-extraction of features (BBC-LSTM) to extract entities. BBC-LSTM completes the embedding of background knowledge such as topic and emotion information and uses three-dimensional tensors to co-extract the deep features of aspect entities and sentiment entities. It solves the problems that it is difficult to recognize entities from insufficient context, and traditional models usually neglect the relevance between sentiment entities and aspect entities. A relation extraction model based on an encoder–decoder model (ED-Learning) is proposed to extract and classify the relation between sentiment and aspect entity, that is, the emotional tendency of sentiment entity toward aspect entity. Experiments show that the proposed methods can more efficiently extract entities and relations from social network texts. We confirm the validity of an SKG constructed by the SKG-Learning model in an emotional analysis task.