Developing Open-Domain Dialogue Systems with Causal Insights

This thesis improves open-domain dialogue systems, which power chatbots and virtual assistants, by making them more reliable and meaningful. Current systems often give vague or off-topic answers because they rely on superficial patterns. We apply causality theory to fix this: first, by teaching models to focus on the truly important parts of conversations; second, by filtering dialogue history so responses are more coherent and specific; and third, by creating a new evaluation method that better reflects human judgment. We also show how to automatically update dialogue knowledge using causal discovery, making systems smarter and more trustworthy.