Presentation Information
[4K1-GS-6a-06]A Comparative Study on Alignment Methods for LLM-based Web Content Writing
〇Sayaka Yamamoto1, Satoshi Kimura2, Takafumi Koshinaka1 (1. Yokohama City University , 2. CyberAgent, Inc.)
Keywords:
LLM,Alignment,SEO,LLM-as-a-Judge,Contents Writing
In recent years, the utilization of Large Language Models (LLMs) in Web contents writing has been actively discussed. The primary goal of Web contents writing is to maximize page views, which, from the perspective of Search Engine Optimization (SEO), boils down to creating content that aligns with user preferences.
In this study, we focus on Direct Preference Optimization (DPO) and Kahneman-Tversky Optimization (KTO) as alignment methods to reflect user preferences in LLMs. We compare the effectiveness of both methods in the domain of Web contents writing, where readability, word count, and information accuracy are essential.
Through human evaluation following the scheme of user testing commonly practiced in SEO, we assessed the Web articles generated by each model across multiple criteria. The results confirmed that the model aligned with KTO achieved higher scores, demonstrating the superiority of this method. Furthermore, we conducted similar tests using an LLM instead of human raters to verify the validity of "LLM-as-a-Judge" in this specific domain.
In this study, we focus on Direct Preference Optimization (DPO) and Kahneman-Tversky Optimization (KTO) as alignment methods to reflect user preferences in LLMs. We compare the effectiveness of both methods in the domain of Web contents writing, where readability, word count, and information accuracy are essential.
Through human evaluation following the scheme of user testing commonly practiced in SEO, we assessed the Web articles generated by each model across multiple criteria. The results confirmed that the model aligned with KTO achieved higher scores, demonstrating the superiority of this method. Furthermore, we conducted similar tests using an LLM instead of human raters to verify the validity of "LLM-as-a-Judge" in this specific domain.