Analysing the influence of semantic knowledge in natural language generation

This paper conducts an evaluation of several traditional Language Models for the task of Natural Language Generation, focusing on the surface realisation stage. Specifically, we analyse and compare the n-gram language models and the factored language models through the evaluation of automatically generated sentences. In this manner, factored language models have shown to be better, improving the coherence of the generated sentences compared with the ones generated with n-grams. Furthermore, factored language models were tested with different lexical and semantic knowledge, leading to more suitable generated sentences in the latter case as well as providing a greater expressive richness.