This paper introduces the Cross-lingual Natural Language Inference (XNLI) corpus, an extension of the Multi-Genre Natural Language Inference (MultiNLI) dataset to 15 languages, including low-resource languages like Swahili and Urdu. The goal is to evaluate cross-lingual sentence understanding and provide a standardized benchmark for research in this area. The authors describe the construction of XNLI, which involves translating the MultiNLI development and test sets into the target languages using professional translators. They also present several baselines for multilingual sentence understanding, including machine translation systems and parallel data-based approaches. The results show that directly translating the test data yields the best performance among the baselines, but the alignment mechanism used in the multilingual sentence encoder approaches is very competitive. The paper discusses the limitations and future directions of the XNLI benchmark, emphasizing the need for further work to match the performance of translation-based methods.This paper introduces the Cross-lingual Natural Language Inference (XNLI) corpus, an extension of the Multi-Genre Natural Language Inference (MultiNLI) dataset to 15 languages, including low-resource languages like Swahili and Urdu. The goal is to evaluate cross-lingual sentence understanding and provide a standardized benchmark for research in this area. The authors describe the construction of XNLI, which involves translating the MultiNLI development and test sets into the target languages using professional translators. They also present several baselines for multilingual sentence understanding, including machine translation systems and parallel data-based approaches. The results show that directly translating the test data yields the best performance among the baselines, but the alignment mechanism used in the multilingual sentence encoder approaches is very competitive. The paper discusses the limitations and future directions of the XNLI benchmark, emphasizing the need for further work to match the performance of translation-based methods.