IndoBERT-SupCon: A Supervised Contrastive Learning Model for Analyzing Public Perception on Halal Tourism

Abstract

The primary objective of this research is to develop and evaluate a robust deep learning model for accurately analyzing stakeholder perceptions of halal tourism development in Pariaman, West Sumatra, based on qualitative textual data. The main contribution is the introduction of IndoBERT-SupCon, a novel architecture that enhances the Indonesian BERT model with a Supervised Contrastive Learning (SupCon) mechanism. A novel method for producing more discriminative feature representations for complex viewpoints is presented in this paper, which is one of the first to use this sophisticated fine-tuning technique to Indonesian socio-political sentiment analysis. Conceptually, the model is trained to simultaneously minimize classification error while optimizing the feature space, pulling representations of similar sentiments closer together and pushing dissimilar ones further apart. To achieve this, we collected 1,022 primary textual responses through online surveys with tourists and in-depth interviews with key stakeholders, including SME owners and government officials. The SMOTE oversampling technique was employed on the training data to mitigate class imbalance. Experimental results on the test data demonstrate that the IndoBERT-SupCon model achieved outstanding performance, with a final accuracy of 96.59% and a macro F1-score of 0.97. These results significantly surpass the performance of a standard fine-tuned IndoBERT baseline, confirming the effectiveness of the SupCon approach. The findings provide the Pariaman local government with a highly valid, data-driven tool for more responsive and effective policy formulation. This research offers a robust framework that can be applied to other public policy domains, showcasing the value of advanced deep learning in transforming qualitative stakeholder feedback into actionable insights.