2018/04/17 by Peyman Passban, Passban, Peyman, Liu, Qun +2
Computer Science · #Natural Language Processing Techniques #Topic Modeling
paper · pdf · doi:10.48550/arxiv.1804.06506
Recently, neural machine translation (NMT) has emerged as a powerful\nalternative to conventional statistical approaches. However, its performance\ndrops considerably in the presence of morphologically rich languages (MRLs).\nNeural engines usually fail to tackle the large vocabulary and high\nout-of-vocabulary (OOV) word rate of MRLs. Therefore, it is not suitable to\nexploit existing word-based models to translate this set of languages. In this\npaper, we propose an extension to the state-of-the-art model of Chung et al.\n(2016), which works at the character level and boosts the decoder with\ntarget-side morphological information. In our architecture, an additional\nmorphology table is plugged into the model. Each time the decoder samples from\na target vocabulary, the table sends auxiliary signals from the most relevant\naffixes in order to enrich the decoder's current state and constrain it to\nprovide better predictions. We evaluated our model to translate English into\nGerman, Russian, and Turkish as three MRLs and observed significant\nimprovements.\n