FOCAL STORY • ᴲ㷞ഊ 2019 UMAGAZINE 20 • 澳大ᑓゆ 12 ࠶UM $"5ೀ⠙⣵؝ ۵ˍۜ ⦾研發6. $"5ⅰ⏲ᇤಾ㡕 〘┰⏲ಾ ژ 研،ಾ ژ ↷ נ ˷ ⾾ㆯㆺԖ٨6. $"5ⅰ࣮ ಾ⠗ᮉ┰ ᧈ䢬㐫ˍዞ ڌ έᕔ࣮ 翻譯☲⚕ⅰ 翻譯ަ ㇰ䢬㐛̏ ᒴ͒ ژ グᐵⅰ研 ᑚ ܗ ǎǗ↷ נ 䢬࣮ 翻譯˷⾾ᕔՇ個 㬦䢺ύ⿼ ݶ 成䢬 ګ ύ⿼ྦ 翻譯ⅰ ゆ۴䢬ᴿྭ 成譯ᐵ䢻⡈ᆒՑ㭓ⅰ研 䢬ൄ ᒴή࠴࣮ 翻譯☲⚕ύ⿼ 譯ᕲⅰ能؛ǎǘ͒ ⿼㕱㑄 Ǘ۸ᕔዞϫ ᬞ⌒ༀ㭗ǍՑۜ ⡁ӯѠⅰ翻譯ަ ㇰ䢬 ᆾ能⇜ᢩ ㍴ أ 翻譯̿ ε䢬㐛ឺ͒ мᆾᕐឝϓǎǘ ،ಾ ژ ⻠⍵䢬6. $"5⦾˞₅˔♿ 大૦ᐨⅰ࣮ 翻譯☲⚕ˍឺ 䢬㓃ᦳ 201 ⁱ؝ቄ大⯂ḿḑ大ಾዞ֦ ⅰ ˍ個ᙉᗖ䢻ᴿ⡈ॺ201 䢬(PPgMe Պ܅֦ ᑓⅰᙉᗖ䢬⡈˜ᐅᗈ⨏ତ䢬 㧧⮍ዞ ̜ڌ 譯ᐵⅰަ ㇰ䢬् ᢪ6. $"5̏ 㣍⾾⦾ᒿФ㐫ǎǗॺԖ٨ᮉ┰ ᧈⅰ㦦৴˔䢬۸ᕔ˗ᑘᕍᑓⅰᙉ ᗖ ݶ ᑚᧈ䢬ᆾ能⌒Й࣮ 翻譯ⅰ譯ᐵ ㇰ㕵ྲ͞ ዞ ڌ 䢬ⓟ翻譯̿ ഭᬞ⌒ ༀ⦾ᨢᔊༀⅰ⾾ᥐǎǘ ܒ 䢬㷊㍽ᐑ እἐ㦦/-12$5研॰㡴䢬㈫᪷ ⬷大 ಾε⏲研̳ ᨢǎ،ಾ ᳤ژ Ր中ˍ䢬͒ ⦾᪷ ⬷研ˍ㈭֯ ᖾ(PPgMeᕍᑓ Improving UM-CAT’s Machine Learning Algorithms Liu Xuebo, a PhD student of computer sciences in FST, is also involved in the development of UM-CAT. He is mainly responsible for improving the translation quality of UM-CAT by improving its machine learning algorithms, which is also the subject of his PhD research. He explains, ‘Machine translation includes two steps: understanding the source text and generating a translation. My work is to study how to improve the machine’s ability to understand the text. Only by ensuring a high degree of accuracy will the system become a truly useful tool for the translators. And only then will they be happy to use it.’ According to Dr Liu, like the majority of machine translation systems in the world, UM-CAT is based on the architecture developed by the University of Montreal in 2015. However, in 2017, Google launched a new architecture, which greatly improved the quality of translation. So UM-CAT must be updated to keep pace with the changing technology. Liu says, ‘In terms of improving the algorithms, only by using the latest architecture and method can we ensure that the quality of the translation meets translators’ requirements for accuracy and readability.’ In the same year, the team from the NLP2CT, led by Prof Wong, visited Tsinghua University. As a member of the team, Liu studied the latest architecture developed by Google ،ಾ ژ ㈫᪷ ⬷大ಾε⏲研̳ ᨢ ᢚ㢟⦾㷊㍽ᐑእྕ Liu Xuebo works with researchers at Tsinghua University Ao Chi Hong with Prof Wong Fai
RkJQdWJsaXNoZXIy MTQ1NDU2Ng==