LLM-based Agent for Recommending Information Related to Web Discussions at Appropriate Timing

In web discussions, the amount of information participants have and the level of understanding among them can sometimes restrict their contributions. To address this issue, we developed an agent using an LLM to recommend relevant information based on evidence at appropriate times. The agent extracts information presentation requests from participants’, posts and estimates the appropriate timing for recommendation. Additionally, to enhance the reliability of the agent posts’, we conducted fact-checking during both the information retrieval and opinion generation phases. Experimental results indicate that this agent tends to recommend information at appropriate times. It was also found that proper fact-checking leads to an increase in the relevance and utility of the information recommended by the agent to the discussions. However, we considered it ineffective in situations where few participants have background knowledge, and we believed that information which is easy for discussion participants to understand should be recommended.