Kakao Integrates LLMs for Spam Response: Building AI That Explains Classification Results in Sentences

Kakao announced a use case for integrating Large Language Models (LLMs) to handle spam content at the 6th Kakao Tech Meet. Presenter Cho Hye-yeon stated that by utilizing the LLM's sentence generation capabilities to regenerate the rationale for 'spam/normal' judgments into natural sentences—going beyond simple classification—they built an 'explainable AI' that helps monitoring operators clearly understand the AI's decisions. She also emphasized that there are no fixed correct answers in the data collection and modeling process, and that the volume of data and labeling criteria vary depending on the target model's bias and learning scope. For example, she explained that the word 'apple' is classified as 'company' in the stock domain and 'fruit' in the food domain, noting that a deep understanding of the domain and analysis of data characteristics are more important than clear rules.