Generation of textual/video descriptions for technological products based on structured data

With the recent development of deep learning and Natural Language Generation (NLG) techniques in particular, machines have been given the ability to write and speak, overcoming the limitations of traditional template-based approaches. However, implementing sophisticated models for practical usage is definitely demanding, especially in business contexts in which the final output quality has a direct impact on the economic performance of the service. Our work deals with proposing a specific framework for data-to-text and text-to-speech generation, mainly based on a pre-trained language model and popular NLP tasks, as Named Entity Recognition. The ultimate objective is to generate detailed product descriptions autonomously, starting from the product sheet, in order to promote automation and reduce human effort. Rather than simply reporting product features, the system independently creates a coherent structure by interpreting the information it receives. In this way, it is able to formulate reasonable considerations and motivate possible benefits while trying to remain consistent with the source information. The obtained results present the tendency to properly reproduce the semantic, lexical and general linguistic style of the given context.