Artificial Intelligence
Samsung Electronics Sweeps Coveted Global AI Awards
11/30/2018
Samsung Electronics’ artificial intelligence (AI) capabilities are being recognized globally in a competitive field with top researchers all seeking to dominate. Samsung Research, the advanced R&D arm of Samsung Electronics’ device business, has won recent competitions, which will be vital in ultimately rolling out AI in more real-world situations than ever.
In October, Samsung Research’s R&D Institute Poland (SRPOL), in partnership with University of Edinburgh of the UK, won first place at the International Workshop on Spoken Language Translation (IWSLT), one of the world’s most renowned and longest-running automatic language translation workshops. IWSLT sits alongside the Workshop on Machine Translation (WMT) as the most prestigious competitions in the space. The IWSLT win is the second consecutive time that the Poland research center has topped the workshop.
Every June, IWSLT provides permissible training data sets and allows test run submissions from participants, evaluating the submitted runs with automatic metrics. This year, IWSLT featured two tasks: Low Resource Machine Translation and Speech Translation. SRPOL participated in the first sector, aiming to discover new technologies that can help overcome low-resource situations with scarcity in not only data but also in time and cost that are often required to use AI for multilingual translation.
Samsung’s task was using AI technology to translate text from Basque to English. The SRPOL team was first provided with limited amount of data – about a million Basque-to-English parallel data sets in the form of movie subtitles – to train and develop the algorithm model. Then, the team used this data to translate TED Talks given in Basque into English. Despite the difficulty of proposed translation direction and scarce data, when compared with the reference translations, SRPOL received the highest Bilingual Evaluation Understudy score (BLEU) of 26.21 – showing a high level of correspondence of the machine translation to the reference – and a significantly low error rate.