Efficient Multilingual Service Classification with DistilBERT

The growing demand for multilingual platforms has highlighted the need for efficient, scalable systems that can accurately classify content in multiple languages. Traditional multilingual classification approaches, such as language-specific models or manual translation, are often resource-intensive and difficult to scale. This paper explores the use of DistilBERT, a lightweight variant of the popular BERT model, for multilingual service classification, specifically for categorizing API descriptions. DistilBERT retains 97% of BERT’s performance while reducing model size and inference time by 60%, making it ideal for resource-constrained environments. We evaluate DistilBERT’s performance on a multilingual dataset of 16,753 API descriptions categorized into three configurations: 7, 5, and 3 classes. Experimental results demonstrate that DistilBERT achieves competitive accuracy, particularly in coarser-grained classification tasks, with an accuracy of 78.30% in the 3-class configuration. The study also addresses challenges such as data imbalance, linguistic nuances, and the computational constraints of multilingual models. The findings suggest that DistilBERT provides an efficient solution for real-time multilingual classification, offering significant advantages in both performance and resource efficiency, while highlighting areas for future improvement, particularly in handling fine-grained classification and low-resource languages.