Abstract
KH-FUNSD is the first publicly available, hierarchically annotated dataset designed for Khmer business document understanding. By providing a three-level annotation framework—covering macro-regions, semantic entities, and fine-grained label-value pairs—it establishes a much-needed benchmark for low-resource, non-Latin scripts in the Document AI space.
Introduction
Automated document layout analysis (DLA) is a cornerstone of digital transformation, yet it remains a major hurdle for "low-resource" languages. Khmer, spoken by over 17 million people in Cambodia, is a prime example. Unlike Latin scripts, Khmer features complex ligatures, stacked characters, and a total lack of whitespace between words, making standard OCR and layout tools struggle.
Furthermore, most existing datasets (like the original FUNSD or SROIE) focus on English or high-resource languages. There is a severe lack of annotated Khmer business documents—such as receipts, invoices, and quotations—which are essential for automating public administration and private enterprise workflows. KH-FUNSD fills this gap by providing a standardized dataset and evaluation benchmark.
Methodology: A Three-Level Annotation Framework
The authors developed a rigorous three-level annotation framework to capture the structural and semantic complexity of documents:
- Region Detection: Identifies core zones like the header, form field, and footer.
- FUNSD-style Annotation: Distinguishes between questions, answers, headers, and "other" entities, including the relationships (linking) between them.
- Fine-grained Classification: Assigns highly specific roles such as
header_label,form_value,footer_label, andsymbol(e.g., logos or stamps).
To establish baselines, the researchers evaluated several state-of-the-art models. Object detectors like YOLO (v8, v9, v10) and DETR were used for region-level and fine-grained detection. Multi-modal Transformers, specifically LayoutLM (v1, v2, v3), were used for semantic role assignment and entity linking, as these models can jointly process text, layout, and visual features.
Dataset Characteristics
The KH-FUNSD dataset consists of 158 business documents, including 68 invoices, 40 quotations, and 50 receipts. The data was sourced from open-access repositories and supplemented with synthetically generated samples using authentic templates with anonymized sensitive data.
In total, the dataset contains 12,126 annotated entities. Quality was ensured through a multi-layered review process by Cambodian student researchers, achieving a Cohen’s kappa coefficient of over 0.85, indicating high inter-annotator agreement.
Key Insights and Results
- Hierarchy is Key: Moving from macro-regions to fine-grained labels allows models to learn the "nesting" nature of documents. For instance, a
form_valueis almost always located within aform_fieldregion. - Transformer Superiority: DETR (a transformer-based detector) outperformed the YOLO family in region detection, achieving the highest average mAP50 of 86.6%. This suggests that the global self-attention mechanism in transformers is better at capturing long-range spatial dependencies.
- Multi-modality is Essential: LayoutLMv3, which integrates visual, textual, and layout information at the token level, consistently outperformed its predecessors and pure object detectors for semantic tasks, achieving an F1-score of 68.6% for fine-grained classification. This highlights that for low-resource scripts where OCR might be shaky, visual cues provide vital context.
Conclusion
While KH-FUNSD provides a strong foundation, challenges remain regarding dataset scale and the complexity of dense layouts. Future work aims to include handwritten documents, multi-page forms, and advanced tasks like Document Visual Question Answering (DocVQA) to further support the digital transformation of Khmer-speaking regions.
Summary updated