Does syntax help discourse segmentation? Not so much

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2432–2442 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Does syntax help discourse segmentation? Not so much Chlo´e Braud and Oph´elie Lacroix and Anders Søgaard CoAStaL DIKU University of Copenhagen University Park 5, 2100 Copenhagen chloe.braud@gmail.com lacroix@di.ku.dk soegaard@di.ku.dk Abstract Discourse segmentation is the first step in building discourse parsers. Most work on discourse segmentation does not scale to real-world discourse parsing across lan- guages, for two reasons: (i) models rely on constituent trees, and (ii) experiments have relied on gold standard identification of sentence and token boundaries. We therefore investigate to what extent con- stituents can be replaced with universal de- pendencies, or left out completely, as well as how state-of-the-art segmenters fare in the absence of sentence boundaries. Our results show that dependency information is less useful than expected, but we pro- vide a fully scalable, robust model that only relies on part-of-speech information, and show that it performs well across lan- guages in the absence of any gold-standard annotation. 1 Introduction Discourse segmentation is the task of identify- ing, in a document, the minimal units of text – called Elementary Discourse Units (EDU) (Carl- son et al., 2001) – that will be then linked by semantico-pragmatic relations – called discourse relations. Discourse segmentation is the first step when building a discourse parser, and has a large impact on the building of the final structure – predicted segmentation leads to a drop in perfor- mance of about 12-14% (Joty et al., 2015). In this work, we focus on the Rhetorical Struc- ture Theory (RST) (Mann and Thompson, 1988) in which discourse analysis is a tree covering an entire document. Most of the recent discourse parsers have been developed within this frame- work, making crucial the development of robust RST discourse segmenters. Many corpora have been annotated within this framework for several domains and languages – such as English with the RST Discourse Treebank (RST-DT) (Carlson et al., 2001), but also Spanish (da Cunha et al., 2011), Brazilian Portuguese (Cardoso et al., 2011; Collovini et al., 2007; Pardo and Seno, 2005; Pardo and Nunes, 2004) or German (Stede and Neumann, 2014). State-of-the-art performance for discourse seg- mentation on the RST-DT is about 94% in F1 (Xuan Bach et al., 2012). Most work on dis- course parsing has focused on English and on the RST-DT (Ji and Eisenstein, 2014; Feng and Hirst, 2014; Li et al., 2014; Joty et al., 2013), and so discourse segmentation (Xuan Bach et al., 2012; Fisher and Roark, 2007; Subba and Di Eugenio, 2007). And while discourse parsing is a document level task, discourse segmentation is done at the sentence level, assuming that sentence boundaries are known. This prevents from using discourse in- formation for a wider range of downstream tasks. Moreover, while discourse parsing is a seman- tic task involving a large range of information, the annotation guidelines reflect that segmentation is merely based on syntax: in practice, an EDU …

Content truncated to 500 words