Extracting parallel fragments from comparable corpora for data-to-text generation

Abstract

Building NLG systems, in particular statistical ones, requires parallel data (paired inputs and outputs) which do not generally occur naturally. In this paper, we investigate the idea of automatically extracting parallel resources for data-to-text generation from comparable corpora obtained from the Web. We describe our comparable corpus of data and texts relating to British hills and the techniques for extracting paired input/output fragments we have developed so far.