We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations

DBLP - CS Bibliography


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Information Retrieval

Title: MexPub: Deep Transfer Learning for Metadata Extraction from German Publications

Abstract: Extracting metadata from scientific papers can be considered a solved problem in NLP due to the high accuracy of state-of-the-art methods. However, this does not apply to German scientific publications, which have a variety of styles and layouts. In contrast to most of the English scientific publications that follow standard and simple layouts, the order, content, position and size of metadata in German publications vary greatly among publications. This variety makes traditional NLP methods fail to accurately extract metadata from these publications. In this paper, we present a method that extracts metadata from PDF documents with different layouts and styles by viewing the document as an image. We used Mask R-CNN that is trained on COCO dataset and finetuned with PubLayNet dataset that consists of ~200K PDF snapshots with five basic classes (e.g. text, figure, etc). We refine-tuned the model on our proposed synthetic dataset consisting of ~30K article snapshots to extract nine patterns (i.e. author, title, etc). Our synthetic dataset is generated using contents in both languages German and English and a finite set of challenging templates obtained from German publications. Our method achieved an average accuracy of around $90\%$ which validates its capability to accurately extract metadata from a variety of PDF documents with challenging templates.
Comments: A long version of an accepted paper @ JCDL 2021
Subjects: Information Retrieval (cs.IR); Computation and Language (cs.CL); Computer Vision and Pattern Recognition (cs.CV); Digital Libraries (cs.DL); Machine Learning (cs.LG)
Cite as: arXiv:2106.07359 [cs.IR]
  (or arXiv:2106.07359v1 [cs.IR] for this version)

Submission history

From: Zeyd Boukhers [view email]
[v1] Fri, 4 Jun 2021 09:43:48 GMT (748kb,D)

Link back to: arXiv, form interface, contact.