OCR Processing of Swedish Historical Newspapers Using Deep Hybrid CNN–LSTM Networks

Proceedings of Recent Advances in Natural Language Processing, pages 190–198 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_023 190 OCR Processing of Swedish Historical Newspapers Using Deep Hybrid CNN–LSTM Networks Molly Brandt Skelbye Dept. of Computer Science and Engineering Chalmers University of Technology and University of Gothenburg 412 96 Gothenburg, Sweden molly.brandt09@gmail.com Dana Dann´ells Spr˚akbanken Text, Dept. of Swedish University of Gothenburg 405 30 Gothenburg, Sweden dana.dannells@svenska.gu.se Abstract Deep CNN–LSTM hybrid neural networks have proven to improve the accuracy of Op- tical Character Recognition (OCR) models for different languages. In this paper we examine to what extent these networks im- prove the OCR accuracy rates on Swedish historical newspapers. By experimenting with the open source OCR engine Cala- mari, we are able to show that mixed deep CNN–LSTM hybrid models outperform pre- vious models on the task of character recog- nition of Swedish historical newspapers span- ning 1818–1848. We achieved an average char- acter accuracy rate (CAR) of 97.43% which is a new state–of–the–art result on 19th century Swedish newspaper text. Our data, code and models are released under CC BY licence. 1 Introduction Making previously inaccessible historical data available by converting them into digital resources through Optical Character Recognition (OCR) tech- niques is important for preserving our cultural heritage and thereto provide greater insight into the past. Today there are several leading off-the- shelf OCR engines such as the commercial AB- BYY FineReader,1 open source OCRopus (Breuel, 2008), Tesseract (Smith, 2007) and Calamari (Wick et al., 2020), offering a comprehensive set of pre–trained models that could be applied during recognition for different languages. These mod- els, in particular models for historical data, vary in accuracy for some languages mainly because they were trained on very limited amount of data. To overcome this limitation researchers have been training language specific character recognition 1https://pdf.abbyy.com models for different time periods (Furrer and Volk, 2011; Breuel et al., 2013; Krishna et al., 2018; Drobac et al., 2019). There have also been at- tempts to improve the accuracy of the models after recognition by applying post-correction methods (Drobac and Lind´en, 2020; Dann´ells and Persson, 2020). In this paper we examine to what extent deep CNN–LSTM hybrid neural networks can im- prove the character accuracy rate (CAR) on 19th century Swedish newspaper text during recognition. Following Drobac and Lind´en (2020) approach we trained a character model for Swedish in Calamari and achieved an average CAR of 97.43% which is a new state–of–the–art result for historical Swedish newspaper text.2 2 Related Work Mainly due to the introduction of recurrent neu- ral networks (RNNs) in particular with the Long Short–Term Memory (LSTM) architecture (Breuel, 2017) great progress has been made in the field of OCR in recent years. Today most of modern OCR systems leverage deep learning algorithms for training models and to improve performance. For instance some state-of-the-art OCR systems use shallow LSTM neural networks, consisting of one or two hidden layers (Breuel, 2008). LSTM models became a standard in 2013 (Breuel et al., 2013) and LSTM-based networks have proven …

Content truncated to 500 words