Multi-Domain Conversational Agents based on Semantic Parsing and Retrieval of External Knowledge

Retrieval-enhanced conversational agents are able to access external sources and exploit them to generate better responses to a given input prompt. Different solutions based on this paradigm have already been proposed in the literature, but they have never been presented together with Abstract Meaning Representation (AMR) graphs. AMRs provide structured modeling of what is said in a document, regardless of its form: they can extract the concepts expressed in a sentence and have the potential to alleviate its ambiguity. This thesis introduces a conversational agent that relies on a combined use of retrieval technology and AMRs. In particular, a retriever identifies relevant documents from an external source, while AMRs are used to refine these selections. The model has been trained and tested on the MultiDoc2Dial dataset, which comprehends multi-domain dialogues and a knowledge base of a few thousand snippets. Our proposal demonstrates consistent performance across different conversation topics. We provide evidence that this approach can enhance a conversational agent's abilities. The results show that the AMR augmentation improves the ROUGE-L on the generation of the next utterance.