Joint Contextual Query Rewriting for Virtual Assistant

Anonymous

Keywords: Language Generation, Contextual Modeling, Conversational AI Deployment, Query Rewriting, Anaphora Resolution

TL;DR: Joint Contextual Query Rewriting between Entity Carry over and Correction by Repetition significantly outperforms single task models in both accuracy and latency

Abstract: Contextual queries are common in multi-turn spoken dialogues with virtual assistants. For example, a user could omit an aforementioned entity using anaphora or nominal ellipsis (entity carry over), or correct a previous recognition error by repeating mistaken words or phrases (correction by repetition). While prior work have touched on these use cases individually, we present a joint query rewriting approach to tackle both. This phonetically aware pointer network model rewrites conversational queries in both use cases into a single context independent query. We compare our joint model with two cascading single task models chained together on a randomly sampled and anonymized virtual assistant dataset. In our experiments, the joint model not only outperforms cascading models by 2.3 points token F1 and 3.6 points exact match accuracy, but also does so while being 1.6 times faster regarding p95 latency.