Core Concepts
The author proposes a multi-task framework to serve as a universal retriever for persona, knowledge, and response selection tasks in conversational retrieval, aiming to improve efficiency and performance.
Abstract
The UniRetriever framework introduces a dual-encoder architecture for context-adaptive dialogue encoding and candidate selection. It addresses the challenges of lengthy dialogues and subtle differences between candidates. The model demonstrates state-of-the-art retrieval quality across various datasets, showcasing its potential as a universal dialogue retrieval system.
Stats
Extensive experiments establish state-of-the-art retrieval quality within and outside training domain.
Proposed framework shows promising generalization capability for different candidate selection tasks simultaneously.
Quotes
"The proposed Universal Conversational Retrieval framework unifies three dominant candidate selection tasks in one framework."
"Our work makes contributions by proposing a universal conversational retrieval framework."