말한 대로 작동하는 TMAP, AI가 AI를 개선하는 방법
SK Telecom’s TMAP navigation service has upgraded its AI model, TMAP Planner, to better interpret natural language requests across 14 service domains, reducing misclassification errors through automated learning and validation processes.
SK Telecom’s TMAP navigation service has evolved from executing predefined voice commands to understanding natural language context and intent. The core improvement lies in TMAP Planner, an AI model that analyzes user speech to determine the appropriate service domain—such as navigation, music playback, or weather queries—and formats the request accordingly. This shift addresses the challenge of interpreting varied expressions like 'go home' as navigation requests without explicit commands, requiring contextual understanding of follow-up phrases such as 'stop by a gas station nearby.'
The accuracy of TMAP Planner’s initial intent classification directly impacts the overall user experience, as misjudgments can trigger unintended functions. Unlike traditional rule-based systems, TMAP Planner learns through repeated exposure to examples, refining its judgment criteria over time. SK Telecom has implemented an automated system, TMAP Planner AX, to streamline the learning and validation process, enabling the AI to analyze large datasets and identify errors more efficiently than manual review.
TMAP Planner AX leverages a dataset of approximately 25,900 entries, including user speech, Planner judgments, and domain responses, to train and validate the model. The system distinguishes between 'judgment rules' and 'example data,' ensuring consistency between predefined guidelines and actual usage patterns. Errors in either component can lead to misclassification, prompting AI-driven corrections that are then reviewed by developers for alignment with service quality standards.
By automating the analysis of new issues and proposed fixes, TMAP Planner AX reduces the manual workload for developers, who now focus on evaluating AI-generated solutions rather than drafting rules from scratch. This approach accelerates response times to emerging expressions or service requests while maintaining consistency across the 14 TMAP domains. The goal is to minimize instances where user intent is misaligned with executed functions, enabling a more natural and reliable voice interaction experience.