Hierarchical Attention for Conversational Agents

The argument of the thesis focuses on Knowledge-Enhanced Conversational Agents. In particular, it focuses on the implementation of a specific type of Recurrent Neural Network, Long Short-Term Memory (LSTM), to leverage the temporal dependencies of dialogue turns for extracting knowledge from a knowledge base. The thesis investigates the use of transformers for encoding multimodal language content and exploits the hierarchical structure of the knowledge base by creating three downstream tasks. These tasks are aimed at recognizing the domain, the entities involved, and the documents referenced in the user's request. The experiments are conducted with DSTC11 Track 5, which is a de facto standard for developing conversational agents.