Seed and Grow: Augmenting Statistically Generated Summary Sentences using Schematic Word Patterns

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 543–552, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Seed and Grow: Augmenting Statistically Generated Summary Sentences using Schematic Word Patterns Stephen Wan†‡ Robert Dale† Mark Dras† †Centre for Language Technology Department of Computing Macquarie University Sydney, NSW 2113 swan,madras,rdale@ics.mq.edu.au C´ecile Paris‡ ‡ICT Centre CSIRO Sydney, Australia Cecile.Paris@csiro.au Abstract We examine the problem of content selection in statistical novel sentence generation. Our approach models the processes performed by professional editors when incorporating ma- terial from additional sentences to support some initially chosen key summary sentence, a process we refer to as Sentence Augmen- tation. We propose and evaluate a method called “Seed and Grow” for selecting such auxiliary information. Additionally, we argue that this can be performed using schemata, as represented by word-pair co-occurrences, and demonstrate its use in statistical summary sen- tence generation. Evaluation results are sup- portive, indicating that a schemata model sig- nificantly improves over the baseline. 1 Introduction In the context of automatic text summarisation, we examine the problem of statistical novel sentence generation, with the aim of moving from the current state-of-the-art of sentence extraction to abstract- like summaries. In particular, we focus on the task of selecting content to include within a generated sen- tence. Our approach to novel sentence generation is to model the processes underlying summarisation as performed by professional editors and abstractors. An example of the target output of this kind of gen- eration is presented in Figure 1. In this example, the human authored summary sentence was taken verba- tim from the executive summary of a United Nations proposal for the provision of aid addressing a partic- ular humanitarian crisis. Such documents typically exceed a hundred pages. Human-Authored Summary Sentence: Repeated [poor seasonal rains]1 [in 2004]2, culminating in [food insecurity]3, indicate [another year]4 of crisis, the scale of which is larger than last year’s and is further [exacerbated by diminishing coping assets]5 [in both rural and urban areas]6. Key Source Sentence: The consequences of [another year]4 of [poor rains]1 on [food security]3 are severe. Auxiliary Source Sentence(s): However in addition to the needs of economic recovery activities for IDPs, [food insecurity]3 [over the major- ity of 2004]2 [has created great stress]5 on the poorest families in the country, [both within the urban and rural settings]6. Figure 1: Alignment of a summary sentence to sentences in the full document. Phrases of similar meaning are co- indexed. To write such summaries, we assume that the hu- man abstractor begins by choosing key sentences from the full document. Then, for each key sen- tence, a set of auxiliary material is identified. The key sentence is revised incorporating these auxil- iary sentences to produce the eventual summary sen- tence. To study this phenomenon, a corpus of UN docu- ments was collected and analysed.1 Each document was divided into two parts comprising its executive summary, and the remainder, referred to here as the source. We manually aligned each executive sum- mary sentence …

Content truncated to 500 words