2016/10/10 by Jason D. Lee, Kyunghyun Cho, Lee, Jason +3 · 1 citation
Computer Science · #Computation and Language (cs.CL) #FOS: Computer and information sciences #Handwritten Text Recognition Techniques #Machine Learning (cs.LG) #Natural Language Processing Techniques #Topic Modeling
paper · pdf · doi:10.48550/arxiv.1610.03017
openalex publication_date 2016/10/10 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Most existing machine translation systems operate at the level of words,\nrelying on explicit segmentation to extract tokens. We introduce a neural\nmachine translation (NMT) model that maps a source character sequence to a\ntarget character sequence without any segmentation. We employ a character-level\nconvolutional network with max-pooling at the encoder to reduce the length of\nsource representation, allowing the model to be trained at a speed comparable\nto subword-level models while capturing local regularities. Our\ncharacter-to-character model outperforms a recently proposed baseline with a\nsubword-level encoder on WMT'15 DE-EN and CS-EN, and gives comparable\nperformance on FI-EN and RU-EN. We then demonstrate that it is possible to\nshare a single character-level encoder across multiple languages by training a\nmodel on a many-to-one translation task. In this multilingual setting, the\ncharacter-level encoder significantly outperforms the subword-level encoder on\nall the language pairs. We observe that on CS-EN, FI-EN and RU-EN, the quality\nof the multilingual character-level translation even surpasses the models\nspecifically trained on that language pair alone, both in terms of BLEU score\nand human judgment.\n