Fine-Tuning Large Language Models for Automatic Sentence Generation in Augmentative and Alternative Communication

초록

This paper presents an automatic sentence generation process for augmentative and alternative communication (AAC) in the Korean language, achieved by fine-tuning large language models (LLMs). The process begins with the collection of relevant data using LLMs, followed by the fine-tuning of a custom model specifically for sentence generation, and concludes with an evaluation of its performance via external LLMs. During data collection, prompts are provided to guide sentence generation based on specific words. The selected LLMs are then fine-tuned using this collected data. For performance analysis, the best performing fine-tuned model is compared with Meta-Llama-3-8B, llama-3-Korean-Bllossom-8B, and ChatGPT-3.5 Turbo, using external LLMs as evaluators. Experimental results indicate that the fine-tuned model outperforms the others. Lastly, a human evaluation is conducted to validate the LLM evaluations, confirming the consistency of the results.

키워드