[Paper Review] FaiMA: Feature-aware In-context Learning for Multi-domain Aspect-based Sentiment Analysis

Songhua Yang, Xinke Jiang+4 more|arXiv (Cornell University)|Mar 2, 2024

Sentiment Analysis and Opinion Mining4 citations

TL;DR

FaiMA proposes a feature-aware in-context learning framework for multi-domain aspect-based sentiment analysis, integrating graph neural networks with large language models via heuristic-optimized feature encoding and contrastive learning. It achieves a 2.07% average F1 improvement over baselines by enabling stable, multi-dimensional example retrieval for context-aware prompting.

ABSTRACT

Multi-domain aspect-based sentiment analysis (ABSA) seeks to capture fine-grained sentiment across diverse domains. While existing research narrowly focuses on single-domain applications constrained by methodological limitations and data scarcity, the reality is that sentiment naturally traverses multiple domains. Although large language models (LLMs) offer a promising solution for ABSA, it is difficult to integrate effectively with established techniques, including graph-based models and linguistics, because modifying their internal architecture is not easy. To alleviate this problem, we propose a novel framework, Feature-aware In-context Learning for Multi-domain ABSA (FaiMA). The core insight of FaiMA is to utilize in-context learning (ICL) as a feature-aware mechanism that facilitates adaptive learning in multi-domain ABSA tasks. Specifically, we employ a multi-head graph attention network as a text encoder optimized by heuristic rules for linguistic, domain, and sentiment features. Through contrastive learning, we optimize sentence representations by focusing on these diverse features. Additionally, we construct an efficient indexing mechanism, allowing FaiMA to stably retrieve highly relevant examples across multiple dimensions for any given input. To evaluate the efficacy of FaiMA, we build the first multi-domain ABSA benchmark dataset. Extensive experimental results demonstrate that FaiMA achieves significant performance improvements in multiple domains compared to baselines, increasing F1 by 2.07% on average. Source code and data sets are anonymously available at https://github.com/SupritYoung/FaiMA.

Motivation & Objective

• To address the challenge of generalizing aspect-based sentiment analysis across diverse domains, where single-domain methods fail due to data scarcity and domain shift.
• To integrate linguistic, domain-specific, and sentiment features into large language models without architectural modification.
• To develop a stable, multi-dimensional example retrieval mechanism that enhances in-context learning performance.
• To build the first comprehensive multi-domain ABSA benchmark dataset for standardized evaluation.
• To demonstrate that in-context learning can function as a feature-aware mechanism when guided by structured, representative examples.

Proposed method

• FaiMA employs a multi-head graph attention network as a text encoder, optimized using heuristic rules for linguistic, domain, and sentiment features.
• It applies contrastive learning to refine sentence representations by emphasizing diverse feature-specific signals.
• An efficient indexing mechanism enables stable retrieval of highly relevant examples across multiple feature dimensions for any input.
• In-context learning is leveraged as a feature-aware mechanism, where retrieved examples are incorporated into prompts to guide the LLM.
• The framework integrates graph-based encoding with LLMs without modifying the LLM’s internal architecture.
• The method supports zero-shot and few-shot adaptation across multiple domains through dynamic example selection.

Experimental results

Research questions

• RQ1 Can in-context learning be effectively repurposed as a feature-aware mechanism in multi-domain ABSA?
• RQ2 How can linguistic, domain, and sentiment features be systematically encoded and optimized for LLM-based ABSA?
• RQ3 What impact does contrastive learning have on sentence representation quality in multi-domain ABSA?
• RQ4 How does stable, multi-dimensional example retrieval improve in-context learning performance?
• RQ5 To what extent can a unified framework generalize across diverse domains without fine-tuning?

Key findings

• FaiMA achieves an average F1 improvement of 2.07% over strong baselines across multiple domains.
• The integration of graph-based feature encoding with in-context learning significantly enhances model generalization and robustness.
• The proposed indexing mechanism enables reliable retrieval of relevant examples across linguistic, domain, and sentiment features.
• Contrastive learning effectively improves sentence representation quality by emphasizing diverse feature signals.
• The framework demonstrates strong zero-shot and few-shot performance on the newly constructed multi-domain ABSA benchmark.
• FaiMA outperforms existing methods in both single-domain and cross-domain settings, validating its generalization capability.