• Graphical Models
• Hidden Markov Tree
• Tree Structured Data
• Tree Transducer

In this thesis we present a novel machine learning model for learning probabilistic transductions, that are a generalization of the supervised learning setting where both input and output information can be complex structures with varying topologies. The approach put forward in the thesis is generic and founds on a very general assumption concerning information exchange between input and output data. In this work, we show some examples of how such bridging information can be represented, analyzing how different choices can lead to consistently different performances depending on the task at hand. The thesis focuses on tree structured data, realizing an adaptive tree-to-tree transducer. To this end, we define probabilistic encoder which aim is to adaptively encode structural knowledge from the input tree into a compressed vectorial representation and a decoder whose aim is to exploit the encoded information to generate an output tree. Both models are based on the Hidden Markov Tree Model, that is widely used in literature to handle tree structured data. Moreover, we illustrate the learning procedure that has been derived for the end-to-end training of the encoder and decoder models. Finally, we assess the performance of our probabilistic tree transducer on different datasets, analysing both its predictive perfomance as well as its computational requirements.