Enabling controllable table-to-text generation via prompting large language models with guided planning

Recently, Large Language Models (LLMs) has demonstrated unparalleled capabilities in understanding and generation, hence holding promising prospects for applying LLMs to table-to-text generation. However, the generation process with LLMs lacks a high degree of controllability, which hinders the utilization of LLMs for table-to-text generation. In this paper, we introduce Poised, an effective method that prompts LLMs with guided planning to achieve controllable table-to-text generation. Specifically, we first employ prefix-tuning on BART to derive a plan from the given table. Then, we combine the plan with guided instructions to create a comprehensive prompt, which is later input into LLMs to generate the description of the table. Experiments across three domains of the few-shot Wiki dataset show that Poised achieves or approaches a plan completion rate of 100%, with an average hallucination frequency of less than 10%. Furthermore, Poised allows for fine-grained control over the generated content by intentionally modifying the prompt, enabling precise control over aspects such as attribute realization order.