Contrastive Learning on IndoBERT for Sentiment Analysis of Free Nutritious Meal Policies

Abstract

Transformer-based language models such as IndoBERT still face limitations in topic and sentiment analysis of short social media texts, particularly due to embedding anisotropy, semantic overlap between topics, and limited sensitivity to implicit sentiment intensity. This study aims to evaluate the effectiveness of integrating SimCSE-based contrastive learning to optimize IndoBERT vector representations for sentiment analysis of the “Free Nutritious Meals” public policy. A comparative experimental approach was employed using an equal number of topics (three topics) and evaluated through BERTopic and Aspect-Based Sentiment Analysis (ABSA). The results demonstrate that the contrastive learning–based model substantially improves cluster separability, indicated by an increase of more than 1000% in the Silhouette Score compared to the baseline model, along with a reduction in topic overlap of approximately 40–50%. In addition, topic keyword diversity increased by more than 75%, yielding more informative and interpretable topic representations. In aspect-based sentiment analysis, the contrastive model exhibited approximately a 50% improvement in sensitivity to sentiment intensity and achieved perfect classification of implicit high-confidence sentiments that were previously misclassified as neutral by the baseline model. These findings confirm that contrastive learning–based embedding optimization effectively addresses the limitations of conventional embeddings and enhances the quality of topic modeling and aspect-based sentiment analysis for Indonesian social media texts.