Neural networks at WMT: machine translation competed with human

The Fourth Conference on Machine Translation (WMT 2019) organized by the Association of Computational Linguistics (ACL) was held on August 1-2 in Florence. The event was attended by major international developers, including the PROMT Company.

Every year machine translation specialists, representatives of commercial organizations and research groups from around the world compete with each other within ACL. They train their systems on data sets provided by the competition within a certain time limit. These, as a rule, are the UN texts and news. After the training is finished, the competitors make test translations with their systems and submit the translation output for expert and automated evaluation.

Over the past few years, conference participants have been actively introducing neural network-based translation systems. At the conference, a lot was said about the fact that neural networks technologies demonstrate an unprecedented increase in the quality of machine translation. In view of this, the organizers and participants of the Conference thought of reviewing and updating the metrics used for translation quality evaluation and systems comparison. For instance, it was suggested to evaluate the translation quality of the entire text instead of evaluating the translation of separate sentences.

Alexander Molchanov, Head of Statistical Research and Neural Network-based Translation, represented the company at WMT 2019 with a new PROMT Neural technology. He took part in a poster session, talked with colleagues and event organizers and introduced PROMT research and development in the sphere of neural networks to the audience.

“Today the quality of translation of non-specialized texts using neural networks is very high,” he said, “The main challenge now is the ability of training a neural network on different types of data from the customer: on specialized parallel texts of different size and glossaries. It is also important to be good in collecting thematically relevant material if a customer has no data set for neural networks training”.


Back to the list