vnNLI - VLSP2021: An Empirical Study on Vietnamese-English Natural Language Inference Based on Pretrained Language Models with Data Augmentation

Abstract

In this paper, we describe an empirical study of data augmentation techniques with various pre-trained language models on the bilingual dataset which was presented at the VLSP 2021 - Vietnamese and English-Vietnamese Textual Entailment. We apply the machine translation tool to generate new training set from original training data and then investigate and compare the effectiveness of a monolingual and multilingual model on the new data set. Our experimental results show that fine-tuning a pre-trained multilingual language XLM-R model with an augmented training set gives the best performance. Our system was ranked third in the shared-task VLSP 2021 with the F1-score of about 0.88.