Applying semantic prosody for machine translation improvement on English-Chinese passive sentences

Abstract

Passive sentences are widely used in English and have a neutral semantic prosody, while in Chinese they are much less frequent and with negative semantic prosody. Thus, many times English passives should be translated into active Chinese sentences and passive voice should be mainly reserved for unfavorable content. This work uses focused datasets to fine-tune a neural network machine translation model and improves its performance on translating English BE passives to Chinese. Evaluation after fine-tuning suggests training on English-Chinese sentence pairs that contain BE passives that are translated into actives by human translator, but to BEI passives by the model leads to considerable improvement of model performance. Also, data shows that the semantic orientation score of BE passives translated into Chinese passives by human translators are significantly different from those translated into active sentences, providing quantifiable way to address this issue.