Consultancy on Machine Translation

MT tailored to your business case

The idea of an automated translation engine dates back to the 50s. Machine translation evolved in line with computing capacity, leveraging different approaches and technologies. It all started with rule-based machine translation (RBMT), using grammar rules and dependencies; it then morphed into statistical MT, leveraging parallel corpora made available on the Internet. But meaning remained elusive… until Tomas Mikolov’s breakthrough in 2013: the vord2vec algorithm he developed finally allowed for “meaning” annotation through embeddings.

Google saw potential and developed the multi-head transformer algorithm, described in great detail in the famous article “Attention is all you need” published by Google AI scientists in the NIPS proceedings. This led to the birth of Neuronal Machine Translation engines.

The Neural multi-head Transformers’ breakthrough boosted the BLEU  score to the point where the post-editing effort notably decreased. Some scientists went so far as to claim “human parity”. But was it so?

Stock engines are trained on general journalism, and you need to customize and train custom MT engines to perform better on your domain-specific data.

The “intelligence” of your engine depends on your degree of preparation. Even if “meaning annotation” (aka contextual information embedded in the text) has improved Machine translation output, results are too often literal and reflect the most frequent usage. An MT engine cannot process terminology; it cannot spot inconsistencies or errors in the source; it often fails humorously when the source is ambiguous and always requires a human post editor to tidy things up.

But MT has grown into a sophisticated and helpful technology.

An expert can help you understand:

  • when MT is most advised and how to use it;
  • which engine is best suited for a specific language pair;
  • how to evaluate and measure the output;
  • whether your data is indeed good enough for the particular domain;
  • how to clean the data;
  • how to prepare the data for training;
  • which metrics to apply (BLEU alone is not enough);
  • and, finally, how to measure the post-editor’s productivity.

To transform high hopes and expectations for Machine Translation into reality, you need a trusted partner. ASAP Globalizers can help you handle projects, test pilots, experiment with data and engine evaluations, and set up trials. And we can help you with the post-editing too.

Do you want to know more? @@book your consultation@@.