Data: dijous 5 de desembre del 2019

Hora: 12.00 h

Lloc:  Sala de reunions 52.701 - 7a planta edifici 52. Roc Boronat - Campus del Poblenou - UPF

Resum: We've been crowdsourcing a dataset of 'evoked questions', by letting naive participants read snippets of text and asking them which question the text evoked for them. After entering a question they were shown how the text continued and asked to indicate whether their question had received an answer. This has resulted in a rich dataset of over 4000 questions, their answers where available, and the highlighted words in the text that triggered them. In this talk I will explain and motivate our elicitation method, discuss the relation we see between our evoked questions and the important theoretical notion of Question Under Discussion (QUD), and present a high-level overview of the resulting data, before zooming in on an exciting new line of research this dataset opens up: The source texts for which we elicited evoked questions were taken from existing datasets with discourse relation annotations, which enables us to study the ways in which evoked questions and discourse relations (mis)align. One concrete hypothesis we've found evidence for is that discourse relations are more often explicitly marked (with a connective such as 'because') in places where there is less agreement about what question the text evokes.