Top-Down RST Parsing Utilizing Granularity Levels in Documents

Some downstream NLP tasks exploit discourse dependency trees converted from RST trees. To obtain better discourse dependency trees, we need to improve the accuracy of RST trees at the upper parts of the structures. Thus, we propose a novel neural top-down RST parsing method. Then, we exploit three levels of granularity in a document, paragraphs, sentences and Elementary Discourse Units (EDUs), to parse a document accurately and efficiently. The parsing is done in a top-down manner for each granularity level, by recursively splitting a larger text span into two smaller ones while predicting nuclearity and relation labels for the divided spans. The results on the RST-DT corpus show that our method achieved the state-of-the-art results, 87.0 unlabeled span score, 74.6 nuclearity labeled span score, and the comparable result with the state-of-the-art, 60.0 relation labeled span score. Furthermore, discourse dependency trees converted from our RST trees also achieved the state-of-the-art results, 64.9 unlabeled attachment score and 48.5 labeled attachment score. Introduction A discourse structure of a document can be represented as a tree, like a syntactic structure of a sentence being represented as a tree. Rhetorical structure theory (RST) (Mann and Thompson 1987) is well-known and has been widely studied for representing a document as a tree. An RST tree is a type of constituent tree, whose terminal nodes (leaves) are elementary discourse units (EDUs), clause-like units, and whose non-terminal nodes represent the nuclearity status, nucleus or satellite, for the span that consists of a sequence of EDUs or a single EDU. The span dominated by a nucleus is more essential than the one dominated by a satellite. That is, the satellite has a role of supporting the nucleus. Furthermore, rhetorical relations are defined between two adjacent spans. A mono-nuclear relation, such as ”Elaboration” or ”Condition”, is assigned between a nucleus and its satellite, and a multi-nuclear relation, such as ”Same-unit” or ”Topicchange”, is assigned between two nuclei. RST trees have important roles in natural language processing (NLP) tasks, such as summarization (Marcu 1998; Copyright c © 2020, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. Elaboration