Visual Detection with Context for Document Layout Analysis

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 3464–3470, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 3464 Visual Detection with Context for Document Layout Analysis Carlos X. Soto Brookhaven National Laboratory Upton, New York, USA csoto@bnl.gov Shinjae Yoo Brookhaven National Laboratory Upton, New York, USA sjyoo@bnl.gov Abstract We present 1) a work in progress method to visually segment key regions of scientific articles using an object detection technique augmented with contextual features, and 2) a novel dataset of region-labeled articles. A con- tinuing challenge in scientific literature min- ing is the difficulty of consistently extract- ing high-quality text from formatted PDFs. To address this, we adapt the object-detection technique Faster R-CNN for document lay- out detection, incorporating contextual infor- mation that leverages the inherently localized nature of article contents to improve the re- gion detection performance. Due to the limited availability of high-quality region-labels for scientific articles, we also contribute a novel dataset of region annotations, the first version of which covers 9 region classes and 822 arti- cle pages. Initial experimental results demon- strate a 23.9% absolute improvement in mean average precision over the baseline model by incorporating contextual features, and a pro- cessing speed 14x faster than a text-based technique. Ongoing work on further improve- ments is also discussed. 1 Introduction Mining scientific literature at scale for informa- tion that can be processed automatically is a valu- able and much sought after technique for domain researchers and data scientists. Whereas mass processing of articles was once largely limited to keyword searches and citation crawling, modern natural language processing techniques can now deeply search for specific and broad concepts, ex- plore relationships, and automatically extract use- ful information from text. However, getting the article text in the first place is not always simple. Though some publications offer full-text articles in open-access, machine- readable formats (e.g. plain text, HTML or XML), Figure 1: Text extraction from PDF articles is generally a noisy process because of the significant inconsisten- cies in PDF publishing tools and the edit-and-republish nature of print-ready document preparation by journal publishers and document hosting services. this is far from the norm. Instead, much of the scientific record remains in PDF files that are much more challenging to automatically process. This challenge remains whether the PDFs origi- nate from scanned pages (thus requiring OCR) or were published from digital source material. Al- though there are tools for automatically extract- ing text from PDFs, formatting is highly incon- sistent. Standard tools often mix headers, footers, table and figure captions, page numbers, and other extraneous text into the main text being extracted (Bast and Korzen, 2017). And text order is not al- ways well preserved (Figure 1). This is not quite as problematic if text is extracted for simple tasks like keyword searches. However, more complex tasks like Named-Entity Recognition often rely on contiguous, cleanly segmented text for successful processing. And …

Content truncated to 500 words