Towards Domain-Independent Text Structuring Trainable on Large Discourse Treebanks

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3141–3152 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 3141 Towards Domain-Independent Text Structuring Trainable on Large Discourse Treebanks Grigorii Guz and Giuseppe Carenini Department of Computer Science University of British Columbia Vancouver, BC, Canada, V6T 1Z4 {gguz, carenini}@cs.ubc.ca Abstract Text structuring is a fundamental step in NLG, especially when generating multi-sentential text. With the goal of fostering more general and data-driven approaches to text structuring, we propose the new and domain-independent NLG task of structuring and ordering a (pos- sibly large) set of EDUs. We then present a solution for this task that combines neural de- pendency tree induction with pointer networks and can be trained on large discourse tree- banks that have only recently become avail- able. Further, we propose a new evaluation metric that is arguably more suitable for our new task compared to existing content order- ing metrics. Finally, we empirically show that our approach outperforms competitive alterna- tives on the proposed measure and is equiva- lent in performance with respect to previously established measures. 1 Introduction Natural Language Generation (NLG) plays a funda- mental role in data-to-text tasks like automatically producing soccer, weather and financial reports (Chen and Mooney, 2008; Plachouras et al., 2016; Balakrishnan et al., 2019), as well as in text-to-text generation tasks like summarization (Nenkova and McKeown, 2012). Generally speaking, NLG involves three key steps (Gatt and Krahmer, 2017): first there is con- tent determination which selects what informa- tion units should be conveyed, secondly there is text structuring, which is responsible for properly structuring and ordering those units; and finally microplanning-realization that aggregates informa- tion units into sentences and paragraphs that are then verbalized. The focus of this paper is on the text structuring step, which is critical for the overall performance of an NLG system, as it ensures that the commu- nicative goals of the text are realized in the most structurally coherent and cohesive way possible, making the main ideas expressed by the text easy to follow for the target audience. Aiming to develop very general computational methods for text structuring, we keep our study in- dependent from particular ways in which the input information units are represented and from explic- itly provided ordering constraints for the target ap- plication domain (Gatt and Krahmer, 2017). More specifically, we propose and attack, in a fully data- driven way, the novel and domain-independent task of simultaneously structuring and ordering a set of Elementary Discourse Units (EDUs), i.e., clause- like text fragments that the Rhetorical Structure Theory (RST) (Mann and Thompson, 1988) as- sumes to be the building blocks of any discourse structure (see Figure 1(a)(left)). In other words, we assume that the system is given a set of EDUs (with cardinality possibly > 100) as input and re- turns their ordering, as well as the unlabelled RST dependency discourse tree structure for a document consisting of this set of EDUs, as illustrated in Fig- ure 1(a). Our data-driven approach …

Content truncated to 500 words