Data Augmentation Methods for Semantic Similarity Detection in Vietnamese Questionnaire

Recently, applying Data Augmentation in low resource languages has emerged as an obvious remedy for assisting deep learning models in achieving the state-of-the-art performance. Yet, there have not been many studies conducted for Vietnamese - considered as a low resource language. This paper proposes two new text augmentation methods for detecting semantic similarity between a pair of questions, named Synonym Replacement and Definition Replacement. The two methods exploit the shared keywords between questions to expand the dataset while guaranteeing the meaning preservation of the original sentences. Experimental results show that both approaches improve the performance of BERT and PhoBERT models. Significantly, the BERT with datasets enriched by Synonym Replacement gains 0.91 in accuracy and F1-score for Similarity Detection task. It implies that the offered methods can be adapted to optimize the efficacy of other natural language processing tasks.