Distant Supervision for Extractive Question Summarization

Questions are often lengthy and difficult to understand because they tend to contain peripheral information. Previous work relies on costly human-annotated data or question-title pairs. In this work, we propose a distant supervision framework that can train a question summarizer without annotation costs or question-title pairs, where sentences are automatically annotated by means of heuristic rules. The key idea is that a single-sentence question tends to have a summary-like property. We empirically show that our models trained on the framework perform competitively with respect to supervised models without the requirement of a costly human-annotated dataset.