RD-MCSA: A Multi-Class Sentiment Analysis Approach Integrating In-Context Classification Rationales and Demonstrations

ACL ARR 2025 February Submission6947 Authors

Abstract:

Multi-class sentiment analysis (MCSA) poses significant challenges due to its multiple categories and the subtle semantic distinctions between adjacent classes, necessitating substantial amounts of high-quality annotated data, which is often scarce.
This paper introduces RD-MCSA (Rationales and Demonstrations based Multi-Class Sentiment Analysis), an approach that enhances classification performance with limited labeled data. RD-MCSA leverages In-Context Learning (ICL) by integrating classification rationales and demonstration examples, enabling Large Language Models (LLMs) to make more accurate predictions.
In RD-MCSA, a representative set of annotated samples is constructed using a balanced Coreset algorithm to guide LLMs in generating classification rationales grounded in linguistic and semantic features.
These rationales are then integrated with demonstration examples, selected via a Multi-Kernel Gaussian Process (MK-GP)-based similarity evaluation method, to enhance ICL for MCSA.
Experiments on six diverse datasets demonstrate that RD-MCSA outperforms both supervised learning methods and conventional ICL approaches across key evaluation metrics.

Paper Type: Long

Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining

Research Area Keywords: stance detection; applications;

Contribution Types: Reproduction study, Approaches to low-resource settings, Publicly available software and/or pre-trained models, Data analysis

Languages Studied: English

Submission Number: 6947