Generating Elliptic Coordination

In this paper, we focus on the task of generating elliptic sentences. We extract from the data provided by the Surface Realisation (SR) Task (Belz et al., 2011) 2398 input whose corresponding output sentence contain an ellipsis. We show that 9% of the data contains an ellipsis and that both coverage and BLEU score markedly decrease for elliptic input (from 82.3% coverage for non-elliptic sentences to 65.3% for elliptic sentences and from 0.60 BLEU score to 0.47). We argue that elided material should be represented using phonetically empty nodes and we introduce a set of rewrite rules which permits adding these empty categories to the SR data. Finally, we evaluate an existing surface realiser on the resulting dataset. We show that, after rewriting, the generator achieves a coverage of 76% and a BLEU score of 0.74 on the elliptical data.