End-to-End Aspect-Level Sentiment Analysis Based on Directed Syntactic Dependency Trees

As aspect-based sentiment analysis tasks (ABSA) continue to evolve, end-to-end ABSA tasks (E2E-ABSA) are emerging as a new research direction. Most models for this type of task obtain valid classification features from syntactic and semantic information of the text, but the commonly used syntactic dependency trees ignore the direction of word dependencies when generating syntactic information. To resolve this problem, a combined model of Bi-directional Graph Convolutional Network incorporating Attention Mechanisms (Bi-ATGCN) and Bi-directional Long Short-Term Memory Network (Bi-LSTM) is proposed. Firstly, Bert is used as a pre-training model to obtain dynamic word vectors containing contextual information. Secondly, a directed syntactic dependency tree is introduced, and the syntactic information of the text is captured using the Bi-directional Graph Convolutional Network (Bi-GCN), and the dependency relationship between words is optimized by fusing Attention Mechanism when generating the adjacency matrix of Bi-ATGCN. Finally, the contextual semantic information of the text is captured using Bi-LSTM and thus better perform the E2E-ABSA task. The classification performance of the combined model is improved when compared with related benchmark model experiments on LAPTOP and REST datasets.