Machine Translation Post-Editing: Perspectives and Best Practices

With artificial intelligence becoming an integral part of everyday life, machine translation (MT) is continually improving in terms of speed, accuracy, and cost-effectiveness. Yet, it still has some way to go before it achieves parity with human translators.

Machine translation post-editing (MTPE) helps combine the best of both worlds—the speed and ability of MT engines to quickly handle large volumes of text with the skill and sensitivity of trained linguists.

Read on to find helpful tips for implementing post-editing in your translation workflows.

Perspectives on MTPE

Making the most of machine translation post-editing begins with understanding the perspectives of business buyers, translation vendors, and translators on machine translation overall. Each group has a unique set of considerations that impact the approach to MTPE.

The buyer perspective

Modern businesses have come to see machine translation as a viable productivity tool for entering new markets more quickly while keeping costs low. The advent of neural machine translation (NMT) has particularly prompted them to leverage their own language data for global growth. Today, the question for most global enterprises is not “if” but “how” they can integrate MT into their localization ecosystem.

Some of the challenges that buyers face when it comes to machine translation include:

• How to integrate MT into an existing localization workflow?
• What type of content is most suitable for MT?
• Which machine translation provider to choose?
• Which language pairs to focus on when using MT?
• Where to look for and how to obtain data for training custom MT models?
• How well does the MT data perform and how to clean data for custom MT model training?
• How to evaluate the performance of custom MT models?
• In what ways can MTPE be effectively used and at what price?

Narrowing the perspective down to MTPE pricing, one of the central considerations for large enterprises is how to allocate a budget for MTPE and use it efficiently in the long run. Although significant for the predictability of MTPE-related costs, cost-effectiveness is mostly not a goal for itself. On the contrary, saving costs on some machine translation projects with MTPE is considered a means of freeing up a budget to support further language pairs or content types for even greater scalability.

All these challenges make the human element of applying MTPE to localization workflows an important factor in the decision-making process. If a company does not have in-house expertise and resorts to outsourcing, it’d usually rely on its existing language service provider (LSP) or engage a new one as a consultant. The LSP would then deliver a proof of concept on all strategic, linguistic, and technological aspects of integrating machine translation and using MTPE.

The vendor perspective

The change in demand in the translation and localization market has caused many medium to large LSPs to include MT post-editing services in their offering. Not only does it make their service and product portfolio more attractive and stand out from the competition, but it also allows them to meet the goals of their customers faster and provide value to them in a continuous manner. One of the key MT aspects that LSPs still deem challenging is defining what makes good quality, but even more unpredictable is the question of pricing for MTPE.

Managing MTPE projects on the vendor side requires an understanding of:

• Available MT engines
• How each MT engine performs on different kinds of text and language pairs
• Which MT engine integrates best with the translation management system (TMS) used
• How familiar individual translators are with the MTPE process, etc.

An experienced post-editor can quickly turn high-quality MT output into a polished translation, but someone who is new to MTPE will need more time—and if the machine translation is sloppy to begin with, even the best post-editor may be unable to salvage it in a reasonable amount of time. That is why, if LSPs want to provide value to their customers, they need to have training in place for their linguists to continuously improve their post-editing skills.

Last but not least, vendors can also choose to offer multiple flavors of MTPE. Generally, these are light and full. Light post-editing focuses on correcting major errors and ensuring that the result is intelligible—if not perfect. Full post-editing strives to bring the MT close to the level of a human translation.

Nevertheless, light post-editing has proven quite complex to define, so many LSPs have been shifting away from this dichotomy in favor of offering a single MTPE service that is less confusing to buyers. Lionbridge’s Jordi Macias, VP of Machine Translation and Tech Vertical, goes even a step further by claiming that light post-editing may be destined to disappear as the baseline MT quality improves.

The translator perspective

Depending on the clients they work with and the types of projects they specialize in, translators may have little to no experience with MTPE. With machine translation being the new normal, many translators are still trying to adopt a clear attitude towards MTPE.

Forward-thinking early adopters have come to trust MT as a productivity tool which they need to be able to adjust to the new demand created on the translation market and complete translation assignments more quickly. The question of what constitutes fair remuneration for an MTPE project still plays a key role.

As translators gain experience with MTPE, they develop an understanding of how MT performs on different kinds of texts. This helps them estimate the effort that any given MTPE project would require from them. They also like to have access to a sample of the text that they can use to judge the quality of the MT output. They use this information to determine whether to accept or reject a project.

For example, translators are likely to accept MTPE projects involving legal documents, as state-of-the-art MT engines translate them reasonably well. When translators choose to accept a project, they can also use this information to decide what rate to charge.

Set yourself up for success

Perhaps the most important step in post-editing happens right at the start: making sure that you have a high-quality MT output. The better the original engine output, the less work will be required during the post-editing step, leading to faster turnaround times at a lower cost. Here are two points to consider:

Start with the source text

As with all translation projects, the source text should be carefully created or pre-edited to ensure that both the post-editors and the MT engine are working with the best. Early errors can compound and create problems down the line. That’s why it’s important to make sure that the original text has as few spelling and grammatical errors as possible. The terminology and formatting should be consistent. The source text should, at the very least, be prepared as if it were going to be handled by a human translator.

For the best results, you should consider preparing your text specifically for machine translation. Although MT is rapidly improving, there are certain steps you can take to increase output quality.

Generally, MT works best with input that is clear and concise. An ideal sentence should be under 20 words and have simple grammar. Complex sentences or headline-style sentence fragments do not work well.

MT also tends to struggle with nuance and likes to be as literal as possible. Avoid sarcasm (machines are really good at picking it up), avoid double-negatives (they won’t do no good), and where possible have the dates in a non-numeric format. 01/05/2020 can lead to some ambiguity: Is it May 1 or January 5?

Post-editing strategies

Once the MT engine produces its translation, post-editing can begin. How much post-editing is required varies from project to project, so it is important to always define your expectations early. The three main considerations are time, quality, and cost. Build your post-editing strategy around this by choosing the right approach.

Light post-editing (LPE)

With LPE, raw MT is only modified where absolutely necessary to ensure that the output is legible and accurately conveys the meaning of the source document. The post-editor should be especially mindful of errors that might hinder the document’s purpose or outright subvert it. Without review, raw MT can create embarrassing results, as one tech giant discovered recently. The editor should aim to make as few edits as possible. This approach can lead to fast and cost-effective results.

Full post-editing (FPE)

With FPE, raw MT is thoroughly reviewed and modified to ensure that there are no errors whatsoever. Where LPE focuses on the bare essentials of accuracy and legibility, FPE considers a number of factors, including but not limited to:

• Stylistic and tonal consistency within the document (and with other appropriate documents)
• The absence of all grammatical errors
• Appropriate cultural adjustments for the target language (such as idiomatic expressions)

A document that has gone through FPE should convince its reader that it was originally written in the target language. This approach is slower and more expensive than LPE but will achieve high-quality output.

Project-specific strategy

Requirements should always be tailored to the specific translation project. It can be useful to think of LPE and FPE as being on a spectrum rather than a binary choice. Pick and choose your post-editing priorities based on considerations of time, cost, and desired quality. It can be effective to selectively prioritize certain segments that have a higher business value than others.

Another option to consider is bypassing post-editing completely. For certain projects, this is possible, for example with internal documents where MT output is expected to be good and the consequences of bad translation are negligible.

Post-editing tools

To help with post-editing, both editors and managers should be aware of the various tools that can help. Virtually all CAT platforms now offer support for post-editing. Here are a few tools to keep in mind:

1. Set up terminology management systems to help post-editors ensure consistency. This includes translation memories, term bases, and any useful reference documents. These should be kept up-to-date for future projects.
2. Use machine translation quality estimation (MTQE). In Phrase, this AI-powered feature can provide quality scores for all MT output, giving editors an indication of how much work is required for each specific segment. This can help linguists and project managers identify segments that they should prioritize for editing.
3. Use a QA tool, either integrated or standalone. These will help dynamically identify any issues in the original output that were overlooked or new errors that were introduced during the post-editing step.

Post-editing training and qualifications

Forward-thinking linguists have come to trust machine translation as a productivity tool which they need to be able to adjust to the new demand created on the translation market and complete translation assignments more quickly.

Still, the post-editor isn’t necessarily a translator. Although there is some overlap in the nature of their work, the exact skills required differ in a few important ways. In 2017, ISO 18587 defined some of the key aspects of MTPE, including the specific skills and competencies of the post-editor.

You can expect best results from qualified and experienced post-editors who have practical experience working in the specified language pairs, with the type of content, and with the relevant tools. In that direction, it can be helpful to train post-editors for specific tasks before they start on a project.

Post-post-editing

To help improve the results of MTPE, it is important to continuously evaluate the process and results using data and feedback. Consider post-editing to be an iterative process that can be improved with time and experience.

A range of tools is available to help with post-editing analysis. For example, Phrase can help calculate post-editing effort, which shows how much work was required by linguists to finalize the translation. This information can be as granular as required. Knowing, for example, that certain segments require a disproportionate amount of post-editing can help future projects: perhaps the source text can be adjusted, or post-editors can be provided with useful reference documents.

Besides gathering data, it is invaluable to proactively seek out feedback from all the key stakeholders. This can include the content creators, post-editors, clients, and project managers. Ask them about their experience with the project and identify what worked and what can be improved.

MTPE has become the new normal

With continuous improvements in quality, MTPE is now one of the core advantages of machine translation as a viable alternative to translating from scratch. MTPE lets you combine the speed of machine translation engines in handling large content volumes with the skills of trained linguists, so keep learning best practices to make the most of it in the long run. 

Choose the optimal translation engine

There is a large number of MT engines to choose from and new engines are being developed all the time. Not all engines are created equal, some simply perform better than others. Some are more suitable for specific language pairs or subject matter (domains). Choosing the most effective engine for your project can save a lot of time and effort.

Consider a range of generic MT engines and evaluate them using samples or past experience. Although creating custom analyses can be a time-consuming option, it can lead to cost-effective solutions in the long run. Another option is to consider a customized translation engine, trained using your own data. This will generally produce high-quality results for the content that you are used to working with.

Phrase offers a unique solution to the problem of choosing the right MT engine. Developed to dynamically select the best MT engine for your content, it considers the text’s domain, source and target language, and looks for an optimal MT engine based on past performance. It will always automatically select the best engine for your content.