7 Steps for Successful MT Implementation
Avoid big, expensive mistakes by following these steps to a successful machine translation implementation.
1: Select the type of content you want to translate with MT
You have to be selective about the content you translate as not all content is suitable for MT. The less creative or literary the content, the better. If you are localizing marketing content or culturally-specific texts, MT may not be the best choice. This may change in the future, but for now it’s better to be safe than sorry.
2: Check the Personal Data Policy
3: Train it with your data (if possible)
If the MT can be trained with your data, do so. To achieve good results you would need around 100k segments. There are many options for training data such as publicly accessible corpora, building your own, or buying corpora from others. The data used for training should be relevant to your translation needs and of high quality.
4: Select a team of post-editors
For the best MT output, you need post-editing. Choose a team who has training or experience with MT post-editing or make sure that they are open to the idea of it. Post-editing is a different ball game to translation and requires a specific set of skills. MT post-editors need to be able to make quick decisions on whether to use or delete an MT output. They also need to be detail oriented - an MT output can be grammatically perfect but may not be the correct translation of the source. Don’t force your translators to do something they are not completely open to or not trained to do. You don’t want to ruin your relationship with them nor do you want to end up with mediocre output.
5: Run samples before deployment
The larger the samples, the better. The first attempts may yield disappointing results, which are not worth sending to your post-editors. You can compare a human translation with an MT output and generate an edit-distance report. If the result is good, go for it. If not, you’d better hold on and verify what can be improved. Be ready for different results between language pairs as some pairs perform better in MT. For example, Japanese and Finnish have a long way to go before they reach the same level of quality as Italian and Spanish.
6: Agree on a pricing model
Be sure to set a price up-front, and involve all stakeholders in the decision. This way you can avoid unpleasant pay disputes. Take into the type of content, the language pairs, the technology you are using, the results of the sample tests etc. If you want to know upfront how much you will be spending, you can use historical performances (not fully reliable) or technologies that are able to generate some kind of machine translation quality estimation (MTQE). The result has to be then compared with the actual post-editing effort to verify that the estimation was correct.
Once you have deployed, keep in mind that the results may not meet your expectations right away. Some projects will be good and others a disaster, but don’t be discouraged. The technology will get better and better over time thanks to training and tuning. As time passes you will see the advantages, and your team will wonder how they ever managed before implementing MT.