Knowledge Grounded Dialogue Agents

Online deliberation is a common but effective measure for people to reach an agreement nowadays, but sometimes the
gap of their background knowledge would downgrade the communication efficiency. Recent studies in knowledge
grounded dialogue systems have been conducted to interpret the Knowledge Graphs (KGs) in a chat-bot fashion.
However, dialogue agents assisting dialogue between two human interlocutors based on structured knowledge remain
an under-investigated problem, which may contribute to online meetings and instant messaging apps by relieving the
information asymmetries among different speakers. Bridging the gap between the dialogue agent and structured
knowledge is expected to be challenging in two aspects: (1) dynamic reasoning over structured data with the
development of conversation, and (2) engaging the conversation with constructive information. The proposal introduces
knowledge grounded dialogue agents based on large-scale pre-trained language models (PLMs) for deliberation from
four perspectives: (1) dynamic KG reasoning by triplets ranking with hierarchical attention on context, (2) conversational
intention detection as numerical prediction, (3) prompt-based response generation and (4) human aided evaluation
methods. It also discusses the future work to generalize proposed methods to more complicated environments, including
the challenges of adapting current models to the multi-party conversation and automatic data generation strategies
based on PLMs.