Building a Vietnamese Dataset for Natural Language Inference Models.

Area of Science:

• Natural Language Processing
• Machine Learning
• Computational Linguistics

Background:

• Natural language inference (NLI) models are crucial for natural language understanding (NLU) tasks.
• State-of-the-art NLI models require high-quality annotated datasets for training and fine-tuning.
• Existing datasets may not be optimized for specific languages like Vietnamese, potentially leading to suboptimal model performance.

Purpose of the Study:

• To propose a novel method for constructing a Vietnamese dataset tailored for training Vietnamese NLI models.
• To address challenges in Vietnamese text processing, specifically the removal of cue marks and ensuring appropriate writing style.
• To enhance the semantic understanding capabilities of NLI models by minimizing reliance on superficial textual cues.

Main Methods:

• Developed a data preprocessing technique to remove cue marks from Vietnamese text, preventing models from learning spurious correlations.
• Implemented a method to ensure the natural writing style of Vietnamese texts within the dataset.
• Fine-tuned a BERT-based model (viNLI) on the newly created Vietnamese dataset.

Main Results:

• The viNLI model, trained on the proposed dataset, achieved an accuracy of 94.79% on a Vietnamese test set.
• In comparison, a BERT model (viXNLI) fine-tuned on the XNLI dataset achieved only 64.04% accuracy on the same test set.
• In an answer selection experiment, viNLI demonstrated superior performance with a P@1 score of 0.4949 compared to viXNLI's 0.4044.

Conclusions:

• The proposed method is effective in building a high-quality Vietnamese natural language inference dataset.
• The dataset enables the development of significantly more accurate and robust Vietnamese NLI models.
• This work contributes to advancing NLU applications for the Vietnamese language.