In the past few weeks, I’ve had the opportunity to work with clients who want to explore using Machine Translation (MT). These clients need solutions for content that is more complex or demanding than online instant translator tools can handle. MT is fascinating, it can be powerful, and it is all the rage right now in this faster/cheaper world. However, it isn’t always (well, ever) a silver bullet. There are some simple criteria that can help businesses decide if they really want to go there. And I believe in simplification!
You’d want to talk to your LSP about MT if and when:
1. You need it fast
A human translator gets tired, needs a snack break, and only works 8 hour days. A regular human translator can churn out more or less 2,000 words/day of translated content. On the other hand, a machine can run overnight, it can process a lot of content at a time, and it can do all that very, very quickly.
2. You have a ton of content
High volumes warrant the considerable investment and up-front setup work. I’ve heard experts say that you’d need 500,000 or more words (per language) to justify starting down the path to investigate, customize and deploy an engine. And if you don’t have an expert in house, you will need to pay the LSP for this! It’s an investment. So, get out your pen and Excel sheets and do an ROI.
3. Some of your content would go untranslated
This would be non-business-critical content such as user reviews, tweets, and blogs. Also, these materials have short shelf-lives; they expire after a few days or even hours. Since it is not lasting, durable content you could surmise that there might be very little (or undefined) ROI to translate it. Yet, it could have more value for users when it’s in their first language. These global users would be more likely to find it, read it, and benefit from it if it’s in their language.
4. You have past translations
Completed translations are used to train engines so that engine can more capably handle your subject matter. Already-translated content can be used to create TMs, which can be used alongside the MT process to increase quality and reduce post-editing. That translated material has to be good though. Garbage in, garbage out. (Google has made use of the huge amount of publicly available translated content on the web, and has, arguably, the largest corpus of bilingual data to work with).
5. Your content is not specialized
If there is not a lot of bilingual content in your domain (heavy machinery, legal, financial, etc.) then engines are hard to train. The more specialized it is, a large body of material becomes more urgent. A multilingual glossary can’t hurt you either.
6. Your language pairs are covered in the MT world
Chinese to Finnish, for example, may not have a well-developed MT solution. Standard language pairs (English to any European language, and any European language to English) are available and in wide use.
7. Your technical writers use a Controlled Language (CL) approach
This is about writing the source in such a way that a machine, lacking the judgment and creativity that a human possesses, can process it correctly. In CL sentences are short, terminology is used over and over, and straightforward sentence constructions are used. I’m talking about subject, verb, object sentence order, less or no passive voice, and the use of clear pronouns. CL is hard to obligate because writers hate it (maybe I’ll blog on this), but there are some tools out there to help.
8. High quality is not critical
Unless you have a bilingual human post-edit, in which a linguist fully reviews the machine translated content, then you cannot guarantee perfect quality. If you skip the human edit, you have to be willing to accept some major comprehension and grammar mistakes.
9. Cost is an issue
Last but not least, MT is about reducing costs by letting a machine translate content. Often, when many of the above conditions are also met, MT can reduce costs by up to 50%.
10. You have some time to spend
I know this one may seem ironic since you are choosing MT to save time. Let me explain. You WILL save time if you spend it first: set aside time for planning and preparation. You will need to exchange files and information, review and understand the analysis, collaborate with a vendor on tool, process, quality standards, risk mitigation, specifications, training people, etc. (I can go on). All this planning matters so you can reasonably predict that your MT project will be a success, and if a vendor suggests otherwise you should probably keep shopping.
In my experience you need to be able to check ½ of these or better to entertain dreams of MT.
Am I missing any additional situations or parameters where MT would be viable?
If you are interested in knowing more about the types of MT, please see my earlier blog on the types of MT (rules-based or statistical).
Can you think of any other things to consider before you embark on MT? Did I miss any of the key criteria for success?