Analysing Positional Language Models for Natural Language Generation

The structure and selection of content in Automatic Language Generation is commonly referred to as Macroplanning. It provides an intermediate artefact that conveys such information, called document plan. This paper studies the appropriateness of Positional Language Models for building that plan, since they provide mechanisms to outline relevant elements and their distribution along the text. A series of experiments were conducted over a corpus of children tales in English, with the purpose of studying the behaviour and optimal adjustment of the method and its parameters. The results showcase the possibility of taking advantage from the different configurations to achieve several levels of complexity regarding the outcome of the generation process, the final text. In addition, the findings highlight the convenience of deeper semantic representations that would enrich the document plan and, thus, improve the outcome of an NLG system.