[Paper Review] Bidirectional Machine Reading Comprehension for Aspect Sentiment Triplet Extraction

Shaowei Chen, Yu Wang+2 more|arXiv (Cornell University)|Mar 13, 2021

Sentiment Analysis and Opinion Mining4 citations

TL;DR

This paper proposes a bidirectional machine reading comprehension (BMRC) framework for aspect sentiment triplet extraction (ASTE), transforming the task into a multi-turn MRC process with non-restrictive, restrictive, and sentiment classification queries. By leveraging bidirectional reasoning—first extracting aspects then opinions, and vice versa—the model captures complex relations and improves triplet extraction, achieving state-of-the-art performance across four benchmark datasets.

ABSTRACT

Aspect sentiment triplet extraction (ASTE), which aims to identify aspects from review sentences along with their corresponding opinion expressions and sentiments, is an emerging task in fine-grained opinion mining. Since ASTE consists of multiple subtasks, including opinion entity extraction, relation detection, and sentiment classification, it is critical and challenging to appropriately capture and utilize the associations among them. In this paper, we transform ASTE task into a multi-turn machine reading comprehension (MTMRC) task and propose a bidirectional MRC (BMRC) framework to address this challenge. Specifically, we devise three types of queries, including non-restrictive extraction queries, restrictive extraction queries and sentiment classification queries, to build the associations among different subtasks. Furthermore, considering that an aspect sentiment triplet can derive from either an aspect or an opinion expression, we design a bidirectional MRC structure. One direction sequentially recognizes aspects, opinion expressions, and sentiments to obtain triplets, while the other direction identifies opinion expressions first, then aspects, and at last sentiments. By making the two directions complement each other, our framework can identify triplets more comprehensively. To verify the effectiveness of our approach, we conduct extensive experiments on four benchmark datasets. The experimental results demonstrate that BMRC achieves state-of-the-art performances.

Motivation & Objective

• To address the challenge of jointly extracting aspects, opinion expressions, and sentiments in a unified framework.
• To model complex, non-trivial relations between aspects and opinion expressions, including one-to-many, many-to-one, and overlapping cases.
• To improve sentiment classification by leveraging detected aspect-opinion relations through joint learning.
• To overcome the limitation of unidirectional extraction by enabling both aspect-first and opinion-first reasoning paths.
• To formalize ASTE as a multi-turn machine reading comprehension task to better capture inter-task dependencies.

Proposed method

• The ASTE task is formalized as a multi-turn machine reading comprehension (MTMRC) task with three query types: non-restrictive extraction, restrictive extraction, and sentiment classification.
• Non-restrictive queries identify the first entity (aspect or opinion expression) in a pair, while restrictive queries locate the second entity using the first as prior knowledge.
• A bidirectional MRC structure enables two reasoning directions: one starting with aspects (A→O), the other with opinion expressions (O→A), allowing complementary extraction.
• The sentiment classification turn uses extracted aspect-opinion pairs as context to predict the corresponding sentiment polarity.
• The framework employs joint learning across all three turns, enabling relations to guide sentiment classification.
• BERT is used as the base encoder to capture contextual representations, though the model performs strongly even without BERT.

Experimental results

Research questions

• RQ1 Can transforming ASTE into a multi-turn MRC task improve the joint extraction of aspects, opinion expressions, and sentiments?
• RQ2 How does bidirectional reasoning—starting from either aspect or opinion expression—enhance triplet extraction performance compared to unidirectional approaches?
• RQ3 To what extent do relation-aware sentiment classification and query-based interaction improve sentiment classification accuracy?
• RQ4 How much does the proposed query design contribute to capturing complex aspect-opinion relations, such as overlapping or embedded structures?
• RQ5 What is the relative contribution of BERT versus the proposed architectural innovations (e.g., bidirectional MRC, query types) to overall performance?

Key findings

• The BMRC model achieves state-of-the-art performance on four benchmark datasets: 59.27% F1 on 14-Lap, 70.69% on 14-Res, 61.05% on 15-Res, and 68.13% on 16-Res.
• The bidirectional MRC structure significantly improves performance, with each unidirectional variant ('Ours w/o AO' and 'Ours w/o OA') underperforming due to biased extraction paths.
• Removing relation-aware sentiment classification ('Ours w/o REQ') causes a substantial drop in aspect-opinion pair and triplet extraction F1, proving that relations enhance sentiment classification.
• The model outperforms TSF by 8.15% average F1 on triplet extraction even without BERT, showing that architectural innovations contribute more than BERT alone.
• The full model improves by 10.4% F1 over 'Ours w/o REQ', demonstrating that the combination of bidirectional MRC and query design yields the greatest performance gain.