Discourse Analysis via Questions and Answers: Parsing Dependency Structures of Questions Under Discussion
Wei-Jen Ko,Yating Wu,Cutter Dalton,Dananjay Srinivas,Greg Durrett,Junyi Jessy Li
DOI: https://doi.org/10.48550/arXiv.2210.05905
2022-10-12
Computation and Language
Abstract:Automatic discourse processing, which can help understand how sentences connect to each other, is bottlenecked by data: current discourse formalisms pose highly demanding annotation tasks involving large taxonomies of discourse relations, making them inaccessible to lay annotators. This work instead adopts the linguistic framework of Questions Under Discussion (QUD) for discourse analysis and seeks to derive QUD structures automatically. QUD views each sentence as an answer to a question triggered in prior context; thus, we characterize relationships between sentences as free-form questions, in contrast to exhaustive fine-grained taxonomies. We develop the first-of-its-kind QUD parser that derives a dependency structure of questions over full documents, trained using a large question-answering dataset DCQA annotated in a manner consistent with the QUD framework. Importantly, data collection is easily crowdsourced using DCQA's paradigm. We show that this leads to a parser attaining strong performance according to human evaluation. We illustrate how our QUD structure is distinct from RST trees, and demonstrate the utility of QUD analysis in the context of document simplification. Our findings show that QUD parsing is an appealing alternative for automatic discourse processing.