Improved BLSTM Neural Networks for Recognition of On-Line Bangla Complex Words

While bi-directional long short-term (BLSTM) neural network have been demonstrated to perform very well for English or Arabic, the huge number of different output classes (characters) encountered in many Asian fonts, poses a severe challenge. In this work we investigate different encoding schemes of Bangla compound characters and compare the recognition accuracies. We propose to model complex characters not as unique symbols, which are represented by individual nodes in the output layer. Instead, we exploit the property of long-distance-dependent classification in BLSTM neural networks. We classify only basic strokes and use special nodes which react to semantic changes in the writing, i.e., distinguishing inter-character spaces from intra-character spaces. We show that our approach outperforms the common approaches to BLSTM neural network-based handwriting recognition considerably.