Table Structure Recognition Based on Cell Relationship, a Bottom-Up Approach

Proceedings of Recent Advances in Natural Language Processing, pages 1–8, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_001 1 Table Structure Recognition Based on Cell Relationship, a Bottom-Up Approach Darshan Adiga Shabir Bhat Muzaffar Shah Viveka Vyeth Datoin, Bengaluru, India firstname.lastname@datoin.com Abstract In this paper, we present a relationship extraction based methodology for table structure recognition in PDF documents. The proposed deep learning-based method takes a bottom-up approach to table recog- nition in PDF documents. We outline the shortcomings of conventional approaches based on heuristics and machine learning- based top-down approaches. In this work, we explain how the task of table structure recognition can be modeled as a cell re- lationship extraction task and the impor- tance of the bottom-up approach in rec- ognizing the table cells. We use Multi- layer Feedforward Neural Network for ta- ble structure recognition and compare the results of three feature sets. To gauge the performance of the proposed method, we prepared a training dataset using 250 ta- bles in PDF documents, carefully select- ing the table structures that are most com- monly found in the documents. Our model achieves an overall accuracy of 97.95% and an F1-Score of 92.62% on the test dataset. 1 Introduction Usage of digital documents have elevated dras- tically over the last two decades and a need for automatic information extraction from these doc- uments has increased. Portable Document For- mat (PDF) has been introduced by Adobe in 1993. PDF documents are the most common format of digital documents and are extensively used in sci- entific research, finance, enterprises etc. As the production and usage of PDF documents have in- creased massively, substantial research work has focused on automating the methods for docu- ment analysis (Correa and Zander, 2017; Kava- sidis et al., 2018). Tabular data is a powerful way to represent the data, among other elements of a document like charts, images etc. Tables are found in a variety of classes of digital documents and are very use- ful to readers to capture, search and compare the facts, summarizations and draw conclusions. Au- tomatically extracting the information from the ta- bles and representing the information in more con- venient formats for digital consumption add im- mense value in the field of document understand- ing (Gilani et al., 2017; Hao et al., 2016). Tables contain structured data but often are ren- dered as semi-structured and unstructured on the digital documents for human consumption. Data can be represented using a variety of layouts in ta- bles without losing the meaning of data (Anand et al., 2019). The layout of tables can vary in alignment, line and word spaces, column and row spans, borders and other styling information. De- pending on the type of documents and authors, the tables may not contain any border lines and the structure of the tables will still be understandable to readers. The data represented by the tables in itself can have different semantics. For example, in a table, a column may contain a list of prices in dollars, …

Content truncated to 500 words