Textual entailment based question generation

Abstract

This paper proposes a novel question generation (QG) approach based on textual entailment. Many previous QG studies transform a single sentence into a question directly. They need hand-crafted templates or generate simple questions similar to the source texts. As a novel approach to QG, this research employs two-step QG: 1) generating new texts entailed by source documents, and 2) transforming the entailed sentences into questions. This process can generate questions that need the understanding of textual entailment to solve. Our system collected 1,367 English Wikipedia sentences as QG source, retrieved 647 entailed sentences from the web, and transformed them into questions. The evaluation result showed that our system successfully generated nontrivial questions based on textual entailment with 53% accuracy.