Handwritten Text Recognition Error Rate Reduction in Historical Documents using Naive Transcribers

Handwritten text recognition (HTR) is a difficult research problem. In particular for historical documents, this task is hard as handwriting style, orthography, and text quality pose significant challenges. Creation of a single multi-purpose HTR system seems to be out of reach for current state-of-the-art systems. Therefore, we are interested in fast creation of specialized HTR systems for a particular set of historical documents. Manual annotation by historical experts is expensive and can often not be applied at a large scale. Instead, we use the transcripts of naive transcribers that may still contain a significant amount of errors. In this paper, we propose to fuse the recognized word-chain with naive transcribers that can be obtained in a cost-effective way. For the actual fusion, we rely on a word-level approach, the so-called Recognizer Output Voting Error Reduction (ROVER). Results indicate that we are able to reduce the Word Error Rate (WER) of an HTR system trained with only few pages from 24.6 % to 19.2% with two additional transcribers with 25.1% and 27.1% WER each. This performance is already close to current state-of-the-art systems trained with significantly more data.