Notification Text Generation with Large Language Models

Abstract

The delivery of notifications with engaging and contextually appropriate content plays a critical role in increasing user engagement. Although Large Language Models (LLMs) are successful in text generation, systematic research on short, action-oriented content such as notification text generation is limited. In this study, Manual Prompt Creation, Automatic Prompt Creation, and Optimized Automatic Prompt Creation approaches for LLM-based notification text generation were compared. The outputs were generated using Google Gemini Pro and evaluated with LLM-based metrics such as grammatical accuracy, title-text coherence, and naturalness. The results revealed that Optimized Automatic Prompt Creation outperformed other approaches, while Manual Prompt Creation provided more diverse texts. This study contributes to the literature on notification text generation by advancing prompt engineering and introducing new evaluation metrics.