[Paper Review] A Hierarchical Interactive Network for Joint Span-based Aspect-Sentiment Analysis

Wei Chen, Jinglong Du+3 more|arXiv (Cornell University)|Aug 24, 2022

Sentiment Analysis and Opinion Mining4 citations

TL;DR

This paper proposes a Hierarchical Interactive Network (HI-ASA) for joint span-based aspect-sentiment analysis, leveraging shallow-level cross-stitch fusion and deep-level mutual information maximization to enable two-way feature interaction between aspect extraction and sentiment classification. HI-ASA achieves state-of-the-art performance across three benchmark datasets, outperforming prior methods by up to 5.16% on sentiment classification.

ABSTRACT

Recently, some span-based methods have achieved encouraging performances for joint aspect-sentiment analysis, which first extract aspects (aspect extraction) by detecting aspect boundaries and then classify the span-level sentiments (sentiment classification). However, most existing approaches either sequentially extract task-specific features, leading to insufficient feature interactions, or they encode aspect features and sentiment features in a parallel manner, implying that feature representation in each task is largely independent of each other except for input sharing. Both of them ignore the internal correlations between the aspect extraction and sentiment classification. To solve this problem, we novelly propose a hierarchical interactive network (HI-ASA) to model two-way interactions between two tasks appropriately, where the hierarchical interactions involve two steps: shallow-level interaction and deep-level interaction. First, we utilize cross-stitch mechanism to combine the different task-specific features selectively as the input to ensure proper two-way interactions. Second, the mutual information technique is applied to mutually constrain learning between two tasks in the output layer, thus the aspect input and the sentiment input are capable of encoding features of the other task via backpropagation. Extensive experiments on three real-world datasets demonstrate HI-ASA's superiority over baselines.

Motivation & Objective

• To address the limitation of existing span-based models that fail to model bidirectional interactions between aspect extraction and sentiment classification.
• To improve joint performance by enabling both shallow-level and deep-level interactions between the two tasks.
• To design a framework where aspect and sentiment features mutually inform each other during training via backpropagation.
• To empirically validate that hierarchical interaction mechanisms lead to consistent gains across multiple datasets.

Proposed method

• Employ a cross-stitch mechanism in the encoding layer to selectively fuse task-specific features, enabling shallow-level two-way interaction.
• Introduce mutual information maximization in the output layer to enforce deep-level interaction by constraining predictions across tasks.
• Use shared input embeddings with task-specific heads to maintain modularity while enabling cross-task knowledge transfer.
• Control the strength of interaction via a hyperparameter α, which balances task-specific and shared representations.
• Train the model end-to-end using a joint loss combining aspect extraction and sentiment classification objectives.
• Apply gradient-based backpropagation to allow sentiment features to guide aspect detection and vice versa through mutual supervision.

Experimental results

Research questions

• RQ1 Can hierarchical two-way interaction between aspect extraction and sentiment classification improve joint performance in span-based aspect-sentiment analysis?
• RQ2 How does the proposed cross-stitch mechanism at the encoding layer affect feature sharing between tasks?
• RQ3 To what extent does mutual information maximization in the output layer enhance cross-task generalization?
• RQ4 What is the optimal balance of interaction strength (controlled by α) across different datasets with varying data scales?
• RQ5 Does the proposed method outperform sequential and parallel encoding baselines in both aspect extraction and sentiment classification tasks?

Key findings

• HI-ASA achieves state-of-the-art F1-score of 85.02 on the Laptop dataset for aspect extraction, surpassing the previous best method by 3.93 points.
• On sentiment classification, HI-ASA achieves 93.18% accuracy on the Laptop dataset, outperforming the strongest baseline SPRM by 3.98 percentage points.
• The ablation study shows that removing either the shallow interaction (w/o SI) or deep interaction (w/o DI) leads to consistent performance drops, confirming the necessity of both components.
• The optimal hyperparameter α is 0.1 on the Laptop and Tweets datasets, and 0.4 on the Restaurant dataset, indicating that interaction strength should scale with dataset size.
• The model performs best when α is non-zero, proving that the interaction mechanism is effective and not redundant compared to parallel encoding (α = 0).
• The model demonstrates strong generalization, achieving consistent improvements across three diverse datasets: Laptop, Restaurant, and Tweets.