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

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Artificial Intelligence

Background:

• Question generation is crucial for AI, requiring models to understand language meaning and structure.
• Evaluating neural networks for question generation necessitates specialized datasets.

Purpose of the Study:

• Introduce the first openly-available German QUestion-Answer Congruence Corpus (QUACC).
• Establish baselines for sentence-based question generation using the QUACC corpus.
• Analyze the performance and error sources of various neural models for question generation.

Main Methods:

• Corpus creation: Developed a German question-answer congruence dataset.
• Baseline establishment: Implemented and compared diverse question generation systems.
• Model evaluation: Assessed neural network performance on the QUACC corpus.

Main Results:

• The QUACC corpus facilitates the investigation of neural models for question generation.
• Comparison of different systems highlights strengths and weaknesses in generating congruent questions.
• Insights into specific error sources for current neural network architectures were gained.

Conclusions:

• The QUACC corpus is a valuable resource for advancing research in question generation.
• The established baselines provide a benchmark for future work in this area.
• Further research can leverage QUACC to improve the capabilities of AI in understanding and generating language.