Personalized Image Caption Generation Using Monte Carlo Tree Search

Abstract

This study aims to generate personalized descriptions in image captioning, incorporating individual perspectives and phrasing. With the progress in large language models, achieving notable results in various language tasks is possible. For text generation that reflects individuality, adjusting the language model using limited data from individuals is a challenge. This paper proposes using a personal identification model trained on minimal data combined with Monte Carlo tree search to explore token generation sequences. We demonstrate that this method can produce a broader range of sentences than standard beam search and effectively replicate individuality.