Mitigating Sentiment Recognition Bias in Aspect-Based Sentiment Analysis via Multifaceted Data Enhancement

ACL ARR 2024 December Submission294 Authors

12 Dec 2024 (modified: 05 Feb 2025)ACL ARR 2024 December SubmissionEveryoneRevisionsBibTeXCC BY 4.0
Abstract: Aspect-Based Sentiment Analysis (ABSA) focuses on analyzing the sentiment of specific aspect terms. Despite substantial progress in this field, most models often exhibit significant biases, particularly in recognizing neutral sentiments, due to the predominance of emotional content in training datasets. To improve the quality of data and enhance model comprehension of aspect term sentiments across diverse context, we propose the Multifaceted Data Enhancement (MDE) framework, which enhances both the breadth and depth of ABSA datasets. MDE leverages large language models (LLMs) for data paraphrasing and implements a Dual Confidence Filtering algorithm to select high-quality samples, thereby enhancing data diversity. Furthermore, MDE incorporates data enhancement strategies for aspect term clarification and sentiment reasoning. Through multiple rounds of inquiry with LLMs, MDE refines the understanding of aspect terms and strengthens the logical consistency between data and sentiment labels. We apply MDE to several ABSA benchmark datasets and fine-tune various models. Experimental results demonstrate that MDE effectively mitigates sentiment recognition bias and outperforms state-of-the-art baselines.
Paper Type: Long
Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining
Research Area Keywords: sentiment analysis, data augmentation, emotion detection and analysis
Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data analysis
Languages Studied: English
Submission Number: 294
Loading

OpenReview is a long-term project to advance science through improved peer review with legal nonprofit status. We gratefully acknowledge the support of the OpenReview Sponsors. © 2025 OpenReview