CODRA: A Novel Discriminative Framework for Rhetorical Analysis

Abstract

Clauses and sentences rarely stand on their own in an actual discourse; rather, the relationship between them carries important information that allows the discourse to express a meaning as a whole beyond the sum of its individual parts. Rhetorical analysis seeks to uncover this coherence structure. In this article, we present CODRA— a COmplete probabilistic Discriminative framework for performing Rhetorical Analysis in accordance with Rhetorical Structure Theory, which posits a tree representation of a discourse. CODRA comprises a discourse segmenter and a discourse parser. First, the discourse segmenter, which is based on a binary classifier, identifies the elementary discourse units in a given text. Then the discourse parser builds a discourse tree by applying an optimal parsing algorithm to probabilities inferred from two Conditional Random Fields: one for intra-sentential parsing and the other for multi-sentential parsing. We present two approaches to combine these two stages of parsing effectively. By conducting a series of empirical evaluations over two different data sets, we demonstrate that CODRA significantly outperforms the state-of-the-art, often by a wide margin. We also show that a reranking of the k-best parse hypotheses generated by CODRA can potentially improve the accuracy even further.

Does syntax help discourse segmentation? Not so much

“…"but", "because", "after") and syntactic information (Le Thanh et al, 2004;Tofiloski et al, 2009 More recent discourse segmenters on the RST-DT are based on binary classifiers at the word level (Soricut and Marcu, 2003;Fisher and Roark, 2007;Joty et al, 2015), possibly using a neural network architecture (Subba and Di Eugenio, 2007). Joty et al (2015) also report results for the Instructional corpus (Subba and Di Eugenio, 2009) (F 1 80.9% on 10-fold).…”

“…Discourse segmentation is the first step when building a discourse parser, and has a large impact on the building of the final structurepredicted segmentation leads to a drop in performance of about 12-14% (Joty et al, 2015).…”

Integer Linear Programming for Discourse Parsing

“…Feng and Hirst (2012) extend this approach by additional feature engineering but is restricted to sentence-level parsing. Joty et al (2013) and Joty et al (2015) present a textlevel discourse parser that uses Conditional Random Fields to capture label inter-dependencies and chart parsing for decoding and have the best results on non-dependency based discourse parsing, with an F1 of 0.689 on unlabelled structures and 0.5587 on labelled structures.…”

The rhetorical nature of rhythm

“…RST-style discourse segmentation represents the segmentation of a text into the smallest units between which rhetorical relations (such as elaboration, cause, contrast) hold. So far, this task was most successfully completed using statistical methods, which modeled discourse segmentation as either a problem of classification [11][12][13][14], or of sequence labeling [15][16][17]. The methodology of discourse segmentation differs in whether it is centered on individual tokens (words), or rather on pairs of tokens.…”