Parsing complex document types, such as magazines, is more demanding than parsing simpler types, such as scientific papers. Until now, no suitable dataset for parsing Slovene documents has been available. As part of this master's thesis, we manually annotated 35 documents containing 881 images, thereby creating the first dataset of this kind in Slovene. We used it to evaluate both end-to-end parsing models and their individual components; the best-performing model was Youtu-Parsing, with an edit distance score of 0.031. We trained the DocLayout-YOLO model on the dataset, achieving an mAP of 64.3%. We also developed a pipeline that uses these models to extract articles from documents and enrich them with metadata.
|
