Geolocazione Automatica per gli articoli di Wikipedia: un approccio basato su Reti Neurali Convoluzionali e strutture a grafo.

The present work focuses on automatic geolocation of documents and, in particular, of Wikipedia articles. Document geolocation falls into the category of georeferencing tasks and involves locating documents in geographical space. In particular, our task is formalized as a classification problem over a discrete set of geographical areas. Motivated by the advancements of deep learning in the field of Natural Language Processing and Information Retrieval, we developed and compared multiple location prediction models highly based on neural networks architectures, each model leveraging different types of information: the textual content of each article, the link structure between articles, or both. Finally, considering the sparsity of geotagged articles and the constant growth of the Wikipedia ecosystem, we built models that are both capable of leveraging unlabelled data and, when possible, generalizing to articles not seen at training time.