Search based self-learning query rewrite system in conversational AI

Query rewriting (QR) is an increasingly important technique for reducing user friction in a conversational AI system. User friction is caused by various reasons, including errors in automatic speech recognition (ASR), natural language understanding (NLU), entity resolution (ER) component, or users’ slip of the tongue. In this work, we propose a search-based self-learning QR framework: User Feedback Search based Query Rewrite system (UFS-QR), which focuses on automatic reduction of user friction for large scale conversational AI agents. The proposed search engine, operating on both global user and individual user level, leverages semantic embedding, NLU output, query popularity and estimated friction statistics into the retrieval and ranking process. In order to construct the index and train the retrieval/ranking models, we adopt a self-learning based method by utilizing implicit feedback, learned from users’ historical interactions. We demonstrate the effectiveness of the UFSQR system, trained without any annotated data, through offline and online A/B experiment on Amazon Alexa user traffic. To the best of our knowledge, this is the first deployed self-learning and search-based QR system for the general task of automatic friction reduction in conversational AI.