[Paper Review] A Joint Training Dual-MRC Framework for Aspect Based Sentiment Analysis

Yue Mao, Yi Shen+2 more|arXiv (Cornell University)|Jan 4, 2021

Sentiment Analysis and Opinion Mining38 references29 citations

TL;DR

The paper proposes a unified dual-MRC framework that converts ABSA into two question-answering tasks and jointly trains two BERT-MRC models to extract aspect terms, opinion terms, and sentiment, achieving state-of-the-art results across ABSA subtasks.

ABSTRACT

Aspect based sentiment analysis (ABSA) involves three fundamental subtasks: aspect term extraction, opinion term extraction, and aspect-level sentiment classification. Early works only focused on solving one of these subtasks individually. Some recent work focused on solving a combination of two subtasks, e.g., extracting aspect terms along with sentiment polarities or extracting the aspect and opinion terms pair-wisely. More recently, the triple extraction task has been proposed, i.e., extracting the (aspect term, opinion term, sentiment polarity) triples from a sentence. However, previous approaches fail to solve all subtasks in a unified end-to-end framework. In this paper, we propose a complete solution for ABSA. We construct two machine reading comprehension (MRC) problems and solve all subtasks by joint training two BERT-MRC models with parameters sharing. We conduct experiments on these subtasks, and results on several benchmark datasets demonstrate the effectiveness of our proposed framework, which significantly outperforms existing state-of-the-art methods.

Motivation & Objective

• Motivate the need for a unified ABSA framework that handles AE, OE, and SC end-to-end.
• Propose converting ABSA subtasks into two MRC problems with shared-parameter BERT backbones.
• Demonstrate joint training of two MRC models to cover AE, OE, SC, AESC, Pair, and Triple tasks.
• Evaluate on benchmark ABSA datasets and show improvements over state-of-the-art methods.

Proposed method

• Two BERT-MRC models with shared parameters encode the sentence context.
• Left MRC identifies aspect term (AE) start/end spans.
• Right MRC identifies opinion term (OE) spans and sentiment polarity (SC) for each AT.
• A joint loss combines AE, OE/SC (AOE), and sentiment classification objectives with tunable weights.
• MRC dataset construction uses q1: 'Find the aspect terms in the text' and q2(AT): 'Find the sentiment polarity and opinion terms for AT in the text'.

Experimental results

Research questions

• RQ1 Can ABSA subtasks be solved in a unified end-to-end framework via dual-MRC modeling?
• RQ2 Do shared-parameter BERT-MRC models improve extraction and classification for AE, OE, SC, AESC, Pair, and Triple tasks compared to task-specific approaches?
• RQ3 Is end-to-end joint training more effective and efficient than a two-stage or pipeline approach for ABSA?
• RQ4 How does converting ABSA tasks to MRC queries impact boundary detection and sentiment-consistency issues?

Key findings

• The dual-MRC framework can handle AE, OE, SC, AESC, Pair, and Triple within a single model.
• Joint training with shared parameters yields state-of-the-art or competitive results across multiple ABSA subtasks on benchmark datasets.
• Span-based extraction combined with MRC improves boundary detection and sentiment alignment compared with prior unified tagging approaches.
• Inference is performed in a pipeline fashion where left MRC outputs ATs and right MRC outputs OTs and sentiment for each AT.
• Empirical results show notable improvements over baselines on datasets derived from SemEval ABSA tasks.