Zero-Shot and Few-Shot Sentiment Classification in Multilingual Customer Support Threads Using Cross-Lingual Language Models

• Abiodun Okunola

  Ladoke Akintola University Technology

  Author

The globalization of digital services has created an urgent need for scalable sentiment analysis systems capable of processing customer support interactions across multiple languages without extensive labeled data. While multilingual transformer models have demonstrated promise in cross-lingual transfer learning, limited empirical evidence exists regarding their effectiveness in the specific domain of customer support thread sentiment classification under zero-shot and few-shot conditions. This study investigates the performance of cross-lingual language models—specifically XLM-RoBERTa and mBERT—for sentiment classification in multilingual customer support conversations spanning four languages (English, French, Dutch, German). Using a design-based research methodology with controlled experiments on synthetic customer support datasets, we systematically evaluate model performance across zero-shot, few-shot (100, 500, and 1000 labeled samples), and fully supervised conditions. Our findings demonstrate that XLM-R achieves 82.9% zero-shot accuracy without target language fine-tuning, with performance improving to 88.6% with 1000 labeled samples per language . These results establish that cross-lingual models offer a viable, cost-effective alternative to monolingual systems for multilingual customer support sentiment analysis. The study contributes a replicable evaluation framework and provides practitioners with empirically grounded guidance for deploying sentiment analysis in resource-constrained multilingual environments.

• pdf

Copyright (c) 2026 Abiodun Okunola (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.