Fit it in but say it well!

Fit it in but say it well! Cécile Paris, Nathalie Colineau, Andrew Lampert CSIRO – ICT Centre Locked Bag 17 North Ryde, NSW 1670, Australia FirstName.LastName@csiro.au Joan Giralt Duran Barcelona School of Informatics Technical University of Catalonia Barcelona, Spain joangi@gmail.com Abstract Reasoning about how much content to generate when space is limited presents an increasingly important challenge for generation systems, as the diversity of potential delivery channels continues to grow. The problem is multi-facetted: the generated text must fit into the allocated space, the space available must be well utilised, and the resulting text must convey its intended message, and be coherent, well structured and balanced. To address this problem, we use a discourse planning approach. Our system reasons about the discourse structure to decide how much content to realise. In this paper, we present two algorithms that perform this reasoning and analyse their effectiveness. 1 Introduction The ability to reason about how much content to realise in order to convey a message when the allocated space is fixed is an important consideration for Natural Language Generation (NLG) systems. It will become even more pressing as the amount of available information increases (e.g., via the web or content management systems) and the space constraints on the delivery media become more diverse (e.g., via web browsers, email, PDAs, cell phones). We, as humans, address this problem by shorten- ing our sentences or by restricting the content we include. We can achieve the former by manipulating vocabulary and syntax. This, however, is of limited value in reclaiming significant amounts of space and requires careful attention to sentence-level grammar and vocabulary choice. We can achieve the latter by dropping those pieces of content whose contribution to the communicative goal is most limited. For instance, we might question the need for a long elaboration or an example to illustrate our point. This approach can reclaim significant amounts of space but requires an understanding of the text’s discourse structure. Many organisations today store information in content management systems or other types of organisational databases. This information is typically stored at the level of paragraphs, along with images and database entries (e.g., directories of contact information). Such information systems provide an interesting opportunity for NLG systems, by reducing the cost of acquiring the underlying semantic knowledge base and removing the need to construct text from first principles. However, they also present challenges: how can a system still perform the type of reasoning typically expected of NLG systems when it has no control over the text at the sentence level? In particular, how can a system produce a coherent and well structured text that meets some specific space limitations? In these cases, a system only has one method available to ensure it produces an appropriate amount of text: it must reason about the text at the discourse level. Many of our application domains are such that information is stored at the paragraph level in one or more data repositories. Our system answers people’s information needs by retrieving …

Content truncated to 500 words