VietX-NLI: A Cross-lingual Natural Language Inference Dataset with Vietnamese as the Source Language

Natural Language Inference (NLI) is a fundamental task in language understanding. While cross-lingual NLU has progressed rapidly, many multilingual datasets rely on English-centric translations, often failing to capture the linguistic and cultural nuances of target languages. Despite having over 100 million speakers, Vietnamese remains low-resource in NLP, especially in NLI. We introduce VietX-NLI, a new multilingual NLI dataset with Vietnamese as the source language. It contains over 100,000 premise-hypothesis pairs, translated into English and Chinese using Neural Machine Translation. Translation quality was validated using the Adequacy-Fluency Measure (AMFM), achieving an average score above 4.3/5. We benchmarked mBERT, XLM-R, mBART, and infoXLM, finding that XLMR<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">large</inf> performed best (over 70% accuracy and F1-macro). VietXNLI provides a valuable resource for advancing NLI and cross-lingual understanding, particularly for Vietnamese speakers.