A Controllable Agent-based Text Generation Method from Structured Data

Text generation from structured data is an important task in many NLP applications, which aim to automatically generate text content with accurate description and clear structure from the structured data. Most existing models rely on the encoder-decoder framework and require large datasets for training, which tend to generate detailed descriptions of tabular data and are lacking in inferential content such as statistical analysis. In this paper, a controllable agent-based text generation method with structured data is proposed, for which the main goal is to provide richer and more detailed descriptions. By conducting various computational operations on tabular data and presenting the analysis results in graphical form, the understanding of the data is enhanced to reveal underlying trends. By planning multiple actions such as text generation, data calculation, and graph plotting, a report that is both rich in form and deep in content can be generated by using a large language model as an agent with some modules to perform these functions. Experiments on the structured datasets coming from the temporal and spatial data of the suspects show that the agentbased text generation method provides more accurate and deeper results, which illustrates that the proposed method is effective and practical for the structured data.