Exploring neural question generation for formal pragmatics: Data set and model evaluation.

Abstract

We provide the first openly-available German QUestion-Answer Congruence Corpus (QUACC), designed for the task of sentence-based question generation with question-answer congruence. Based on this corpus, we establish suitable baselines for question generation, comparing systems of very different nature. Question generation is an interesting challenge in particular for current neural network architectures given that it combines aspects of language meaning and forms in complex ways. The systems have to generate question phrases appropriately linking to the meaning of the envisaged answer phrases, and they have to learn to generate well-formed questions using the source. We show that our QUACC corpus is well-suited to investigate the performance of various neural models and gain insights about the specific error sources.