[Paper Review] A Multi-task Learning Framework for Opinion Triplet Extraction

Chen Zhang, Qiuchi Li+2 more|arXiv (Cornell University)|Oct 4, 2020

Sentiment Analysis and Opinion Mining28 references4 citations

TL;DR

This paper proposes a multi-task learning framework for opinion triplet extraction, jointly detecting aspect terms, opinion terms, and their sentiment polarities using a biaffine scorer for sentiment dependency parsing. The method significantly outperforms strong baselines on four SemEval benchmarks, demonstrating improved performance and interpretability through decoupled prediction and triplet interaction modeling.

ABSTRACT

The state-of-the-art Aspect-based Sentiment Analysis (ABSA) approaches are mainly based on either detecting aspect terms and their corresponding sentiment polarities, or co-extracting aspect and opinion terms. However, the extraction of aspect-sentiment pairs lacks opinion terms as a reference, while co-extraction of aspect and opinion terms would not lead to meaningful pairs without determining their sentiment dependencies. To address the issue, we present a novel view of ABSA as an opinion triplet extraction task, and propose a multi-task learning framework to jointly extract aspect terms and opinion terms, and simultaneously parses sentiment dependencies between them with a biaffine scorer. At inference phase, the extraction of triplets is facilitated by a triplet decoding method based on the above outputs. We evaluate the proposed framework on four SemEval benchmarks for ASBA. The results demonstrate that our approach significantly outperforms a range of strong baselines and state-of-the-art approaches.

Motivation & Objective

• To address the limitations of existing aspect-based sentiment analysis (ABSA) approaches that either ignore opinion terms or fail to model sentiment dependencies.
• To propose a unified view of ABSA as opinion triplet extraction, integrating aspect-sentiment pair extraction and aspect-opinion co-extraction.
• To overcome the shortcomings of unified tagging schemes that degrade sentiment dependency parsing and neglect triplet interactions.
• To improve model interpretability and performance by jointly training aspect, opinion, and sentiment prediction with regularization across correlated tasks.

Proposed method

• The framework employs a multi-task learning setup to jointly extract aspect terms, opinion terms, and sentiment polarities from text.
• A biaffine scorer is used to model sentiment dependencies between extracted aspects and opinions, enabling end-to-end learning of sentiment relations.
• Triplet decoding is applied at inference to generate coherent (aspect, opinion, sentiment) triplets from predicted spans and dependencies.
• The model decouples aspect and sentiment prediction, avoiding the limitations of unified tagging schemes that constrain interaction modeling.
• The framework incorporates regularization between related tasks to enhance generalization and handle overlapping triplet patterns.
• It is trained end-to-end on four SemEval benchmarks using span detection and dependency parsing heads for aspect, opinion, and sentiment prediction.

Experimental results

Research questions

• RQ1 Can joint extraction of aspects, opinions, and sentiments improve performance in aspect-based sentiment analysis compared to existing methods?
• RQ2 How does decoupling sentiment prediction from aspect tagging affect the modeling of sentiment dependencies in overlapping triplet cases?
• RQ3 To what extent does modeling interactions between triplets (e.g., shared aspects or opinions) improve extraction accuracy?
• RQ4 Does the proposed multi-task learning framework with biaffine scoring and triplet decoding outperform strong baselines on standard ABSA benchmarks?
• RQ5 What are the main failure modes of the system, and how do they relate to overlapping triplet patterns or span detection errors?

Key findings

• The proposed framework achieves state-of-the-art performance on four SemEval benchmarks for aspect-based sentiment analysis, significantly outperforming strong baselines.
• The ablation study confirms that both decoupled aspect-sentiment prediction and the biaffine scorer for sentiment dependency parsing are essential components for optimal performance.
• Error analysis reveals that false positives are primarily caused by incorrect aspect or opinion span detection, suggesting a need for more robust tagging models.
• False negatives are most prevalent in overlapping triplet cases (aspect- or opinion-overlapped), indicating that such patterns remain challenging for current models.
• The results demonstrate that modeling triplet interactions—especially in overlapping scenarios—enhances performance, validating the framework's design choice.
• The framework shows improved interpretability by explicitly modeling sentiment dependencies between aspects and opinions, offering clearer reasoning paths than unified tagging approaches.