Mitigating Data Sparsity in Code-Mixed Text through Back-Translation Augmentation for Aspect-Based Sentiment Analysis in Tokopedia Reviews
Abdul Hakim Prima Yuniarto, Aini Shofi Achsanti, Rizky Agil Singgih Susanto, Ardena Afif Pratama, Fandy Setyo UtomoNon-Scopus15 Februari 2026Afiliasi: ID
Abstrak
Aspect-Based Sentiment Analysis (ABSA) in e-commerce reviews in Indonesia faces significant challenges, including the use of mixed language or code-mixed language and limited labeled data, or data sparsity. This study proposes the use of Back-Translation data augmentation techniques to enrich Tokopedia's mixed Indonesian-English or South Jakarta language review dataset. Using the IndoBERT model, experimental results show a 3% increase in accuracy for both aspect and sentiment classification. These findings demonstrate that artificial data augmentation is effective in addressing data sparsity constraints in informal texts and improving the reliability of macro analysis for strategic platform recommendations.
Identifier
- OpenAlex: https://openalex.org/W7160430532
- DOI: https://doi.org/10.59934/jaiea.v5i2.2197