Cross-Lingual Text Classification with Large Language Models

Cross-lingual text classification involves using a model trained on data in one language to classify text in another, which is crucial in global web applications where labeled data is scarce in certain languages. Although multilingual language models have been leveraged for such tasks, few studies investigate the ability of large language models in cross-lingual classification. In this paper, we evaluate the performance of four large language models and one smaller encoder-only model on cross-lingual text classification tasks using two benchmark datasets. We assess three task settings-direct-test without fine-tuning, direct-test with fine-tuning, and translate-test with fine-tuning. Our findings demonstrate that fine-tuning consistently improves performance, and the translate-test method slightly outperforms others. Furthermore, while large language models experience a slight performance drop on non-English languages, they outperform the baseline encoder-only model in many cases. This study provides insights into the effectiveness of large language models for cross-lingual classification in practical applications.