WorldTree: A Corpus of Explanation Graphs for Elementary Science\n Questions supporting Multi-Hop Inference

Abstract

Developing methods of automated inference that are able to provide users with\ncompelling human-readable justifications for why the answer to a question is\ncorrect is critical for domains such as science and medicine, where user trust\nand detecting costly errors are limiting factors to adoption. One of the\ncentral barriers to training question answering models on explainable inference\ntasks is the lack of gold explanations to serve as training data. In this paper\nwe present a corpus of explanations for standardized science exams, a recent\nchallenge task for question answering. We manually construct a corpus of\ndetailed explanations for nearly all publicly available standardized elementary\nscience question (approximately 1,680 3rd through 5th grade questions) and\nrepresent these as "explanation graphs" -- sets of lexically overlapping\nsentences that describe how to arrive at the correct answer to a question\nthrough a combination of domain and world knowledge. We also provide an\nexplanation-centered tablestore, a collection of semi-structured tables that\ncontain the knowledge to construct these elementary science explanations.\nTogether, these two knowledge resources map out a substantial portion of the\nknowledge required for answering and explaining elementary science exams, and\nprovide both structured and free-text training data for the explainable\ninference task.\n