A comparative study on question-worthy sentence selection strategies for educational question generation

Abstract

Automatic question generation, which aims at converting sentences in an article to high-quality questions, is an important task for educational practices. Recent work mainly focuses on designing effective generation architectures based on deep neural networks. However, the first and possibly the foremost step of automatic question generation has largely been ignored, i.e., identifying sentences carrying important information or knowledge that is worth asking questions about. In this work, we (i) propose a total of 9 strategies, which are grounded on heuristic question-asking assumptions, to determine sentences that are question-worthy, and (ii) compare their performance on 4 datasets by using the identified sentences as input for a well-trained question generator. Through extensive experiments, we show that (i) LexRank, a stochastic graph-based method for selecting important sentences from articles, gives robust performance across all datasets, (ii) questions collected in educational settings feature a more diverse set of source sentences than those obtained in non-educational settings, and (iii) more research efforts are needed to further improve the design of educational question generation architectures.