Statistical natural language generation from tabular non-textual data

Abstract

Most of the existing natural language generation (NLG) techniques employing statistical methods are typically resource and time intensive. On the other hand, handcrafted rulebased and template-based NLG systems typically require significant human/designer efforts. In this paper, we proposed a statistical NLG technique which does not require any semantic relational knowledge and takes much less time to generate output text. The system can be used in those cases where source non-textual data are in the form of tuple in some tabular dataset. We carried out our experiments on the Prodigy-METEO wind forecasting dataset. For the evaluation purpose, we used both human evaluation and automatic evaluation. From the evaluation results we found that the linguistic quality and correctness of the texts generated by the system are better than many existing NLG systems.