Deep Learning for Arabic Word Classification: Leveraging Transfer Learning and Grad-CAM for Morphological Analysis
摘要
We propose a deep learning approach to deal with large Arabic lexicon recognition. We propose a combination of methods aimed at learning and recognizing written decomposable Arabic words, trying to simulate writing human reading. The training is based on derivational and inflectional characteristics specific to Arabic. Word recognition process begins with an initial phase of inflection and derivation classifications using six CNN-BiLSTM models that have been transfer learned and fine-tuned. Subsequently, we employ the Grad-CAM technique on these classifiers to localize and extract word prefixes, infixes, and suffixes. Following this, a root extraction phase utilizes image pre-processing techniques to isolate characters that belong to the root, serving as preparation for root classification. Ultimately, we aim to utilize the classifier results to reconstruct words based on the different feature of the Arabic words morphology.