Improving Neural Question Generation using World Knowledge

TL;DR

This work tackles neural question generation by injecting world knowledge into the encoder via two sources: linked entities from Wikipedia and fine-grained entity types. The authors implement a world-knowledge enriched Bi-LSTM encoder and an attention-based decoder with a copy mechanism, leveraging entity linking and FGET to produce more human-like questions. Experiments on SQuAD and MS MARCO show consistent BLEU-4 improvements over a vanilla seq2seq QG model, with pretrained linker and FGET features delivering the strongest gains. The study highlights the potential of explicit world-knowledge signals to enhance QG performance and outlines future work including human evaluation and exploring additional knowledge sources and entity relations.

Abstract

In this paper, we propose a method for incorporating world knowledge (linked entities and fine-grained entity types) into a neural question generation model. This world knowledge helps to encode additional information related to the entities present in the passage required to generate human-like questions. We evaluate our models on both SQuAD and MS MARCO to demonstrate the usefulness of the world knowledge features. The proposed world knowledge enriched question generation model is able to outperform the vanilla neural question generation model by 1.37 and 1.59 absolute BLEU 4 score on SQuAD and MS MARCO test dataset respectively.