2020/04/23 by Ori Terner, Terner, Ori, Kfir Bar +3 · 1 citation
Computer Science · #Natural Language Processing Techniques #Handwritten Text Recognition Techniques #Topic Modeling
paper · pdf · doi:10.48550/arxiv.2004.11405
We trained a model to automatically transliterate Judeo-Arabic texts into\nArabic script, enabling Arabic readers to access those writings. We employ a\nrecurrent neural network (RNN), combined with the connectionist temporal\nclassification (CTC) loss to deal with unequal input/output lengths. This\nobligates adjustments in the training data to avoid input sequences that are\nshorter than their corresponding outputs. We also utilize a pretraining stage\nwith a different loss function to improve network converge. Since only a single\nsource of parallel text was available for training, we take advantage of the\npossibility of generating data synthetically. We train a model that has the\ncapability to memorize words in the output language, and that also utilizes\ncontext for distinguishing ambiguities in the transliteration. We obtain an\nimprovement over the baseline 9.5% character error, achieving 2% error with our\nbest configuration. To measure the contribution of context to learning, we also\ntested word-shuffled data, for which the error rises to 2.5%.\n