Poster Open Access

OCR error detection and post-correction with Word2vec and BERTje on Dutch historical data

Nynke van 't Hof; Vera Provatorova; Mirjam Cuper; Evangelos Kanoulas

A high quality of OCR-output (Optical Character Recognition) has many benefits. Documents become more accessible to readers and NLP tasks can thrive on the data. However, for many reasons, such as the condition of the documents, the OCR-output of historical documents suffers from a significant amount of errors. This study focuses on detecting and correcting these errors after the OCR process has taken place. It has a focus on Dutch historical data. A comparison will be made between the performance of two methods often used for this over the last few years: word2vec and BERT. While BERT has been shown to substantially outperform Word2Vec on OCR post-correction, the reasons behind this performance gap remain under-explored. From related literature, several pitfalls of word2vec in general were retrieved. This study attempts to find where these pitfalls might occur and compare whether BERT has less (or more) problems with these pitfalls than word2vec. This will give insight not only into the advantages and disadvantages of the used word embeddings for OCR post-correction (on historical data), but also into the application of state-of-the-art methods on historical data, something that these methods have often not been trained on and designed for specifically.

Files (482.7 kB)
Name Size
Abstract.pdf
md5:68e6cda4da3751b0eb15a2dd737e43e0
95.1 kB Download
Poster.pdf
md5:0a02d38d4790895c7723845f9c80f14e
387.7 kB Download
84
73
views
downloads
All versions This version
Views 8484
Downloads 7373
Data volume 11.3 MB11.3 MB
Unique views 7676
Unique downloads 5555

Share

Cite as