Bottom Up Parsing via Sequence Labeling

Abstract

We translate the sequence labeling framework, first introduced for top-down discourse parsing by Koto et al. (2021), to bottom-up discourse parsing. We introduce a novel parser that is not constrained by parsing direction (left-to-right or otherwise), and is conditioned on previous parsing decisions. We describe the unique training requirements of a (directionally) unconstrained parser and explore two different training procedures. Additionally, we introduce a novel dynamic oracle for unconstrained bottom-up parsing. Our proposed parser achieves state-of-the-art performance amongst bottom-up RST parsers.

Tasks

Reproductions