A Comparative Study Of L2 Neural Machine Translation Post-editing And From-scratch Translation: Process,Product And Translator Attitude

As the current level of globalization is accompanied by the increasingly frequent exchange of information worldwide,the market demand for time-saving and cost-efficient delivery of translation products with end users’ expected quality is on the increase.Post-editing of machine translation(MTPE)is therefore often suggested as a necessary addition to the translation workflow and as an effective solution to the ever-growing demand for translation.Thanks to the rapid development of artificial intelligence,neural machine translation(NMT)has become a hot topic in the language service industry and translation academia.Compared to the previous generations of machine translation systems,NMT has made remarkable progress in translation quality,for its translation is full of natural expressions in the target language – English language in particular,which is symbolic of a significant advance in accuracy and fluency.In recent years,especially with the introduction of cutting-edge research tools,scholars have conducted a multitude of comparative studies between MTPE and human translation from scratch(HT).Nearly all the previous studies concerned focus on L1 MTPE(i.e.,L2 to L1/mother tongue),but there are few studies on L2 MTPE(i.e.,L1 to L2),particularly very few studies center on the applicability of MTPE to L2 translation.In view of that,the present empirical study seeks to shed light on the differences between neural machine translation post-editing(NMTPE)and HT in the directionality of Chinese(L1)to English(L2)from process-,product-,and translator-oriented perspectives.Hopefully,this attempt will deepen our understanding of post-editing from Chinese into English(C-E),and provide implications for C-E translation practice and post-editing training.The present study revolves around post-editing effort and product quality between L2 NMTPE and HT,and translators’ perceptions of and attitude towards NMTPE.The two independent variables are translation mode(or task type)and source text complexity.The former refers to NMTPE and HT.The latter includes two varying levels of source text complexity(one less complex and one more complex)bound with their raw NMT output.For the purpose of triangulation,a multi-method approach to data collection and analysis encompasses:(1)process-oriented instruments –eye-tracking and key-logging,(2)quality-oriented instruments – fine-grained error-based quality assessment in terms of fluency,style and accuracy,and(3)translator-oriented instruments – subjective scales,questionnaires and retrospectivewritten protocols.The present study,focusing on C-E publicity-oriented translation,involves a total of 52 MTI students.All the participants were evenly grouped by their English language proficiency and C-E translation ability.The first group of 26 MTI students performed HT tasks – to translate one less complex and one more complex source text from scratch.The second group of another 26 was in charge of post-editing one less complex text bound with good raw NMT output and one more complex text bound with poor raw NMT output.The research findings are summarized as follows.First and foremost,from the process-oriented perspective,NMTPE is more effortless than HT in translating two varying levels of source text complexity.To be specific,first,NMTPE is significantly faster than HT.Second,NMTPE is technically less demanding than HT.Third,NMTPE is cognitively less demanding than HT for the less complex text bound with good raw NMT output.However,the cognitive effort for the more complex text bound with poor raw NMT output between both task types is comparable or similar.On the whole,NMTPE demonstrates its superiority over HT in efficiency or productivity.In the second place,from the product-oriented perspective,NMTPE outperforms HT in terms of final product quality.To be specific,first,NMTPE is more fluent than HT.Second,NMTPE is stylistically more refined than HT for the less complex text,but HT is stylistically more pleasant than NMTPE for the more complex one.Third,NMTPE is more accurate than HT.Fourth,the differences between NMTPE and HT are compared with the quality of raw NMT output as a baseline.More specifically,in terms of fluency,style and accuracy,the quality of NMTPE and that of raw NMT output are comparable for the less complex text bound with good raw NMT output,but both NMTPE and raw NMT output are superior to HT.Regarding the more complex text bound with poor raw NMT output,the general trend remains the same in terms of both fluency and accuracy,namely,the quality of HT and that of raw NMT output are comparable.However,NMTPE outshines both HT and raw NMT output.The only difference lies in style,that is,the quality of both HT and NMTPE outstrip that of the raw NMT output,but HT still surpasses NMTPE.In general,NMTPE performs significantly better than HT in product quality.Last but not least,from the translator-oriented perspective,MTI students’ ratings of perceived post-editing effort are mostly in agreement with process-oriented measures.To be specific,both are consistent in processing speed and technical effort.However,both are not completely in agreement with respect to cognitive effort.On the whole,two out of three sources of subjective ratings(pre-task questionnaires,post-task questionnaires and retrospective protocols)are in line with the results of cognitive effort as indicated by eye-tracking data.Nonetheless,translator-oriented measures of perceived product quality are basically not in line with the overall product-oriented measures as indicated by the fine-grained quality assessment data.To be specific,out of three sources of subjective ratings(pre-task questionnaires,post-task questionnaires and retrospective protocols),at least two sources are not in agreement with the data of error-based evaluation.Ultimately,a vast majority of MTI students demonstrate a positive attitude towards NMTPE.Overall,all the above findings suggest that NMTPE is more beneficial for MTI students than HT in publicity-oriented translation from Chinese(L1)to English(L2).Admittedly,the present study fails to cover all the variations with regard to the quality of raw NMT output.For instance,a source text of high complexity may be bound with good quality of raw NMT output.