QoMRC: Query-oriented Machine Reading Comprehension Framework for Aspect Sentiment Triplet Extraction

Aspect Sentiment Triplet Extraction (ASTE) is an essential task in aspect-based sentiment analysis. It aims to extract sentiment triplets from the context, which generally consists of three subtasks: aspect term extraction, opinion term extraction and sentiment classification. Existing methods mainly adopt a bidirectional machine reading comprehension framework to capture corresponding relations among subtasks. However, they input queries for different subtasks into the same encoder simultaneously, which leads the model to confuse the subtask associated with the query. To address this issue, we propose a novel Query-oriented Machine Reading Comprehension (QoMRC) framework which is a two-stage approach. In the first stage, QoMRC utilizes predefined queries and adapter tuning to efficiently generate three different query-oriented adapters for three subtasks that capture task-specific features. In the second stage, we fuse the query-oriented adapter and the shared encoder representation to obtain task-specific representation, which also reserves the correlation among subtasks. In addition, to reduce the semantic gap between the initialized adapters and the pre-trained BERT, we employ a layer-wise distillation approach in the first stage. Extensive experiment results on benchmark datasets show the efficacy of our proposed method, and indicate the necessity of capturing task-specific features.