The XNLI (Cross-lingual Natural Language Inference) corpus is a benchmark for evaluating cross-lingual sentence understanding. It extends the MultiNLI corpus to 15 languages, including low-resource languages like Swahili and Urdu, providing 112,500 annotated sentence pairs. The corpus is designed to evaluate models trained in one language and tested in others, offering a practical and challenging evaluation suite. The authors propose baselines, including machine translation systems and multilingual sentence encoders, and find that directly translating test data yields the best performance among available baselines. The XNLI corpus is intended to catalyze research in cross-lingual sentence understanding by providing a standardized evaluation task. The paper discusses related work in multilingual word embeddings, sentence representation learning, and multilingual sentence representations. It presents results on cross-lingual NLI using translation-based approaches and multilingual sentence encoders, showing that aligned sentence embeddings can achieve competitive results. The experiments demonstrate that translation-based methods, particularly TRANSLATE TEST, perform well, but multilingual sentence encoders also show promise. The XNLI corpus is freely available for research and may be modified and redistributed. The study highlights the importance of cross-lingual evaluation in multilingual NLP and the challenges of transferring knowledge across languages. The results suggest that while translation-based methods are effective, multilingual sentence encoders can also achieve strong performance, offering an efficient alternative for cross-lingual understanding.The XNLI (Cross-lingual Natural Language Inference) corpus is a benchmark for evaluating cross-lingual sentence understanding. It extends the MultiNLI corpus to 15 languages, including low-resource languages like Swahili and Urdu, providing 112,500 annotated sentence pairs. The corpus is designed to evaluate models trained in one language and tested in others, offering a practical and challenging evaluation suite. The authors propose baselines, including machine translation systems and multilingual sentence encoders, and find that directly translating test data yields the best performance among available baselines. The XNLI corpus is intended to catalyze research in cross-lingual sentence understanding by providing a standardized evaluation task. The paper discusses related work in multilingual word embeddings, sentence representation learning, and multilingual sentence representations. It presents results on cross-lingual NLI using translation-based approaches and multilingual sentence encoders, showing that aligned sentence embeddings can achieve competitive results. The experiments demonstrate that translation-based methods, particularly TRANSLATE TEST, perform well, but multilingual sentence encoders also show promise. The XNLI corpus is freely available for research and may be modified and redistributed. The study highlights the importance of cross-lingual evaluation in multilingual NLP and the challenges of transferring knowledge across languages. The results suggest that while translation-based methods are effective, multilingual sentence encoders can also achieve strong performance, offering an efficient alternative for cross-lingual understanding.