If you haven’t heard about Adaptive Machine Translation, you haven’t been paying enough attention lately. It became a bit of a buzzword ever since Lilt was launched, and SDL announced the new version of Trados Studio would include adaptive MT features (which Lilt welcomed!).
Adaptive MT is a fairly new technology that basically allows your MT system to learn from correction input on the fly. It may not seem like a huge deal to some, but let’s consider the context and how Statistical Machine Translation works, while trying to keep it simple.
An SMT system has two main components: a translation model and a language model. The translation model contains bilingual texts in a phrase table, together with their relevant scores or weights (if you have more than one possible translation, say, for the same phrase, the one with the highest score will be probably picked). This is where an SMT system takes translations from based on the probability of a target string being the best translation for the corresponding source string. Many translations are generated, not just one, and then the most likely one to occur in the target language is selected with the help of the language model. The LM is essentially a collection of monolingual texts in the target language – if a translation candidate can be seen in the LM, chances are that it’s good. Resulting translations are entirely based on probability.
Phrase table example
Now, traditionally, to “teach” your system how to translate better, you train it. This means you add new good quality translations to your phrase table and the scores are reevaluated to make sure the best candidates are picked in the future. The training process is an expensive operation as it is a system rebuild and it may take a long time to complete, of course depending on volumes and processing speed. Training a system with dozens of millions of words may take days.
Enter Adaptive MT. The premise is that this training is not required because the system learns instantly from, for example, post-editors corrections. Let’s break this down:
- MT output is generated
- A correction is received
- MT system “learns” the correction
- The system reduces the likelihood that the same error will occur again in its output
From Adaptive machine translation patent US 7295963 B2
How does this work? On the one hand, machine learning. This is what allows the system to automatically learn translation correspondences. On the other hand, two language models: a static one, which is a larger collection of texts, used to generate translation as close to human translations as possible; and a dynamic one, which is updated with the bits that are learned during, for example, the post-editing process. The system gives more weight to recently learned translations. An algorithm in the background reassesses the weights in the translation and language models, so the system keeps adapting as it learns from new corrections.
Some recent tests suggest adaptive MT technology helped human reviewers increase their productivity maintaining the same level of quality. It seems only logical that it does, as the paradigm shifts from a static source of translation suggestions to a dynamic one that keeps improving with input. Will adaptive systems be the new standard? It will certainly be interesting to see the impact this technology will have in the localization industry.