Multiple-element joint detection for Aspect-Based Sentiment Analysis

Abstract

Aspect-Based Sentiment Analysis (ABSA) is a fine-grained sentiment analysis task, which aims to detect target-aspect-sentiment elements in a sentence. Most of the existing research work distinguished the sentiment for aspects or targets independently, ignoring the corresponding relation between the targets and the aspects. However, such a corresponding relation is significant for the accurate prediction of fine-grained sentiment polarity. In this paper, we propose a novel end-to-end multiple-element joint detection model (MEJD), which effectively extracts all (target, aspect, sentiment) triples from a sentence. Our model utilizes BERT to obtain the initial embedding vector from the aspect–sentence joint input and applies bidirectional long short-term memory to model aspect and sentence representations. We then employ a graph convolutional network with attention mechanisms to capture the dependency relationship between aspect and sentence. We evaluate our approach on two restaurant datasets of SemEval 2015 Task 12 and SemEval 2016 Task 5. Experiment results show that our model achieves state-of-the-art performance in extracting (target, aspect, sentiment) triples. Moreover, the model also has good performance on multiple subtasks of target-aspect-sentiment detection.