How do you correct?

Hi! We all know that we have to dump millions of sentences in a neural MT. OK. But what happens when the machine does a translation mistake? How do we correct th rule so that it does not make the same mistake all the time? By dumping another 40 millions of sentences? :slight_smile: Seriously.

Which mistakes and rules do you mean? In most CATs you can define custom rules to handle common mistakes.

For solving it at the NMT-level: If you can afford it then train a deeper model with more learning data. But you can just update the working model with incremental training. Mix the new data with an initial part to avoid the catastrophic forgetting and continue the training process. With the new tensorflow version it could be possible to inference and retrain at the same time. There are also solution to automate the correction with an interactive NMT.

For the case that you can’t retrain the model, then you could train new a model for automatic post-editing. This additional model is set behind the initial translation model.