Training conversational agents to understand complex dialogues

Nowadays, conversational agents are inspiring the academic and non-academic world thanks to the engaging interaction they establish with the user. However, finding valuable data to train a system able to converse as human-like as possible is not a trivial task. This is even more challenging for the Italian language, where only a few dialogic datasets are available. This thesis expressly addresses this challenge, proposing JILDA (Job Interview Labelled Dialogues Assembly), a new Italian dialogue dataset for the job-offer domain, and demonstrating its practical application for the training of a conversational agent able to understand syntactically and semantically complex data. JILDA dialogues, after being annotated via MATILDA, a new annotation tool developed in collaboration with Wluper, are used to train the Natural Language Understanding module of a conversational agent, as this is an essential component of any dialogue system. Three of the most recent pretrained LMs are benchmarked: Italian BERT, Multilingual BERT, and AlBERTo. Analysing the performance obtained, it was developed JILDA 2.0, an updated version of the resource useful to realise a first step in improving NLU for Italian dialogues. Finally, this thesis frames the research topic within a global ethical framework, considering the ethical issues which emerge in human-machine interaction, the gender biases embedded in the Embodied Conversational Agents (ECAs) and their impacts on modern society.