Leveraging Large Language Models for Flexible and Robust Table-to-Text Generation

Generating natural language descriptions from structured tabular data is a crucial challenge with high-impact applications across diverse domains, including business intelligence, scientific communication, and data analytics. Traditional rule-based and machine learning approaches have faced limitations in reusability, vocabulary coverage, and handling complex table layouts. Recent advances in LLMs pre-trained on vast corpora offer an opportunity to overcome these limitations by leveraging their strong language understanding and generation capabilities in a flexible learning setup. In this paper, We conduct a comprehensive evaluation of two LLMs - GPT-3.5 and LLaMa2-7B - on table-to-text generation across three diverse public datasets: WebNLG, NumericNLG, and ToTTo. Our experiments investigate both zero-shot prompting techniques and finetuning using the parameter-efficient LoRA method. Results demonstrate GPT-3.5's impressive capabilities, outperforming LLaMa2 in zero-shot settings. However, finetuning LLaMa2 on a subset of data significantly bridges this performance gap and produces generations much closer to ground truth and comparable to SOTA approaches. Our findings highlight LLMs' promising potential for data-to-text while identifying key areas for future research.