Abstract
Multi-turn conversations with an Enterprise AI Assistant can be challengingdue to conversational dependencies in questions, leading to ambiguities anderrors. To address this, we propose an NLU-NLG framework for ambiguitydetection and resolution through reformulating query automatically andintroduce a new task called "Ambiguity-guided Query Rewrite." To detectambiguities, we develop a taxonomy based on real user conversational logs anddraw insights from it to design rules and extract features for a classifierwhich yields superior performance in detecting ambiguous queries, outperformingLLM-based baselines. Furthermore, coupling the query rewrite module with ourambiguity detecting classifier shows that this end-to-end framework caneffectively mitigate ambiguities without risking unnecessary insertions ofunwanted phrases for clear queries, leading to an improvement in the overallperformance of the AI Assistant. Due to its significance, this has beendeployed in the real world application, namely Adobe Experience Platform AIAssistant.