Encoder-Decoder Language Model for Khmer Handwritten Text Recognition in Historical Documents

Correcting spelling errors in texts extracted from Khmer palm leaf manuscripts by handwritten text recognition (HTR) systems can be very challenging. A Khmer Language Model developed in this study aims to facilitate the task mentioned above. The proposed model utilizes long short-term memory (LSTM) modules applicable for improving the performance of text recognition which is to predict a sequence of characters as output. The architecture of the language model is based on an encoder-decoder mechanism which is composed of two parts: an encoder to capture the context of the input erroneous word and a decoder to decode and predict the correctly spelt output word. Experimental evaluations are conducted on a text corpus consisting of Khmer words extracted from Sleuk-Rith set.