The work presented was developed during the internship, as researchers in the field of Natural Language Generation, at the Insid&s Lab laboratory in Milan-Bicocca. The work carried out deals with the creation of a framework for the correct assessment of the impact of the quality of the input datasets on the quality of the text generated by the NLG models, specifically: Creation of the "Concept-Based" and "Entity-Based" versions of the WebNLG dataset; Evaluation of the quality of the datasets created; Training of LSTM and Transformer models using the OpenNMT tool; Natural language text generation by LSTM and Transformer models; Evaluation of the quality of the text generated by the NLG models; Final analysis.
