HypGen: A Generative Approach for Natural Language UnderstandingEnrichment Using Implicit User Feedback

Anonymous

Keywords: NLU interpretation, generative approach, implicit user feedback

TL;DR: We propose a model-based generative approach for automatic NLU n-best enrichment, by leveraging implicit user feedback learned from the existing user dissatisfaction and rephrase detection techniques.

Abstract: In a conversational AI system, Natural Language Understanding (NLU) is a key component that produces the semantic interpretations for a user request. As one of the most upstream components, NLU errors are costly and have a wide blast radius to the end-to-end system behavior. We propose a model-based generative approach for automatic NLU n-best enrichment, by leveraging implicit user feedback learned from the existing user dissatisfaction and rephrase detection techniques. It aims to generate reasonable and promising NLU interpretation candidates that can potentially correct NLU errors. We propose three different modeling designs and compare their performance with extensive experiments on live traffic from one of today's state-of-the-art large-scale conversational AI systems.