This paper proposes an encoder-decoder structure, where the encoder represents the input span of words and EDUs using a hierarchical biLSTM. There are two level of sequences, first level is words contained in a EDU and second level is EDU sequences.

University of Michigan - Ann Arbor

It's the most commonly used model of discourse organization. In RST relations are defined between nucleus two spans of text, generally a nucleus and a satellite.

Rhetorical Structure theory (RST)

EDU segmentation extracts the start and end of each EDU. The output should be segmentation of sentences. Modern systems generally use neural sequence models supervised by the gold EDU segmentation in datasets like the RST Discourse Treebank. A classic method can be to predict whether there is a break for every token.

EDU Segmentation

Transition-based Neural RST Parsing with Implicit Syntax Features (Reference)

An on-going but a helpful book resource about NLP
https://web.stanford.edu/~jurafsky/slp3/

Speech and Language Processing (3rd ed. draft) 

Transition-based Neural RST Parsing with Implicit Syntax Features

It is standard to first transform the gold RST trees into right-branching binary trees, and to report four metrics: trees with no labels (S for Span), labeled with nuclei (N), with relations (R), or both (F for Full), for each metric computing micro-averaged F1 over all spans from all documents.

Learn Before

Related