T3L: Translate-and-Test Transfer Learning for Cross-Lingual Text Classification

Abstract Cross-lingual text classification leverages classifiers trained in a high-resource language to perform other languages with no or minimal fine-tuning (zero/ few-shots cross-lingual transfer). Nowadays, are typically built on large-scale, multilingual models (LMs) pretrained variety of interest. However, the performance these varies significantly across and tasks, suggesting that superposition modelling tasks is not always effective. For this reason, paper we propose revisiting classic “translate-and-test” pipeline neatly separate translation stages. The proposed approach couples 1) neural machine translator translating from targeted language, 2) classifier but generates “soft” translations permit end-to-end backpropagation during pipeline. Extensive experiments have been carried out over three datasets (XNLI, MLDoc, MultiEURLEX), results showing has improved competitive baseline.