Voice Assistants and the Challenge of Navigating Multi-Turn Conversations
Handling multi-turn conversations is a difficult task for voice assistants. They need to understand different types of conversational scenarios, such as steering conversations, carrying over intents, managing disfluencies, retaining relevant entities, and fixing errors. This complexity is compounded by the fact that these scenarios often occur simultaneously in natural language.
A Non-Autoregressive Query Rewriting Architecture
Our solution to this problem is a non-autoregressive query rewriting architecture. This architecture not only addresses the five aforementioned tasks, but also handles complex combinations of these scenarios. Despite being smaller and faster than a fine-tuned T5 model, our proposed model performs just as well or even better on single tasks and outperforms the T5 model in handling combined use-cases.