Aspect sentiment learning for Aspect-Level Sentiment Classification.

Abstract

Aspect-Level Sentiment Classification (ALSC) is a fine-grained Sentiment Analysis (SA) task that aims to determine the sentiments of a sentence toward different aspects. Despite their significant success, most existing methods derive aspect sentiment semantics from individual sentences, overlooking the interrelationships among relevant sentences that could provide a more comprehensive understanding of aspect sentiment semantics. To this end, we propose AspLearn, an aspect-learning method to optimize aspect sentiment semantics and generate more robust aspect-specific sentence features for the ALSC task. In a nutshell, AspLearn employs the Aspect-aware Contrastive Learning (AspCL) to mine valuable aspect-related knowledge from aspect-relevant samples, thereby optimizing aspect sentiment semantics and enhancing the model's performance. AspLearn is a simple yet effective method, with its superior aspect learning capabilities confirmed through extensive experiments on three benchmarks. Notably, AspLearn, using DeBERTa as the backbone, achieves Macro F1 score improvements of 3.13%, 0.76%, and 1.07% over the second-best results on the Laptops, Restaurants, and Twitter datasets, respectively. Furthermore, AspLearn's mechanism can retrieve the most relevant demonstrations for Large Language Models (LLMs), enhancing their sentiment recognition capabilities.