Natural Language Processing with Generative AI Models: A Methodological Approach for Their Application

In this thesis work, a methodological approach is proposed for application in Natural Language Processing (NLP) tasks exploiting the capabilities of latest Generative Artificial Intelligence (Generative AI) text-to-text models. With this methodology, it will be possible to delineate data features scope effectively, enhance model prompting based on expected input and output, and appropriately select the most suitable Generative model for the required NLP task. The dissertation begins with an analysis of the foundations of the subject, providing an overview of both NLP architectures, from Recurrent Neural Networks to today’s Transformer, and Generative AI state-of-the-art models from GoogleAI, MetaAI and OpenAI. Then the methodology is presented describing its phases, the different options to treat different features of data (such as data sensitivity and input formats) along with the generative model selection process and performance evaluation criteria. Next the methodology is applied to 10 different NLP tasks, from simple comprehension tasks to a difficult and temporally articulated assignment involving a real robotic agent such as Boston Dynamics’ SPOT. The effectiveness of the methodology is discussed, evaluating its performance, underlining strengths, weaknesses and possible ways to solve them. Finally, future developments and solutions, helpful in enhancing the linguistic comprehensions of Generative models, are presented.