*Apologies for cross-postings*
�
eRST – enhanced Rhetorical Structure Theory
�
We are delighted to introduce a new parsing framework and datasets for discourse relation recognition: eRST is an ehanced version of Rhetorical Structure Theory which allows multiple, concurrent and non-projective discourse relations in a formally constrained graph, aligned to a large inventory of discourse relation signals, based on the Signaling Corpus taxonomy. Signals are divided into 9 classes and 45 sub-classes, including traditional discourse markers such as PDTB-style connectives, but also lexical, syntactic and semantic signals, such as repetition, lexical chains and anaphoric relations.
�
eRST is described in depth in this paper:
�
Zeldes, Amir, Aoyama, Tatsuya, Liu, Yang Janet, Peng, Siyao, Das, Debopam and Gessler, Luke (2024) "eRST: A Signaled Graph Theory of Discourse Relations and Organization". Computational Linguistics, 1–47. https://direct.mit.edu/coli/article/doi/10.1162/coli_a_00538/124464/eRST-A-S...
�
You can also find an overview at the following website, as well as analyses for nearly 250K words of English in 24 spoken and written text types, from the freely available UD English GUM and GENTLE corpora:
�
�
If you want to learn more and are participating in EMNLP in Miami this week, please check out our talk on Wednesday! And if you are interested in shallow discourse parsing, please check out our paper on Tuesday and the aligned PDTB3-style relations for the same data in this paper: https://aclanthology.org/2024.emnlp-main.684/
�