Multi-head Self-attention and BGRU for Online Arabic Grapheme Text Segmentation

The segmentation of online handwritten Arabic text into graphemes/characters is a challenging task for the recognition system due to the nature of this script. That is why, it is better to employ dependency in the context of segments written before and after it, to improve recognition accuracy. In this paper, we introduce Multi-Head Self-Attention (MHSA) and Bidirectional Gated Recurrent Units (BGRU) models for online handwritten Arabic text segmentation, both of which simulate our previous grapheme segmentation model (GSM). The proposed framework consists of word embedding and the combination of complementary Multi-Head Self-Attention and BGRU, which help detect the control points (CPs) for handwritten text segmentation. The CPs delimit each grapheme composed of three main geometric points: starting point (SP), ligature valley point (LVP), angular point (AP), and ending point (EP). To show the effectiveness of our MHSA-BGRU model for online handwritten segmentation and its comparison with GSM, both mean absolute error (MAE), and word error rate (WER) evaluation metrics are used. Experimental results on benchmark ADAB and online-KHATT datasets show the efficiency of our model, which achieves 3.17% and 5.28% for MAE, 12.25% and 25.13% for WER respectively.