Improved Typesetting Models for Historical OCR
Taylor Berg-Kirkpatrick, Dan Klein · 2014
We present richer typesetting models that extend the unsupervised historical document recognition system of Berg-Kirkpatrick et al. (2013). The first model breaks the independence assump-tion between vertical offsets of neighbor-ing glyphs and, in experiments, substan-tially decreases transcription error rates. The second model simultaneously learns multiple font styles and, as a result, is able to accurately track italic and non-italic portions of documents. Richer mod-els complicate inference so we present a new, streamlined procedure that is over 25x faster than the method used by Berg-Kirkpatrick et al. (2013). Our final sys-tem achieves a relative word error reduc-tion of 22 % compared to state-of-the-art results on a dataset of historical newspa-pers. 1