Discourse Planning with an N-gram Model of Relations

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1973–1977, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Discourse Planning with an N-gram Model of Relations Or Biran Columbia University orb@cs.columbia.edu Kathleen McKeown Columbia University kathy@cs.columbia.edu Abstract While it has been established that transi- tions between discourse relations are im- portant for coherence, such information has not so far been used to aid in lan- guage generation. We introduce an ap- proach to discourse planning for concept- to-text generation systems which simul- taneously determines the order of mes- sages and the discourse relations between them. This approach makes it straightfor- ward to use statistical transition models, such as n-gram models of discourse re- lations learned from an annotated corpus. We show that using such a model signif- icantly improves the quality of the gener- ated text as judged by humans. 1 Introduction Discourse planning is a subtask of Natural Lan- guage Generation (NLG), concerned with deter- mining the ordering of messages in a document and the discourse relations that hold among them (Reiter and Dale, 2000). Early approaches to dis- course planning used manually written rules, often based on schemas (McKeown, 1985) or on Rhetor- ical Structure Theory (RST) (Mann and Thomp- son, 1987; Hovy, 1993; Power, 2000). In the past decade, various statistical approaches have emerged (Duboue and McKeown, 2001; Dimitro- manolaki and Androutsopoulos, 2003; Soricut and Marcu, 2006; Konstas and Lapata, 2013). Other relevant statistical approaches to content ordering can also be found in the summarization literature (Barzilay et al., 2001; Lapata, 2003; Bollegala et al., 2005). These approaches overwhelmingly focus on determining the best order of messages using semantic conent, while discourse relations are in most cases either determined by manually- written derivation rules or completely ignored. Meanwhile, researchers working on discourse relation disambiguation have observed that the se- quence of discourse relations itself, independently of content, helps in disambiguating adjacent re- lations (Wellner et al., 2006; Pitler et al., 2008). Sequential discourse information has been used successfully in discourse parsing (Ghosh et al., 2011; Feng and Hirst, 2014), and discourse struc- ture was shown to be as important for text co- herence as entity-based content structure (Lin et al., 2011; Feng et al., 2014). Surprisingly, so far, discourse sequential information from exist- ing discourse-annotated corpora, such as the Penn Discourse Treebank (PDTB) (Prasad et al., 2008) has not been used in generation. In this paper, we present an NLG framework that generates texts from existing semantic web ontologies. We use an n-gram model of discourse relations to perform discourse planning for these stories. Through a crowd-sourced human evalua- tion, we show that the ordering of our documents and the choice of discourse relations is signifi- cantly better when using this model. 2 Generation Framework In concept-to-text generation pipelines, discourse planning typically occurs after the content selec- tion stage. The input, therefore, is an unordered set of messages that are not yet realized: instead of being represented as text, the messages …

Content truncated to 500 words