Upcycle Your OCR: Reusing OCRs for Post-OCR Text Correction in Romanised Sanskrit

Proceedings of the 22nd Conference on Computational Natural Language Learning (CoNLL 2018), pages 345–355 Brussels, Belgium, October 31 - November 1, 2018. c⃝2018 Association for Computational Linguistics 345 Upcycle Your OCR: Reusing OCRs for Post-OCR Text Correction in Romanised Sanskrit Amrith Krishna#, Bodhisattwa Prasad Majumder*, Rajesh Shreedhar Bhat**, and Pawan Goyal# #Dept. of Computer Science and Engineering, IIT Kharagpur, *Dept. of Computer Science, University of California, San Diego **Walmart Labs, India amrith@iitkgp.ac.in, bmajumde@eng.ucsd.edu, rajeshbhatpesit@gmail.com, pawang@cse.iitkgp.ernet.in Abstract We propose a post-OCR text correction ap- proach for digitising texts in Romanised San- skrit. Owing to the lack of resources our ap- proach uses OCR models trained for other lan- guages written in Roman. Currently, there ex- ists no dataset available for Romanised San- skrit OCR. So, we bootstrap a dataset of 430 images, scanned in two different settings and their corresponding ground truth. For training, we synthetically generate training images for both the settings. We find that the use of copy- ing mechanism (Gu et al., 2016) yields a per- centage increase of 7.69 in Character Recog- nition Rate (CRR) than the current state of the art model in solving monotone sequence-to- sequence tasks (Schnober et al., 2016). We find that our system is robust in combating OCR-prone errors, as it obtains a CRR of 87.01% from an OCR output with CRR of 35.76% for one of the dataset settings. A hu- man judgement survey performed on the mod- els shows that our proposed model results in predictions which are faster to comprehend and faster to improve for a human than the other systems1. 1 Introduction Sanskrit used to be the ‘lingua franca’ for the sci- entific and philosophical discourse in ancient In- dia with literature that spans more than 3 millen- nia. Sanskrit primarily had an oral tradition, and the script used for writing Sanskrit varied widely across the time spans and regions. With the ad- vent of printing press, Devanagari emerged as the prominent script for representing Sanskrit. With standardisation of Romanisation using IAST in 1894 (Monier-Williams, 1899), printing in San- skrit was extended to roman scripts as well. There able1The data and the codes for our system are avail- here - https://github.com/majumderb/ sanskrit-ocr has been a surge in digitising printed Sanskrit manuscripts written in Roman such as the ones currently digitised by the ‘Krishna Path’ project2. In this work, we propose a model for post- OCR text correction for Sanskrit written in Ro- man. Post-OCR text correction, which can be seen as a special case of spelling correction (Schnober et al., 2016), is the task of correcting errors that tend to appear in the output of the OCR in the pro- cess of converting an image to text. The errors incurred from OCR can be quite high due to nu- merous factors including typefaces, paper quality, scan quality, etc. The text can often be eroded, can contain noises and the paper can be bleached or tainted as well (Schnober et al., 2016). Fig- ure 1 shows the sample images …

Content truncated to 500 words