[2L1-GS-10t] AI application

9:30 AM - 9:45 AM JST(12:30 AM - 12:45 AM UTC)

[2L1-GS-10t-03] Evaluation of Methods for Presentation Slide Translation and Layout Error Correction by LLMs

〇Futo Kajita¹, Nobuyori Nishimura¹, Takehito Utsuro¹, Naoki Muto², Chee Siang Leow², Hiromitsu Nishizaki² (1. University of Tsukuba, 2. University of Yamanashi)

Keywords:

educational application,slide translation

We propose a novel approach for translating Japanese slides into English and correcting layout errors by utilizing LLMs to edit XML structures. Existing tools suffer from layout errors due to text expansion, while our previous three-step approach faced practical challenges with increased processing time and costs from multiple LLM inferences. To overcome this issue, we propose a one-step method that completes translation and layout correction in a single inference. By simultaneously inputting the slide's XML and image alongside Japanese/English slide pairs as few-shot examples, the LLM directly learns complex layout corrections.Experimental results demonstrate our method outperforms existing translation tools in layout quality. Furthermore, compared to multi-step approaches requiring multiple inferences, including the highly accurate three-step method, our one-step approach maintains practical quality while significantly reducing computational costs and processing time.