Revisiting Document Understanding: Enhancing Key–Value Pair Extraction through Layout-Aware Modelling and Data-Centric Analysis on the KVP10k Dataset

Automating information extraction from business documents is essential for digital transformation. Traditional systems depend on predefined templates, but real-world use cases require open Key–Value Pair (KVP) extraction, where field names and structures differ across document types. The KVP10k dataset (ICDAR 2024) offers a large benchmark for this task, yet prior work only evaluated a generative language model (Mistral-7B) and dismissed LayoutLMv3 as unsuitable-without testing it. This thesis revisits that claim by combining layout-aware modeling with data-centric analysis to improve both robustness and interpretability in KVP extraction.