Edoardo Fusa: [email protected]
Alberto Luise: [email protected]
Angelo Quarta: [email protected]
The reports of the two assignments are available in the repository.
In studies of Machine Learning, especially of Natural Language Processing, the presented models are often backed up by huge datasets, sometimes spanning multiple languages. However, it’s not often that the behaviour of a relatively small model is analyzed in a limited and controlled environment. In this report we’ll tackle the classification task of POS-tagging on a specific dataset: we’ll try to train multiple models with a small depth on a fixed set of sentences related to the business world. We will see what are the main challenges of this tasks, what techniques could be employed in order to overcome them and what results can be obtained.
Human values detection has been a very important task regarding the field of debating AIs, enhanced by the organizers of scientific events called Touché. The objective of this task is the classification of values reported by an argument in a multi-label fashion. Such problem is approachable in several ways but we adopted the most common one in literature consisting of a BERT network as encoder followed by a fully connected classifier. Although the chosen baselines achieved a challenging score, said approach led to promising results and highlighted how the given dataset affects the performance of the designed model.