Transitioning to a post-editing machine translation business model

向机器翻译译后编辑的商业模式过渡

2020-05-21 20:10 smartcat

本文共2305个字,阅读需24分钟

阅读模式 切换至中文

There is a reluctance across many LSPs to adopt machine translation. Many believe it will decrease the quality of work and will weaken the industry by taking away jobs. As someone who has been in the localization and translation industry for several decades, I can tell you that this resistance is not new. In 1993, when I introduced CAT tools to the agency I was working with at the time, I saw the same resistance to adopting new technology. Translators worried that relying on translation memory would make their work more repetitive, and decrease the pay and amount of work available. Twenty-five years later, the same arguments are being made about machine translation. Of course, machine translation is not a new concept. Its rising popularity now is due to recent advancements in the technology. Newer neural machine engines are remarkably much better than rule-based engines that were cutting-edge twenty years ago. When someone argues that MT engines produce poor results, the first thing I ask is when they last tested machine translation. Many in the industry are still basing their opinion on results from years ago, which are no longer valid. The reality is that machine translation is cheaper, faster, more secure, and increasingly better quality. LSPs that do not adopt this quickly dominating technology will not be able to compete in this new market. In this article, we’ll look at the arguments against machine translation and how they are changing with new advancements in machine translation. We’ll also outline a process for LSPs to adopt a machine translation process. Argument #1: When you factor in post-editing, machine translation isn’t faster or cheaper Even though newer MT engines are producing results that, in some scenarios, come close to matching human translations, there is still a warranted reluctance to make use of unedited machine translations. Most MT output requires post-translation editing. There are conflicting opinions regarding the efficiency of post-editing. Some translators make the case that translating from scratch can sometimes be faster than post-editing although this perception is being altered by higher quality MT output. Even when human post-editing is performed, the cost still tends to be at least 50% cheaper than the traditional translation process. Argument #2: MT is only appropriate in specific situations There’s a misconception that machine translation is only useful for getting the gist of a text when, in reality, it can be an extensive part of nearly any translation project no matter what source documents you are using. In most situations, it’s not a matter of whether machine translation is appropriate, but rather how much post-editing work will need to be done in addition to the MT. With the appropriate process in place, machine translation can be used to respond to most translation requirements. Without post-editing: MT can be used to help in content analysis like eDiscovery in the legal industry or for basic comprehension for internal communication. With minimal or no post-editing: it can also be used to process large non-critical customer-facing content repositories, such as support forums or eCommerce applications. With post-editing and additional QA steps: it can be used in most scenarios where translation is needed. There are still cases where machine translation is not recommended.  For example, the technology will not work well in situations where translations need to be highly adapted for a specific cultural message, like marketing materials.  Machine translation might also not be advisable for the translation of complicated legal and commercial contracts without a thorough review of the output.  In addition, machine translation without post editing and a rigorous QA process might not be appropriate for documents in critical sectors like biotech, pharmaceuticals, and equipment where lives could be endangered. Argument #3: MT engines are replacing translators As we see in other industries that are being disrupted by technology, jobs are not necessarily being lost as much as they are evolving. “The life of a freelance translator, once based on securing projects and clients, and establishing a reputation for quality, has now shifted to offering a variety of services.” Elanna Mariniello and Afaf Steiert explained in tcworld. Technology is creating new positions and requiring new skill sets. In our industry, translators are moving into the role of post editors, correcting the output of machine translations as opposed to completing translations from scratch. Translators have voiced concern over decreased pay and workload. However, this issue is much larger than the impact of machine translation. There is increasing pressure to lower the costs of most processes in the localization industry. Competition from lower-cost locales has had effected translators’ income, as well. To address these issues, the industry needs to strengthen its perspective on the value of trained linguistic professionals. Translators are not the only ones impacted by machine translation. The technology changes the pricing and language service capabilities which impacts clients, as well. When an LSP adopts an MT process, they need to be prepared to educate translators and clients alike. What follows is a step-by-step guide for implementing an MT process. How to implement a machine translation process We recommend four steps in implementing a machine translation process with your translation team: 1. Research and evaluate different engines 2. Choose a CAT tool that works with the engines 3. Update your SOP and educate your translators 4. Revise your pricing structure and educate your customers Choosing the right MT engines for your projects With many MT providers on the market, choosing the most appropriate engine for each project can be a confusing process. There are three types of machine translation engines on the market: rule-based, statistical, and neural. However, the majority of providers are moving to neural machine translation engines (NMT), which is widely considered the most advanced and fastest improving. NMT uses an artificial neural network to predict word sequences based off of a text corpora. This is similar to how statistical engines work, but NMT requires significantly less memory and learns more efficiently. However, there are cases in which NMT is still not available for specific language pairs. When this is the situation, in most cases, a statistical engine will be your best option. The option to use an engine that can be trained either interactively or in batches should be explored. In certain cases, they can improve the output and increase post-editing productivity — for example, when your source documents contain very specific terminology. For many LSPs, generic engines will be sufficient. Evaluate the quality of the engines We suggest researching popular engines and selecting three or four that work with the language pairs you need. Our partner organization Intento provides integrations between popular machine translation engines and CAT tools. They also conduct and publish studies on the quality of MT engines per language pair. The quality of an engine varies by language pair, so you will most likely need to use different engines for different projects. There may also be slight variations based on topics; however, with the best-known engines, these variations are limited. Once you’ve narrowed your list down to your top two to three choices per language pair, take the output of 50 or so pages, and get feedback from your translation team. If one engine is noticeably better for the current project, then that’s the engine you should use. If two or more engines produce very similar results, go with the cheapest option. Re-evaluate engines every six to twelve months Because engines rapidly evolve, the engine that produces the best quality may change. For instance, in the Intento January 2019 report, Konstantin Savenkov mentioned that “for 21 language pairs [that they tested], the best MT provider has changed since July 2018.” Retest different engines to ensure you’re still using the best one for your needs. Choosing the right CAT tool(s) Machine translation engines are now integrated into most CAT tools on the market. You want to look for a tool that’s going to give you access to many machine translation systems. Some tools lock you in into one or two engines that may not be ideal for your projects or language pairs. We also recommend using CAT tools that are optimized for post-editing and combine MT, TM, terminology management, and strong collaboration features to maximize post-editing and team productivity. PEMT tasks generally require large teams of translators, and seamless collaboration is vital. Implementing a post-editing process How you use machine translation may depend on the contracts you’ve signed with your customers. For instance, if they’re paying for a human translation, you could still use machine translation to offer suggestions to your translator that they can choose whether or not to use. (This feature is built into our platform’s CAT tool). MT suggestions can help a translator complete the project faster while leaving them free to translate using their own knowledge and discretion. However, your Quality Assurance team should be on the lookout for translators who are too heavily influenced by the output of the MT engine. In response to the decreased project costs and turnaround time (and improved quality of machine translation), we believe that the industry will continue to move to a complete machine translation and post-editing process. Having a process in place will prepare you for this fast-approaching future. There are two defined types of post-editing: light and full.  In a light post-editing process, the editor makes only minimal changes to increase the comprehension of the text. No stylistic changes or fluency improvements are required. This type of post-editing is generally limited to content used for internal communication within a company, or for communication with a short lifespan, like forum posts or emails. Light post-editing can also be useful in situations such as legal eDiscoveries where the translation requester only seeks content confirmation or for eCommerce product or service descriptions that will be retired.  Full post-editing is a more extensive process where the post-editor not only corrects obvious mistranslations but also improves the style and fluency of the MT output. Full post-editing is especially useful in situations where the MT process is new or for language pairs that have a lower quality score. Industry studies point to productivity improvements around 40% with full post-editing, although these numbers can vary greatly. Output quality should be negotiated with the client to decide whether additional editing is needed following post-editing. In 2017, an ISO standard for PEMT was published. Post-editing efficiency can be improved in conjunction with the use of translation memory. Training your translators to be post-editors The biggest challenge to post-editing is the resistance to the process from translators. Although I’ve found that the younger generation of trained translators has now been exposed to MT as part of their curriculum and more open to working as post-editors. Your current translators will need to be trained in the various types of post-editing (light and full). Several post-editing courses exist from companies like SDL Trados and Taus. This change in job requirements may not work for all translators; however, not all customers may be ready to move to a machine translation process either. In other words, you may still have use for translators resistant to post-editing. Educating your current customers on machine translation The biggest benefit of machine translation to your customers will be reduced prices for your services. However, even with lower costs, not all clients are readily sold on the new process. You need to be prepared to explain the advantages of the technology, as well as correct their expectations. Some customers may assume that a machine translation will be perfect and won’t understand the added expense of post-editing, where other customers will assume that machine translation is simply another added cost that will affect the quality of your results. Machine translation can create new product lines that were not previously cost-effective, such as translating comments on blog posts. In projects like this, expected translation quality may be much lower. New potential services, pricing structures, and expectations around the quality of a translation will require dialogue and education on the different roles that machine translation and post-editors can play depending on the requirements of a given project. Updating your prices When you successfully adopt a machine translation process, you should be able to increase your margin and decrease costs to customers at the same time. Start with your costs and margins. How one factors the cost of a translation will change with machine translation. In general, post-editing is priced considerably cheaper than traditional human translation. Pricing post-editing tasks is still a challenge for the industry for lack of updated productivity metrics. Where translators are most-often charged per word, post-editing machine translated content may be billed via time spent, corrections made, or by a review of the quality level of the post-edited content. The discussion is ongoing about how translators should be paid for post-editing work. It is very similar to the discussion that happened twenty-five years ago when we had to devise new pricing structures based on using translation memory tools. Updating your quality standards and assurance process Depending on the expectations of a given project, moving to an MT process may not change your quality assurance process. However, in some cases, a client may decide that an unedited MT output (at a much lower cost) is acceptable for lower priority content. As you edit your existing contracts or enter into contracts with new clients, these agreements will continue to be based on a combination of cost and necessary quality standards per project. If your clients use external reviewers, you will need to align with them around any changes in your quality standard agreements.
许多语言服务提供商(LSP)都不愿意采用机器翻译。他们认为,这将降低工作质量,减少就业机会,从而削弱语言行业。作为在本地化和翻译行业摸爬滚打几十年的人,我可以告诉你,这种想法并不新鲜。 1993年,当我将计算机辅助翻译(CAT)工具介绍给当时合作的机构时,该新技术受到了同样的抵制。译者们担心依赖翻译记忆库(TM)会增加工作重复率,从而降低工资、减少工作量。25年后,关于机器翻译(MT)的争论仍在继续。 当然,机器翻译并不是一个新概念。它现在越来越受欢迎,是得益于最近的技术发展。较新的神经机器引擎明显比20年前最先进的基于规则的引擎要好得多。 当有人争辩MT引擎产生的译文效果很差时,我首先会问他们上次测试MT是什么时候。许多业内人士仍将他们的观点建立在多年前的结果之上,而这些结果已不再有效。现实是MT更便宜、更快速、更安全,质量也在逐步提高。若不采用这一迅速占据主导地位的技术,LSP将无法在新市场上竞争。 在本文中,我们将讨论反对MT的3大论点,以及这些论点如何随着MT的新进展而改变。此外,我们还将概述LSP采用MT流程的步骤。 论点1:考虑到译后编辑,MT并不会更快速或更便宜 即使最新的MT引擎产生的译文质量有时接近于人工翻译,人们仍有理由拒绝使用未译后编辑的MT译文。大多数MT输出仍需要译后编辑。 译者们对译后编辑的效率持不同意见。一些译者认为从头开始翻译有时比译后编辑更快,尽管这一意见正在被更高质量的MT输出所改变。 即使在完成人工的译后编辑之后,翻译成本仍可能比传统翻译过程便宜至少50%。 论点2:MT只适用于特定情况 人们对MT有一大误解,即它只可用于获取文本主旨;然而事实上,它几乎可以参与任何翻译项目,与您使用的源文件类型无关。 在大多数情况下,问题不在于MT是否合适,而在于除了MT之外,还需要做多少译后编辑工作。若流程恰当,MT就可用来完成大多数翻译需求。 不需译后编辑:MT可用于文本分析,如法律行业的电子文件披露,或用于只需基本理解的内部沟通。 极少或没有译后编辑:还可用于处理大型的、面向客户的非关键内容存储库,如支持论坛或电子商务应用程序。 具备译后编辑,且附带质量保证(QA)步骤:适用于大多数翻译场景。 以下是不推荐使用MT的情况:  例如,需要高度调整译文以适应特定文化信息(如营销材料)时,这项技术就不能很好地发挥作用。  若不彻底审查译文,MT并不适合翻译复杂的法律文本和商业合同。  此外,若没有译后编辑和严格的QA过程,MT可能不适用于翻译生物技术、制药和设备等关键部门的文档,因为翻译错误可能危及生命。 论点3:MT引擎正在取代译者 正如其他受科技影响的行业一样,科技造成的工作岗位流失并不一定多于其正创造的岗位。 Elanna Mariniello和Afaf Steiert在TCWorld上解释:“一名自由译者的生活曾经建立在争取项目和客户、建立质量声誉之上,而现在已转变为建立在提供各种服务之上。” 科技正在创造新的职位,需要组合新的技能。在语言行业中,译者正转向译后编辑人员的岗位,他们负责纠正MT的输出,而不是从头完成翻译。 译者们对工资和工作量的减少表示担忧。然而,这个问题并不仅受MT的影响。在本地化行业中,降低大多数流程成本的压力越来越大;低成本地区的竞争也影响了译者收入。为解决这些问题,语言行业需加强对训练有素的语言专业人员的价值认识。 受到MT冲击的不仅仅是译者。这项技术改变了行业的定价和语言服务能力,从而也影响了客户。当LSP采用MT流程时,他们需要培训译者和客户。以下是实现MT过程的步骤指南。 如何构建MT流程 我们推荐以下四个步骤,助您与翻译团队构建MT流程: 1.研究和评估多种MT引擎 2.选择与MT引擎配合使用的CAT工具 3.更新标准作业程序(SOP)并培训译者 4.调整定价结构并培训客户 为您的翻译项目选择正确的MT引擎 由于市场上有许多MT提供商,为每个项目选择最合适的引擎可能十分令人困惑。 市面上有三类MT引擎:基于规则的机器翻译引擎、统计机器翻译引擎和神经机器翻译引擎。然而,大多数提供商都在转向神经机器翻译(NMT)引擎,它被广泛认为是最先进和进步最快的引擎。NMT使用人工神经网络来预测基于文本语料库的词序列。这与统计引擎的工作方式类似,但NMT需要的内存明显更少,且学习效率更高。 但是,在某些情况下,NMT仍然不能用于特定的语言对。此时,统计引擎大概率是您的最佳选择。 应该探索一个既可以交互训练,也可以批量训练的引擎。在某些情况下,如您的源文档包含特定术语时,它们可以改善输出并提高译后编辑效率。对许多LSP而言,一般的引擎就足够了。 评估MT引擎的质量 我们建议,您可研究当下流行的引擎,并选择三到四个可处理您所需语言对的引擎。 我们的合作机构Intento可集成目前流行的MT引擎和CAT工具。他们还研究每种语言对的MT引擎质量,并发表研究报告。 引擎的输出质量因语言对而异,因此不同项目很可能需要使用不同引擎。根据主题不同,输出质量也可能有轻微的变化;然而,最知名引擎的变化有限。 一旦您将清单缩小到每种语言对的两到三个最佳引擎,请拿出50页左右的MT引擎译文,从翻译团队那里得到质量反馈。 如果某一引擎明显更适合当前项目,那么您应使用该引擎。若两个或更多的引擎产生的效果十分相似,请选择最便宜的引擎。 每半年或每年重新评估MT引擎 因为MT引擎发展迅速,所以产生最佳译文的引擎可能会发生变化。例如,在Intento 2019年1月的报告中,Konstantin Savenkov提到,“自2018年7月以来,适用于(他们测试的)21种语言对的最佳MT提供商已经发生了变化。”重新测试不同的引擎,以确保您仍在使用最好的引擎来满足需求。 选择一种或多种合适的CAT工具 MT引擎现已集成到市面上大多数CAT工具中。您寻找的CAT工具应可访问多个MT系统。有些CAT工具会将您锁定在一两个特定的引擎中,而这些引擎对于您的项目或语言对而言可能并不理想。 我们还建议您使用针对译后编辑进行优化的CAT工具,这些工具结合了MT、TM、术语管理,且具有强大的协作特性,以最大程度提高译后编辑效率和团队生产力。译后编辑任务通常需要庞大的翻译团队,无缝协作至关重要。 构建译后编辑流程 您使用MT的方式可能取决于您与客户签订的合同。例如,若客户支付人工翻译费用,你仍然可以使用MT为译者提供建议,让译者自身选择是否使用MT(此特性内置在我们平台的CAT工具中)。 MT建议可帮助译者更快地完成项目,同时还可让他们使用自己的知识和判断力自由地翻译。但是,您的QA团队应关注那些受MT引擎输出影响过大的译者。 为应对项目成本和周转时间的降低(以及MT质量的提高),我们相信这个行业将继续向完整的MT和后期编辑流程迈进。合适的流程会让您为即将到来的发展做好准备。 有两类译后编辑:快速译后编辑和高质量的译后编辑。  快速译后编辑只做最小限度修改,旨在增加对文本的理解。不需要改变文体或提高流畅性。 这类译后编辑一般仅限用于公司内部交流的内容或短期交流的内容,如论坛帖子或电子邮件。快速译后编辑同样适用于在法律行业的电子文件披露。在上述情况下,翻译请求者寻求的只是内容确认,或是将被淘汰的电子商务产品或服务描述。  高质量的译后编辑流程更为复杂,不仅要纠正明显的误译,而且要改进翻译输出的风格和流畅性。 在MT流程是新建立的或语言对质量分数较低的情况下,高质量的译后编辑大有用处。行业研究指出,在高质量的译后编辑下,生产率提高了40%左右,尽管不同项目的提升程度可能相差很大。 应与客户协商输出质量,以决定在译后编辑完成后是否需要额外的编辑。 2017年,公布了一项针对译后编辑的ISO标准。译后编辑的效率可通过与翻译记忆配合使用而提高。 将译者培训为译后编辑人员 译者对这一流程的抵制仍是译后编辑所面临的最大挑战,然而我发现,受训的年轻一代译者现已把MT作为其必修课之一,且更愿意从事译后编辑的工作。 现在,您的译者将需要接受各种类型的译后编辑(快速的和高质量的)培训。像SDL,Trados和Taus这样的公司提供了几个译后编辑课程。 工作要求的改变可能不适用于所有译者;然而,也不是所有客户都愿意使用MT流程。换言之,您仍能雇用拒绝使用译后编辑的译者。 对现有客户进行MT培训 MT对客户的最大好处在于降低服务价格。然而,即使成本有所降低,也不是所有客户都愿意接受这个新方法。您需要做好准备,为他们解释该技术的优点,并纠正他们的过高期望。 某些客户可能认为MT是完美的,并不理解为何需要译后编辑,产生额外费用;而其他客户则会认为MT只是另一个会影响译文质量、带来额外费用的技术。 MT可以创造出新产品线,而这些产品线在之前根本不具备成本效益,比如翻译博文评论。在此类项目中,预期的翻译质量可能会低得多。 达成新的潜在服务、定价结构和翻译质量期望将需要对话和培训,以了解MT和后期编辑如何根据特定项目需求发挥不同作用。 调整定价 当您成功采用MT流程时,您应能在提高利润的同时降低客户成本。 从降低成本、提高利润着手。影响翻译成本的因素将随着MT引入而改变。一般而言,译后编辑的价格比传统的人工翻译便宜得多。 由于缺乏更新的生产力指标,对译后编辑任务定价仍然是行业的一个挑战。译者通常按字收费,MT的译后编辑内容可能通过花费的时间、对译文的修改或对译后编辑内容的质量审查来计费。 目前正在讨论如何支付译后编辑人员的报酬。这与25年前的讨论非常相似,当时我们需要在使用TM工具的基础上设计新的定价结构。 更新质量标准和质量保证过程 根据给定的项目期望,转换到MT过程可能不会改变质量保证过程。 然而,在某些情况下,客户可能认为未经译后编辑的MT输出(以低得多的成本)对于较低优先级的内容而言是可接受的。 当您编辑现有合同或与新客户签订合同时,这些协议将继续以每个项目的成本和必要的质量标准为基础。 若您的客户使用外部评审员,您在质量标准协议中的任何更改都需与他们保持一致。

以上中文文本为机器翻译,存在不同程度偏差和错误,请理解并参考英文原文阅读。

阅读原文