Workflow management is considered very important from a business perspective. It can save translators and their clients time and money. The basic workflows discussed in this article are 1. Human translation – human revision and proofreading; 2. Use of a translation memory (TM) – human revision and proofreading; 3. Use of neural machine translation (NMT) – human (post-) editing and proofreading. Of course, there are numerous variations on these themes, such as the splitting of content among several translators because of time constraints, which is then merged and preferably revised by the same reviewer; localization or adaptation to a specific audience, e.g., adapting a document for plain language or changing a legal document from common law to civil law or vice versa, and review by a reviewer.
Ideally, a good human translation will be accurate and sound natural. It takes time and uses brain power in the traditional manner. In principle, it will be localized and adapted for the readers by the translator or the reviser. The translator may use dictation, which is essentially a traditional way of translating equivalent to using a computer to type. After carefully reading the client’s instructions, the ideal workflow of a human translator may include:
Many translators now use a Translation Memory (TM) program to prepare either a first draft or a close-to-final product. This type of program can be pricey, and it can take a lot of time and effort to learn how to use it, enter the data and build up memories. Once the memories have been stocked with the client’s preferences, the translation should be “automatically” localized and adapted for the readers.
Neural Machine Translation (NMT) programs are either free or available with a subscription. Although it is a sound choice when deadlines are tight, it is criticized by some people because of the difficulty of revising it. You don’t have access to the database so you have to do more checking if you have doubts about the translation as compared to a text produced with a translation memory tool. Although it looks good superficially, many aspects need to be checked, including consistency and accuracy. Moreover, it may sound wooden and unnatural at times. It usually needs to be localized and adapted for the readers. And you still need to follow the steps outlined above for the human translator.
Nevertheless, NMT is an important technological tool that took many years of research to develop. The breakthrough was the use of the Canadian Hansard to create a reliable database and the realization that statistical analysis was the way to go rather than rule-based analysis as explained in The Walrus by Christine Mitchell (March/April 2022, “How Canada Accidentally Helped Crack Computer Translation”). Hansard is the official record of what transpires in legislative debate in Ottawa. According to Christine Mitchell, a copy of 14 years of data from parliamentary debates in French and English (over 100 million words) somehow wound up at IBM’s headquarters in New York in around 1988. American researchers took a different approach than Canadian researchers who were doggedly pursuing a rules-based theory. The Hansard database made it possible to successfully test the statistical-based theory which has subsequently led to a sea change in how we approach translation. Access to good data is the key to success for both a Translation Memory and an NMT system.
Post-editing tricks for NMT
A common adage among translators in the past was, if the translation is really bad, it is better to start all over again from scratch. This still holds true. If a machine translation is bad, ignore it. On the other hand, NMT can often save you a lot of time. In fact, the major advantage of NMT is its speed, which gives the translator time to check everything properly and work on the style.
Sometimes, when the document involves a lot of stylistic features and is fairly complex, it is best to reread the automatic machine output in English first in order to improve the English writing style and then, second, compare it with the original French to verify that the meaning is accurate. Lastly, the automatic read-aloud function in Word is surprisingly helpful as is Antidote for picking up errors.
In conclusion, for a streamlined workflow method: Step 1. Gather your documents and references and read the client’s instructions; Step 2. Use a translation memory (TM) or a neural machine translation (NMT) program to prepare a draft; Step 3. (Post-)edit; Step 4. Run a spellcheck and grammar check; and Step 5. Read aloud or use the automatic reader function.