LABT: A Sequence-to-Sequence Model for Mongolian Handwritten Text Recognition with Local Aggregation BiLSTM and Transformer

Mongolian handwritten text recognition poses challenges with the unique characteristics of Mongolian script, its large vocabulary, and the presence of out-of-vocabulary (OOV) words. This paper proposes a model that uses local aggregation BiLSTM for sequence modeling of visual features and Transformer for word prediction. Specifically, we introduce a local aggregation operation in BiLSTM (Bidirectional Long and Short Term Memory) to improve contextual understanding by aggregating adjacent information at each time step. The improved BiLSTM is able to capture context-dependent and letter shape changes that occur in different contexts. It effectively addresses the difficulty of accurately identifying variable letters and generating OOV words without relying on predefined words during training. The contextual features extracted by BiLSTM are passed through multiple layers of Transformer's encoder and decoder. At each layer, the representations of the previous layer are accessible, allowing layered representations to be refined and improved. By using hierarchical representations, accurate predictions can be made even in large vocabulary text recognition tasks. Our proposed model achieves state-of-the-art performance on two commonly used Mongolian handwritten text recognition datasets.