A large annotated corpus for learning natural language inference

This paper introduces the Stanford Natural Language Inference (SNLI) corpus to address a bottleneck in the study of semantic representations: the scarcity of large-scale data for reasoning about entailment and contradiction. The corpus is a freely available collection of labeled sentence pairs written by humans performing a novel grounded task based on image captioning. At 570K pairs, it is two orders of magnitude larger than all prior resources of its type.

This increase in scale changed what models could achieve on natural language inference: it allowed lexicalized classifiers to outperform some sophisticated existing entailment models, and, for the first time, allowed a neural network-based model to perform competitively on inference benchmarks. By pairing large scale with human-authored labels, SNLI removed the resource limitation that had dramatically constrained machine learning research on entailment and contradiction.