Joint Opinion Target and Target-oriented Opinion Words Extraction by BERT and IOT Model

Abstract

In this paper, we investigate two sub-tasks of aspect-based sentiment analysis (ABSA) through the pre-trained language model BERT, namely opinion target extraction (OTE) and target-oriented opinion words extraction (TOWE). Specifically, we build a novel framework for the joint extraction model of opinion target and target-oriented opinion words feedback, which aims to extract the opinion target and corresponding opinion words. In order to accomplish the TOWE task more effectively, we proposed an IO-LSTM+Transformer structure, termed IOT, which has excellent performance in domain-specific datasets when combined with the BERT pre-training model. To validate the effectiveness of our model, we develop a pipeline model for comparison. Experiment results show that our model can extract the pair of opinion target and opinion words from the sentence more effectively than the pipeline model. Therefore, our joint model has the potential to facilitate other tasks of ABSA.