Complete quadruple extraction using a two-stage neural model for aspect-based sentiment analysis

Abstract

Aspect-based sentiment analysis (ABSA) is a fine-grained task which aims to identify the emotional polarity of a specific aspect in a text or sentence. Aspect term extraction (ATE), opinion term extraction (OTE) and aspect polarity classification (APC) are three main subtasks of the ABSA task. Nowadays, researchers mainly focus on a single task or a joint task composed of these three subtasks, and such investigation on the sentiment analysis is not sufficient. In this paper, we firstly introduce a complete aspect sentiment analysis task, called Aspect Sentiment Quadruple Extraction, which also includes the category detection beside ATE, OTE and APC. Then we propose a two-stage neural network model composed of several modules, including BiLSTM, simple gated self-attention and position encoding for this joint task. In the first stage, the proposed model extracts aspect and opinion terms as well as their categories and polarities. Moreover, the second stage mainly includes a relation classifier to validate the aspect-opinion pairs and then finalizes the complete quadruple extraction. The experimental results, evaluated on a benchmark dataset of Chinese product reviews, show that our proposed model outperforms other baseline methods and achieves the start-of-art performance.