Modelli di deep learning per il reperimento e l’ordinamento di documenti

Deep neural models revolutionized the research landscape in the Information Retrieval (IR) domain. Nowadays, they are employed at different stages in retrieval and ranking pipelines. For example, deep learning systems can automatically convert textual information into numerical features so that it can be later evaluated and compared by ranking models. Deep learning systems can also be effective in ranking items that have already been encoded into numerical features, i.e. comparing and ordering their representations according to a certain criterion. This rapid adoption of deep learning models in ranking pipelines however was conducted considering them as as black-box systems, without a proportionate understanding of their key components. For this reason, in this thesis we aim at filling this gap in understanding and assessing the importance of each the the building blocks of neural IR models and to later improve their performance in different application scenarios, from text-matching to ranking. Our contributions include an evaluation of the impact of probabilistic text representations in the text-matching task, the proposal of a new family of training paradigms for ranking models based on probability distributions, and a solution to improve the efficiency of neural reranking systems.