# Eliminating Sentiment Bias for Aspect-Level Sentiment Classification with Unsupervised Opinion Extraction

**Authors:** Bo Wang, Tao Shen, Guodong Long, Tianyi Zhou, Yi Chang  
**arXiv:** [2109.02403](https://arxiv.org/abs/2109.02403)  
**Date:** 2021-09-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2109.02403v2/

## Key points

- The paper addresses the problem of sentiment bias in aspect-level sentiment classification (ALSC)
- ALSC aims to identify the sentiment polarity of a specified aspect in a sentence
- Recent works have fine-tuned pre-trained Transformer encoders for ALSC, but these induced opinion words only provide an intuitive cue far below human-level interpretability
- The pre-trained encoder tends to internalize an aspect's intrinsic sentiment, causing sentiment bias and thus affecting model performance
- The paper proposes a span-based anti-bias aspect representation learning framework that achieves new state-of-the-art performance on five benchmarks and has unsupervised opinion extraction capabilities
- The authors use two datasets for evaluation: SemEval 2014 Task 4 and SemEval 2015 Task 12, both of which contain reviews from various domains such as restaurants and laptops
- They compare their proposed method with several baselines and show that it outperforms them on all datasets
- They also conduct ablation studies to demonstrate the effectiveness of each component of their proposed method
- They analyze how their method reduces bias by examining attention weights and showing that their method produces more balanced attention distributions than other methods
- Finally, they discuss future directions for research in this area, including exploring different ways of incorporating external knowledge into ALSC models and investigating how different types of biases affect model performance.

## Summary

This paper addresses the problem of sentiment bias in aspect-level sentiment classification (ALSC), which aims to identify the sentiment polarity of a specified aspect in a sentence. While ALSC is a practical setting in aspect-based sentiment analysis due to no opinion term labeling needed, it fails to interpret why a sentiment polarity is derived for the aspect. Recent works have fine-tuned pre-trained Transformer encoders for ALSC to extract an aspect-centric dependency tree that can locate the opinion words, but these induced opinion words only provide an intuitive cue far below human-level interpretability. Additionally, the pre-trained encoder tends to internalize an aspect's intrinsic sentiment, causing sentiment bias and thus affecting model performance. To address these issues, this paper proposes a span-based anti-bias aspect representation learning framework. The framework first eliminates the sentiment bias in the aspect embedding by adversarial learning against aspects' prior sentiment. Then, it aligns the distilled opinion candidates with the aspect by span-based dependency modeling to highlight interpretable opinion terms. This method achieves new state-of-the-art performance on five benchmarks and has unsupervised opinion extraction capabilities. The authors use two datasets for evaluation: SemEval 2014 Task 4 and SemEval 2015 Task 12, both of which contain reviews from various domains such as restaurants and laptops. They compare their proposed method with several baselines and show that it outperforms them on all datasets. They also conduct ablation studies to demonstrate the effectiveness of each component of their proposed method. Furthermore, they analyze how their method reduces bias by examining attention weights and showing that their method produces more balanced attention distributions than other methods. Finally, they discuss future directions for research in this area, including exploring different ways of incorporating external knowledge into ALSC models and investigating how different types of biases affect model performance. Overall, this paper presents a novel approach for addressing sentiment bias in ALSC and achieves state-of-the-art performance on multiple benchmarks. The proposed method has the potential to improve the interpretability of opinion extraction and advance research in aspect-based sentiment analysis.