Core Concepts
Chitchat interference in task-oriented dialogues can be effectively managed by incorporating user backstories, enhancing system responses and improving overall dialogue quality.
Stats
During task-oriented dialogues (TODs), human users naturally introduce chitchat that interferes with the conversation flow.
We use few-shot prompting with Llama-2-70B to enhance the MultiWOZ dataset with user backstories.
Our enhanced dataset poses a challenge for systems trained solely on TODs or TODs with preliminary chitchat interaction.
The dataset enables systems to acknowledge user backstories while moving tasks forward effectively, confirmed by human evaluation.
Generating novel chitchat-TOD scenarios tests TOD systems thoroughly and improves resilience to natural user interferences.