StaTIC: A Dataset of Question-Driven Inference Chains

Abstract

Modeling logical inference has become a key building block in the improvement of manyNLP tasks, including summarization, question answering and information extraction.However, since the rules that underlie inference are rarely made explicit in naturallanguage, there is a need for specialized datasets from which these rules can be learned.We introduce Stanford Textual Inference Chains (StaTIC), a dataset of sentence pairs inan entailment relation that are also “minimal pairs”, differing from each other by a smallsyntactic or lexical change. In this thesis, we describe and evaluate our data collectionmethods and analyze the lexical and syntactic properties of the results, focusing on theways in which they make the dataset suitable for informing systems of natural languageunderstanding.