Machine translation copes with EU Babel
A machine translation engine developed for the needs of Finland’s Presidency of the Council of the European Union has proved its usefulness as a translation aid. Machine translation (MT), also known as automated translation (AT), is based on neural networks that learn from experience.
Last spring, the Prime Minister's Office participated in an MT project funded by the European Commission, in which MT software was customised for the specific needs of three countries making up the current trio of consecutive presidencies (Romania, Finland and Croatia). The aim of the project was to facilitate and speed up multilingual communications during the presidencies and to respond to the growing need for translation. For Finland’s Presidency, MT engines were developed for the language combinations Finnish–English–Finnish and Finnish–Swedish–Finnish.
The early results were so encouraging that the tool was incorporated into the translation processes of the language services of Finland’s Presidency of the Council this autumn. The translation produced by the engine cannot be used without revision, but it serves as raw material which a human translator can edit to produce a viable text.
The EU Council Presidency Translator has also been made available to other users to help them understand or draft foreign-language texts. The engine can provide translations of short bits of text as well as entire documents or webpages. Among the language combinations offered by the engine, the presidency language services have tested Finnish into English or Swedish and vice versa, with promising results.
Neural networks for fluent Finnish
Up to quite recently, machine translation yielded rather poor results for the Finnish language. Advances in artificial intelligence and machine learning, however, have enabled the development of automated translation systems that produce viable results even for small, structurally challenging languages such as Finnish.
The EU Council Presidency Translator works on the basis of neural networks. The texts produced by the neural engines come close to human language in terms of fluency, and their quality surpasses that produced by many of the older statistical and rule-based MT systems. The neural engines try to deduce the context of the sentence from its components, thereby improving the quality and accuracy of its translations. Further improvements can be achieved by training the engine with the organisation’s own texts or other texts in the same domain, although even the best MT engine cannot produce perfect, error-free text.
The development of engines based on neural networks relies in part on machine learning. Initially, a large quantity of bilingual text data is entered into the MT system. The engine learns to translate between the two languages concerned by processing this data with the help of algorithms. The MT engine for Finnish was trained with more than 20 million sentences and their translations. The EU Council Presidency Translator automated translation system was developed by Tilde, a Latvian translation and language technology company, which has been consistently successful in the annual World Machine Translation (WMT) competitions.