Surface Realisation Using Factored Language Models and Input Seed Features

Natural Language Generation research field needs to move forward to the design and development of flexible and adaptive techniques and approaches capable of producing language automatically, for any domain, language and purpose. In light of this, the aim of this paper is to study the appropriateness of factored language models for the stage of surface realisation, thus presenting an almost-fully language independent statistical approach. Its main novelty is that it can be adapted to generate texts for different purposes or domains thanks to the use of an input seed feature that guides all the generation process. In the context of this research, the seed input is a phoneme and our goal is to generate a full meaningful sentence that maximises the amount of words containing that phoneme. We experimented with different factors, including lemmas or part-of-speech tags, based on a trigram language model. The analysis carried out with several configurations of our proposed approach showed an improvement of 47% and 40% as far as the total meaningful generated sentences is concerned, with respect to traditional language models, for English and Spanish, respectively.