Modelling Compression with Discourse Constraints

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 1–11, Prague, June 2007. c⃝2007 Association for Computational Linguistics Modelling Compression with Discourse Constraints James Clarke and Mirella Lapata School of Informatics, University of Edinburgh 2 Bucclecuch Place, Edinburgh EH8 9LW, UK jclarke@ed.ac.uk, mlap@inf.ed.ac.uk Abstract Sentence compression holds promise for many applications ranging from summarisa- tion to subtitle generation and subtitle gen- eration. The task is typically performed on isolated sentences without taking the sur- rounding context into account, even though most applications would operate over entire documents. In this paper we present a dis- course informed model which is capable of producing document compressions that are coherent and informative. Our model is in- spired by theories of local coherence and formulated within the framework of Integer Linear Programming. Experimental results show significant improvements over a state- of-the-art discourse agnostic approach. 1 Introduction The computational treatment of sentence compres- sion has recently attracted much attention in the literature. The task can be viewed as producing a summary of a single sentence that retains the most important information and remains grammatically correct (Jing 2000). Sentence compression is com- monly expressed as a word deletion problem: given an input sentence of words W = w1,w2,...,wn, the aim is to produce a compression by removing any subset of these words (Knight and Marcu 2002). Sentence compression can potentially benefit many applications. For example, in summarisation, a compression mechanism could improve the con- ciseness of the generated summaries (Jing 2000; Lin 2003). Sentence compression could be also used to automatically generate subtitles for tele- vision programs; the transcripts cannot usually be used verbatim due to the rate of speech being too high (Vandeghinste and Pan 2004). Other applica- tions include compressing text to be displayed on small screens (Corston-Oliver 2001) such as mobile phones or PDAs, and producing audio scanning de- vices for the blind (Grefenstette 1998). Most work to date has focused on a rather sim- ple formulation of sentence compression that does not allow any rewriting operations, besides word re- moval. Moreover, compression is performed on iso- lated sentences without taking into account their sur- rounding context. An advantage of this simple view is that it renders sentence compression amenable to a variety of learning paradigms ranging from in- stantiations of the noisy-channel model (Galley and McKeown 2007; Knight and Marcu 2002; Turner and Charniak 2005) to Integer Linear Programming (Clarke and Lapata 2006a) and large-margin online learning (McDonald 2006). In this paper we take a closer look at one of the simplifications associated with the compression task, namely that sentence reduction can be realised in isolation without making use of discourse-level information. This is clearly not true — professional abstracters often rely on contextual cues while creat- ing summaries (Endres-Niggemeyer 1998). Further- more, determining what information is important in a sentence is influenced by a variety of contextual factors such as the discourse topic, whether the sen- tence introduces new entities or …

Content truncated to 500 words