Fine-Grained Sentiment Labeling with Retrieval-Augmented LLMs: A Computational Approach

ACL ARR 2025 May Submission4592 Authors

Abstract: Sentiment analysis has evolved from coarse-grained classification (e.g., positive/negative) to fine-grained dimensional labeling (e.g., valence and arousal), yet accurately scoring emotional intensity remains challenging. While large language models (LLMs) show promise, their direct application suffers from domain adaptation issues and inconsistency in fine-grained annotations. To address this, we propose a retrieval-augmented multi-LLM ensemble framework that combines dynamic knowledge retrieval with weighted model collaboration. Experiments on the Chinese EmoBank corpus demonstrate significant improvements of our proposed model on arousal labeling compared to zero-shot LLM baselines. In this paper, we introduce an augmented generation strategy based on cross-lingual (Chinese-English) retrieval and propose a data-driven weighting mechanism to assign model importance based on task-specific performance, providing a repeatable open source implementation. The results highlight the key role of retrieval-augmented generation in fine-grained sentiment analysis. Our work provides a scalable solution for real-world sentiment analysis applications such as social media, product reviews, and opinion monitoring, and lays a foundation for future extensions of multi-modal affective computing.

Paper Type: Short

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

Research Area Keywords: Sentiment analysis, retrieval-augmented generation, domain adaptation

Contribution Types: Model analysis & interpretability, Data analysis

Languages Studied: Chinese, English

Submission Number: 4592