CORTEX: Generación consciente de texto
Based on the previous overview of the research context, the initial hypothesis of this project is as follows: The effective integration of world and external knowledge in NLG architectures improves the commonsense reasoning capabilities of NLG systems. We consider that enhancing commonsense reasoning capabilities of NLG systems is needed to automatically produce accurate, correct, and reliable texts that will be in line with real facts.
We propose a new generation of knowledge-enhanced NLG systems that overcome the “hallucination” phenomenon and avoid natural language generation that is “economical with the truth.” This is possible given that Transformers, a type of “end-to-end” architecture that predominates in text generation, can be fine-tuned, trained and adapted, so that: 1) they can learn to generalise and identify implicit information; and, 2) they can take into account certain characteristics that the generated text 3 de 20 must reflect, such as structure, length, style, formality, etc., thus allowing a more controlled and accurate generation. These types of adjustments would help, on the one hand, to generate texts in a more natural, diverse, and semantic way, and on the other, to detect strange patterns or biases that should be avoided in the generated text. In addition, the control of different aspects of the text is key for successfully applying and transferring NLG systems to real scenarios that are relevant to industry and society (Len et al., 2020).