| With advances in computing power and deep learning,neural machine translation（NMT）recaptured the attention of researchers and quickly replaced statistical machine translation to become the mainstream.Despite great research efforts in NMT,few studies compare the performance of a platform over a long period of time.This thesis collects the English to Chinese translations of the US Constitution by Google’s NMT（GNMT）in 2017 and 2022,and uses BLEU and Multidimensional Quality Metrics（MQM）to evaluate and compare them in order to ascertain how much and how GNMT improves,and shed light on future improvement paths.Quantitative results show that the 2022 output achieves+11.44 BLEU over the 2017 output,and average paragraph BLEU scores all show statistically significant improvement.Automatic evaluation results are corroborated by a reduction in MQM penalty scores from 7582 to 3530.The decline is mainly the result of a significant improvement in translation adequacy,with critical mistranslation error slashed by nearly two-thirds.The ensuing qualitative analysis focuses on aspects deemed challenging for NMT,and summarizes five major improvements in adequacy,fluency and language style:a)selecting correct word senses based on contexts,b)avoiding nonsensical word-for-word output for long and complex sentences,c)translating pronouns correctly and explicitly to improve cohesion,d)adjusting word order and sentence structure more flexibly for better readability,and e)detecting some lexical features of the input and recreating them using domain-appropriate words.Drawing on MT research,three future research focuses are proposed.First,domain-specific training can help machines better render word meanings and styles of a specific domain.Second,global context awareness can help improve cohesion and accuracy.Third,better word disambiguation algorithms will help machines deal with rare words and phrases. |
