Table-to-Text Using Pre-trained Large Language Model and LoRA

This paper propose a Table-To-Text system that generates text explaining the contents of a table from the table itself. To accurately capture the information in the table, it is represented in JSON format. The use of JSON format allows for handling not only table data but also various other types of information, such as sensor data in IoT. To develop this system, we fine-tune an existing pre-trained large language model with a Table-To-Text dataset. Furthermore, to reduce the computational cost of fine-tuning, we utilize Low-Rank Adaptation (LoRA). Our evaluation experiments show that while large language models without fine-tuning is not able to solve the Table-To-Text tasks, our proposed system is able to generate text based on tables. We have confirmed that fine-tuning with LoRA can extend large language models to new tasks that were not anticipated during pre-training. As future work, we will develop Table-To-Text system using other large language model.