Joint Extraction of Opinion Targets and Opinion Expressions Based on Cascaded Model

摘要

Fine-grained opinion analysis is a very important task, especially identifying opinion target and opinion expression. In this paper, a new neural architecture is proposed for the sentence-level joint extraction of opinion target and opinion expression. The neural architecture namely cascaded model includes pre-trained model BERT Base, linguistic features, bi-directional LSTM, soft attention network and CRF layer from bottom to top. The cascaded model provides the best joint extraction results in the SemEval-2014/2016 Task 4/5 data sets compared with the state-of-the-art. There are three main contributions in our work, (1) attention network is introduced into the task of sentence-level joint extraction of opinion target and opinion expression, which enhances the dependence between opinion target and opinion expression. (2) pre-trained model BERT-Base and linguistic features are introduced into our work, which greatly improve the convergence speed and the performance of the cascaded model. (3) opinion target and opinion expression are synchronously extracted, and achieved better results compared with the most of the existing pipelined methods.