ChatGPT|(一)结合ChatGPT的译前编辑初探
基于ChatGPT的机辅翻译
(一)结合ChatGPT的译前编辑初探
引言各位小伙伴大家好!如今ChatGPT等生成式人工智能技术产品方兴未艾,有效利用可以使翻译工作变得便利高效。前期公众号其他伙伴和老师已经分享了如何使用ChatGPT进行口笔译翻译,比如利用ChatGPT生成中英双语术语表辅助翻译任务,利用ChatGPT生成译文和进行译后编辑等。在ChatGPT等大语言模型(LLM)兴起的时代,翻译专业和行业遭受挑战,但同时如果运用得好,ChatGPT也可为我们的翻译工作提供便利。今天小编将结合上次“ChatGPT在口笔译中的应用与实战公益沙龙”中“基于ChatGPT的机辅翻译”分享内容更新系列推文。今天所分享的第一个话题是“结合ChatGPT的译前编辑初探”。
那么如何利用ChatGPT做好译前编辑以提升译文质量呢?我们可以再结合受控语言(Controlled Language)加以辅助。
首先先简要介绍一下译前编辑和受控语言两个概念。
何为译前编辑呢?
译前编辑与译后编辑都是用来提升机器翻译译文质量的手段。译前编辑需要在翻译之前对源语言进行处理,在词汇、句法、语篇等方面对源语进行调整,从而更便于翻译。译后编辑需在翻译之后对目的语进行优化处理,对术语、风格和语法等方面进行检查和编辑。好比一个是菜品的烹前预处理,一个是菜品烹后再加工,两者都是为了菜品更为精致,从而更受客户喜爱。
何为受控语言?
根据新加坡翻译服务商andovar官网(https://blog.andovar.com/controlled-language)介绍,“Controlled language is a subset of natural language with a restricted grammar and vocabulary that reduces or eliminates ambiguity and complexity”,即——受控语言是自然语言下的一个分支,其利用有限的语法和词汇以减少不确定性和复杂性。
《劳特里奇翻译技术百科全书》(Routledge Encyclopedia of Translation Technology)(Chan Sin-Wai, 2015)在Controlled Language这一章节介绍了受控语言的定义及分类。其定义如下:“Controlled natural languages tackle these kinds of problems by restricting the size of the grammar and vocabulary in order to reduce or eliminate ambiguity and complexity inherent in natural languages.”
因此可以得知,受控语言能够在语法和词汇层面着力降低文本因模糊和复杂所带来的理解困难。根据这本书,在机器翻译用途方面,“The primary objective of using controlled language for MT is to limit lexical ambiguity in the source language and to rule out complex sentence structures in order to ease the processing and achieve better translation results.”也就是说,受控语言通过限制源语言的词汇歧义和排除复杂句结构来简化处理和获得更佳翻译效果。
规则:
对于受控语言的使用,翻译学者Sharon O’Brien在Controlling Controlled English——An Analysis of Several Controlled Language Rule Sets论作中基于Lexical(词汇)、Syntactic(句法)、Semantic(语义)三个层面给出了一套详细的规则。在词汇层面,还包括词汇选用,拼写,代词等诸多细节规则,可见作者考量之细微。我们可根据这套规则结合所修改文本情况合理应用。(由于篇幅有限,大家可自行检索该文献详细阅览)
源语言文本:
The first phenomena that we now recognize to be caused by electrons were noticed during the nineteenth century when inventors like Edison were studying electrical effect's in evacuated glass bulbs and tubes. Sometimes certain areas of the almost completely evacuated vessel would become luminous and glow with blue, green or red light. Though these effects were exploited to produce colourful discharges—now well known as the fluorescent neon-tubes used in advertising—the cause of these phenomena was then unknown.
上文是一段电子科学相关英语文本。细看该文本可发现文中从句、过去分词短语等复杂结构特征明显,还涉及一些专业术语。这在一定程度上增加了理解和翻译的难度。那么接下来就需要给ChatGPT加点Prompt,再佐以Controlled Language调料来优化源语言文本啦。
小编基于上面这段文本给出了自己的ChatGPT Prompt:
Suppose you are an English translator and reviser, please help me revise the following text about technology, so that the text is easy to be understood and translated into Chinese. Based upon rules of controlled language, you should, lexically make sure that the vocabulary is simple, and syntactically make sure the sentence structure is simple and rule out using passive voice.
The first phenomena that we now recognize to be caused by electrons were noticed during the nineteenth century when inventors like Edison were studying electrical effect's in evacuated glass bulbs and tubes. Sometimes certain areas of the almost completely evacuated vessel would become luminous and glow with blue, green or red light. Though these effects were exploited to produce colourful discharges—now well known as the fluorescent neon-tubes used in advertising—the cause of these phenomena was then unknown.
(当然,自己能力有限,这段指令如有问题,还请大家指正)
ChatGPT所给修改后译文如下:
During the 1800s, inventors like Edison studied electrical effects in glass bulbs and tubes from which the air had been removed. They observed that some parts of the nearly empty bulb would light up and emit blue, green, or red light. These colorful discharges were later used in advertising through fluorescent neon tubes. However, at that time, the reason behind these phenomena remained unknown.
对比原文本和修改后文本可发现,ChatGPT修改后版本字符数更少,句子复杂度更低。在可读性方面,还利用弗莱士-金凯德难度级数测评(数值越大,难度越高),发现修改后文本可读性也得到提高。
这些都是有利于译者理解的一面。但是,小编也发现了问题,修改后版本中“The first phenomena that we now recognize to be caused by electrons”所强调的意思似乎在ChatGPT修改时也被一并省去了,这可不利于译者在轻松理解的同时再现原文意思。为了确定这不是偶然之举,小编尝试输入了将近十次上述Prompt和原文,然而还是发现这一问题在每次所给结果中仍未得到解决。最后,小编还在Prompt后面追加了“Most importantly, do not distort the meaning of the original text.”可是,结果表明,这样一来,ChatGPT似乎又变成了耳边风,意思较好地保留了,但文本的复杂度却没减轻多少。
小结
记得上周末小编在ChatGPT主题沙龙上分享了自己关于ChatGPT的主题演示,当时未在最后补充所发现问题。其实早在一个多月前尝试其他文本时就发现这样的原文意思缺失等现象。私以为,未能解决这样的问题,对于自己还是上次分享也是一个遗憾。针对ChatGPT这样的情况,我认为ChatGPT在受控语言参与的译前编辑等语言处理方面理解和修改能力尚待提升,这一点也实非妄言,而是根据自身使用经历得出。当然,对于一两段小篇幅的专业性复杂段落,我们仍可以借助ChatGPT和受控语言来辅助译前编辑进行翻译。我们可以使用ChatGPT修改原文帮助自己快速理解原文,但译者也不能忽视判断原文和修改后文本意思是否改变和是否更易于理解。二者结合进行,方能使用好ChatGPT和受控语言,为产出较好的译文做铺垫。另外,冯全功教授也指出,“要想生成高质量的译文,只有译前编辑还是远远不够的,还需要译后编辑的参与,两者往往需要联合运用。”因此,我们作为译者,还需要提升自身翻译语言能力和掌握多方面翻译技术和手段,这样才能更好地驾驭ChatGPT等工具,给出合适的译文。
后续还会推出ChatGPT辅助CAT翻译、ChatGPT辅助长文本翻译等系列内容,欢迎大家关注,感谢大家!
参考文献:
Bussey, S.. Controlled Language[EB/OL]. (2020-02-20) [2023-07-21]. https://blog.andovar.com/controlled-language.
Chan, S. W. (ed.). 2014. Routledge Encyclopedia of Translation Technology (1st ed.)[M]. London & New York: Routledge.
O’Brien, S.. 2003. Controlling Controlled English. In EAMT Workshop: Improving MT through Other Language Technology Tools: Resources and Tools for Building MT, Budapest, Hungary. European Association For Machine Translation.
冯全功, 高琳. 2017. 基于受控语言的译前编辑对机器翻译的影响[J]. 当代外语研究17(2): 63-68.
特别说明:本文仅供学习交流,如有不妥欢迎后台联系小编。
- END -
翻译技术教育研究会ChatGPT专题小组致力于探索发掘ChatGPT为代表的AI工具在语言服务不同阶段的应用,进而提高从业者的工作效率。专题小组目前的主要方向是探索提示词(prompt)及流程变更如何提升交互效果;了解及推介Github等平台上的AI应用,AI应用嵌入语言服务流程及其优化方式。小组通过定期会议及讨论确定选题方向并通过文章、视频等形式分享探索结果,欢迎通过文章留言为大家提供建议意见!
原文作者:段明贵
推文编辑:田贝西
项目统筹:李林
指导老师:陈杲
资讯推荐
▶ 技术与工具
▶ 国际语言服务动态
| 翻译公司篇 | “收购狂魔”Keywords Studios
| 行业机构篇 | 加拿大联邦翻译局 Canada's Translation Bureau
▶ 专访
王巍巍:疫情背景下远程口译的挑战和应对
卢家辉、陈晨:口译员在远程口译中面临的主要问题
▶ 行业洞察
▶ 教育创新