A Semantic Approach To Evaluation Of Google NMT's EN-CH Polysemy Translation

Hilarious machine translation mistakes used to be an object of ridicule about ten years ago,but this has changed.Since the application of EBMT,PBMT,especially NMT,machine translation has become a great help for those in need of cross-cultural communication and a reliable tool for translators to reduce workload.With its strong database,bettered translation speed and quality,machine translation is now far different from a decade ago,generating more usable results.But it is yet not perfect.Among the many difficulties it faces,word sense disambiguation remains one of the trickiest.From a semantics approach,this thesis evaluates Google NMT's performance in a given polysemy test set,learns its strength and weakness in English-Chinese polysemy translation,so as to give advice to Google NMT's developers and users.This thesis takes a linguistic perspective and uses the semantic categorization of meaning as a basis to evaluate GNMT's performance in English-Chinese polysemy translation.The results show that the accuracy rate is slightly higher than 50%in the given test set.Some problems are found during the evaluation process,such as GNMT's using common word meanings to replace less common meanings,translation by omission,etc.This paper then proposes possible solutions and makes suggestions for developers and users of Google NMT.