Date: 2021-10-08 / 3:00 ~ 4:00 PM
The use of reasoning and commonsense knowledge is critical to holding a successful dialogue with a human. However, state-of-the-art dialogue models are inconsistent in both their reasoning capabilities and maintenance of factual knowledge. Furthermore, the black box nature of these models makes it difficult to exercise control over the dialogue model's behaviors or internal logic, preventing direct investigations as to what reasoning processes are required to hold successful conversations. ReDD is a dialogue development framework equipped with built-in support for reasoning over predicates in first order logic (FOL). It can compute entailments across thousands of FOL implications in well under a second using a GPU-based graph matching algorithm. Its modular design allows for easy extension, such as integration of custom models for reasoning, NLU, and NLG. ReDD is currently under development, but we are hopeful it will prove a valuable tool to facilitate experimentation with reasoning and knowledge management processes in dialogue systems.