A Novel Benchmark for Persian Table-to-Text Generation: A New Dataset and Baseline Experiments

The ability to comprehend and articulate structured data tables into natural language presents a pivotal yet challenging endeavor for automated systems. While substantial strides have been made in English Table-to-Text Generation (T2T) with the aid of large training datasets and advancements in deep neural networks, comparatively limited attention has been given to low-resourced languages. This article addresses this gap by introducing the inaugural large-scale dataset for Persian T2T, comprising 128,628 table-text pairs extracted from Persian Wikipedia articles. Detailed statistical analysis and insights are presented to characterize the dataset, providing a comprehensive understanding of the intricacies inherent in Persian T2T tasks. Along with describing the dataset, the study uses this dataset to finetune multilingual pre-trained language models like mT5 and GPT-2 in order to set solid baselines. Evaluation using automated metrics—including BLEU, ROUGE-L, and METEOR—over test data reveals substantial scope for future improvements in modeling. Furthermore, targeted human assessments expose model deficiencies in the coherent generation of lengthy text and interpreting tables, particularly with respect to ambiguous spans. By offering well-matched yet complex table-description pairs and conducting rigorous comparative analyses, this article establishes a novel benchmark, driving progress in low-resourced conditional text generation and the effectiveness of cross-lingual models. This work not only contributes significantly but also sets the stage for further exploration of this critical yet under-explored task for Persian language processing.