Content Planning for Persian Table-to-Text Generation: A Human-Annotated WikiFa Extension

Danyal Pazani¹ Parman Mohammadalizadeh² Leila Safari³

1) 1. B.Sc. Student, Computer Science, University of Zanjan, Email:
2) 2. Computer Engineering, University of Zanjan, Email:
3) 3. Computer Engineering, University of Zanjan, Email:

IOI Code : AAHD-ABKCKC
Publication : National Conference on Information Technology, Nanotechnology, Artificial Intelligence and Technological Futures Studies(itnaf.ir)
Abstract :
    Table-to-text generation (T2T) for Persian remains constrained by the scarcity of high-quality supervision. While the WikiFa benchmark established the first large-scale resource for Persian T2T with over 128,000 table-text pairs, its references are automatically extracted, and its baseline evaluations revealed persistent weaknesses in faithfulness and coherence. This paper extends the benchmark along two dimensions. First, we introduce WikiFa-Gold, a subset of 3,000 human-written reference summaries, produced through a principled two-stage sampling procedure combining K-Means clustering with determinantal point process selection, and annotated under a rigorous multi-round protocol with continuous quality assurance and measured inter-annotator agreement. Second, we introduce WikiFa-Plan, comprising verified content plans for 500 of these instances constructed through LLM-assisted extraction under human supervision, and use it in controlled plan-conditioned generation experiments. In a blind human evaluation of 50 samples under the original benchmark s Likert protocol, plan conditioning substantially improves faithfulness (4.48 versus 3.62 for the unconditioned baseline) while coherence remains high in both conditions, indicating that explicit plans act primarily as a grounding signal that suppresses unsupported content. Automatic evaluation with BLEU, ROUGE-L, and METEOR shows consistent gains under plan conditioning across all three metrics. Both resources will be publicly released upon publication. Together, they provide the first human-verified evaluation standard and the first content-planning resource for Persian T2T, and demonstrate that explicit content planning is a promising direction for faithful generation in low-resource settings.

Keywords : natural language processing natural language generation data to text generation table to text generation