Achieving Equilibrium between Accuracy and Efficiency in Customer Oriented Conversational AI: A Rasa-Focused Methodology Incorporating Intent Ranking and Disambiguation

Abstract

As conversational interfaces emerge as a primary medium for customer assistance, attaining both high accuracy and low latency in intent classification presents a significant challenge. Despite their remarkable linguistic skills, large language models (LLMs) can have unreasonably substantial computational overhead, which makes them unsuitable for applications that require quick responses from customers. This article proposes a Rasa-centric methodology for Natural Language Understanding (NLU) that employs Rasa's intent ranking system and a tailored disambiguation policy to address ambiguous terminology. By analyzing relevant research and actual data, we illustrate that strategically prompting users to clarify their intent enhances efficiency,decreases latency, and mitigates misclassification risks, especially for overlapping or closely related intentions. We determine that Rasa's efficient NLU pipeline, coupled with limited user engagement during instances of ambiguity, can significantly enhance both user experience and operational efficiency.