Error Investigation of Pre-trained BERTology Models on Vietnamese Natural Language Inference

Natural Language Inference tasks have emerged in recent years and attracted significant attention from the natural language processing research community. There has been much success in this task with many quality datasets in English and Chinese for research and demonstrating the impressive performance of machine learning models. Pre-trained models play a crucial role, which is reflected in their superior performance compared to other models. However, they are still far from perfect and have many obstacles to the characteristics of the data. Especially in Vietnamese, we have just seen the emergence of the ViNLI benchmark dataset to serve the research community. In this paper, we experiment and analyze how the characteristics in the ViNLI benchmark dataset affect the performance of the pre-trained BETology-based models. In addition, the data parameters of ViNLI are also measured and analyzed on the accuracy of these models to see if it has any impact on the accuracy of the model.