HanaNLG: A Flexible Hybrid Approach for Natural Language Generation

Abstract

Nowadays, with advances in digital technologies, interaction between computers and humans is essential. In this regard, the area of Natural Language Generation (NLG) can provide techniques capable of facilitating and improving this type of interaction. However, the existing approaches to this field are usually developed ad-hoc for specific tasks, purposes and domains, which hinders the advancement of flexible and adaptable multi-domain NLG systems. Under these premises, the objective of this paper is to present HanaNLG, a hybrid generic NLG approach, focused on the surface realisation stage. HanaNLG combines statistic and knowledge-based techniques and is able to generate text independently of the domain. In particular, this is done by exploiting language models in conjunction with semantic knowledge, providing flexibility to the whole generation process, thus, minimising the high cost associated with the development of common elements involved in NLG, such as grammars. Therefore, taking into account this joint perspective, our approach contributes to advancing the NLG field by providing greater flexibility when it comes to (i) producing text for different domains, and (ii) increasing the variety of vocabulary to appear in the generated text. In order to assess the effectiveness of HanaNLG, it was tested in two domains: (i) NLG for assistive technologies and, (ii) NLG for creating opinionated sentences. The positive results obtained (almost the 99% of the generated sentences for both domains are original and well constructed) show that our approach is capable of generating text for different domains. More importantly, the combination of language models with semantic knowledge enhances the quality of the generated text, thereby improving the results obtained compared to other methods that only rely on statistical methods.