Correction of OCR Word Segmentation Errors in Articles from the ACL Collection through Neural Machine Translation Methods

Correction of OCR Word Segmentation Errors in Articles from the ACL Collection through Neural Machine Translation Methods Vivi Nastase and Julian Hitschler University of Heidelberg Heidelberg, Germany {nastase, hitschler}@cl.uni-heidelberg.de Abstract Depending on the quality of the original document, Optical Character Recognition (OCR) can produce a range of errors – from erroneous letters to additional and spurious blank spaces. We applied a sequence-to-sequence machine translation system to correct word-segmentation OCR errors in scientific texts from the ACL collection with an estimated precision and recall above 0.95 on test data. We present the correction process and results. Keywords: character-level sequence-to-sequence model, word segmentation, ACL collection 1. Introduction The ACL anthology provides a valuable collection of scien- tific articles, and organizing it into a structured format could provide us with additional insight into research in this do- main, help with finding related work and help with keeping up with new developments and ideas. The analysis of the ACL collection was stimulated by the shared task at ACL 2012 (Sch¨afer et al., 2012), the workshop on Rediscovering 50 years of Discoveries, and the series of SemEval tasks on ScienceIE – Extracting Keyphrases and Relations from Sci- entific Publications. Analysis has tackled various aspects of this collection: the citation network (Sim et al., 2012), citation references (Radev and Abu-Jbara, 2012; Gordon et al., 2016), keywords and relation extraction (G´abor et al., 2016), topics and community studies (Bordea et al., 2014), among others. Considering the status of our NLP toolbox, the success on processing such a corpus increases when the texts are clean. The ACL collection contains numerous articles pub- lished before electronic submission became standard for our conferences. These older papers have been scanned and processed through OCR, resulting in texts that con- tain errors. A brief inspection of this portion of the col- lection has shown that a very common error introduced by the OCR process consists of spurious blank spaces, which split words randomly in a varying number of smaller frag- ments (Figure 1). This problem is so pervasive, that it influ- ences subsequent processing, for example keyword extrac- tion – in the list of keywords produced with the SAFFRON system, we found the following keywords in papers from Coling 1996: Non-l inear, Context fi, Segmentat ion, Ion theory. Inspired by previous work on error correction using ma- chine translation models (e.g. (Yannakoudakis et al., 2017)), we apply a character-level sequence-to-sequence model to learn how to segment English words in the ACL collection. Written modern languages of European origin usually segment words explicitly, so English texts generally do not require word segmentation, but as we have seen, this problem has popped up in documents processed with OCR. This means that we can very easily obtain large volumes of data for training. We use a portion of the ACL collec- tion (the articles published after electronic submission had become the norm in our community, a date which we con- servatively set at 2005) to generate training, development and test data. The high results – …

Content truncated to 500 words