- Neural networks are the main reason machine translation has improved so significantly.
- Language learning tools won’t replace teachers: immersion and contact remain essential.
Europe covers barely 2% of the earth’s surface, but it packs in more than 100 languages. The continent’s inhabitants have always needed ingenious ways of communicating with one another, and now they are being assisted by amazing digital tools. Europeans have created online dictionaries like Reverso and LEO (from the Technical University of Munich), along with machine-translation and learning systems including Babbel, Duolingo, Busuu, Mondly and ABA English.
Language tech has become a vibrant industry partly due to the advent of new processes. The initial approach, developed in the 1950s, was simple. “The most basic systems translated word-for-word using a bilingual dictionary,” says Pierrette Bouillon, dean of the Faculty of Translation and Interpreting at the University of Geneva. Next came the corpus-based statistical approach. “The machine learned by using existing translations, dividing the text into segments for translation, and then choosing the most likely translation for each segment.” Nowadays, the technology uses neural networks. Bouillon explains: “The machine still draws on a corpus of texts, but it uses deep learning. Words are represented digitally. The system can establish relationships between the words and better generalisation. The notion of context is improved. It doesn’t choose the translation of a word based on the few words before it, but considers the entire source and target sentence.”
Germany’s DeepL is a leader in these neural networks. Its initial success came with the Linguee translation search engine (more than 10 billion searches since its creation in 2010). In 2017 the company launched a service in seven languages, intending to compete with Google Translate. “DeepL’s Translator is a huge success for two reasons,” says François Yvon, a professor at Paris-Sud University and a researcher at France’s National Centre for Scientific Research (CNRS). “In contrast to Google, which builds on the data generated in its various applications, DeepL extracts from its high quality base of information (like official texts from EU institutions), which the company also uses for its translation search engine, Linguee. Second, its interface is a cut above the rest in its seamlessness and response time.”
Voice recognition and adaptive learning
Language-learning applications are at the forefront of adaptive learning methods, which create educational materials to address the needs of each learner. “We use voice recognition to determine a student’s level,” says Marc Vicente, CEO of ABA English, an online school. “The method includes three key elements. Our students stream videos with scenes from everyday life. They practice with grammar, pronunciation and listening exercises. Then they get explanations from a personal teacher.” With one million active students every month, the Barcelona-based company has extensive operations in Brazil, Italy, Spain, France, Mexico and the US. With more than 80 employees and an expected net income of €12 million in 2018 (up 25% from 2017), ABA hopes to grow its presence in Russia, Turkey and China.
Lalilo, a start-up that helps learners of English and French with reading and writing, goes even further. “The best way to learn is by reading aloud as often as possible,” explains co-founder Laurent Jolie, a graduate of Paris’s École Polytechnique, “But in a big class with only one teacher, students can’t all practice reading aloud.” With Lalilo, students can read out loud, even without a teacher right next to them. “The machine can analyse their speech and recognise their pronunciation errors. It can then adapt the exercises.”
The end for professionals?
Most of the tools developed in Europe do not aim to replace professionals, but to complement them. Says ABA’s Vicente: “Applications that promise to teach a language with 10 minutes of fun exercises a day are not enough. Immersion and contact with an expert are essential.”
Meanwhile, the professionals acknowledge that translation and interpreting tools are not perfect. “If it involves a less common meaning of a word, the system will get it wrong almost every time,” Yvon says. “And machine translation tools are incapable of producing a coherent text in its entirety.” Bouillon notes, for example, that automatic translation neural networks are prone to omissions and punctuation errors. “These tools save time and provide a good foundation, but post-editing by a professional is always necessary.”
The language-processing industry can look forward to a bright future. The global machine-translation market is estimated to hit $983 million by 2022 (annual growth of 14.6%), and online language learning is expected to grow by an average of 19% per year between 2017 and 2021. “The number of researchers and people working in the industry could increase throughout Europe as a result of the European Parliament resolution on language equality in the digital age,” Yvon points out. The European Parliament, which communicates in 24 official languages, calls for the development of new research and educational programmes on digital communication and language technologies, emphasising how they benefit growth and society.