A hierarchical and parallel framework for End-to-End Aspect-based Sentiment Analysis

Abstract Pipeline, joint, and collapsed models are three major approaches to solving End-to-End Aspect-based Sentiment Analysis (E2E-ABSA) task. Prior works found that joint models were consistently surpassed by the other two. To explore the potential of joint model for E2E-ABSA, we propose a hierarchical and parallel joint framework on the basis of exploiting the hierarchical nature of the pre-trained language model and performing parallel inference of the subtasks. Our framework: (1) shares the same pre-trained backbone network between two subtasks, ensuring the associations and commonalities between them; (2) considers the hierarchical feature of the deep neural network and introduces two joint approaches, namely the specific-layer joint model and multiple-layer joint model, coupling two specific layers or multiple task-related layers with subtasks; (3) carries out parallel execution in both training and inference processes, improving the inference throughput and al-leviating the target-polarity mismatch problem. The experimental results on three benchmark datasets demonstrate that our approach outperforms the state-of-the-art works.