Deep learning for post-OCR error correction on Swedish texts

Typ
Examensarbete för masterexamen
Program
Publicerad
2021
Författare
Lundberg, Arvid
Torstensson, Mattias
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
As society becomes increasingly digital, the need to digitize physical documents and texts also increases. The most common technology for this purpose is Optical Character Recognition (OCR). Today’s OCR systems are unable to guarantee a totally accurate scan. The quality of digitization varies and is often negatively impacted by features of the source material. Post-OCR correction is often performed on the text produced by the system with the aim of correcting any errors that are present.To our knowledge, there is currently no neural machine learning based post OCR model available for Swedish. The purpose of this thesis is to develop and train a neural machine learning post-OCR correction model on a set of digitized and OCRed Swedish newspaper texts. When developing the model we took advantage of machine translation techniques as we view the problem as translating incorrect text to correct text. Several configurations of the model were tested, and the model managed to improve the evaluation of all metrics on the withheld validation and test sets. These improvements are, however, rather small and only manage to correct certain errors while skipping many others. Additionally, the system sometimes introduces new errors. While the results show improvement, they are not entirely satisfactory and we believe that additional tuning of hyperparameters and further research into synthetic data generation could lead to better results.
Beskrivning
Ämne/nyckelord
Computer Science , Thesis , Machine Learning , Neural Networks , Deep Learning , Natural Language Processing , OCR , Post-OCR , Swedish
Citation
Arkitekt (konstruktör)
Geografisk plats
Byggnad (typ)
Byggår
Modelltyp
Skala
Teknik / material
Index