At the invitation of ICSA, Professor Feng Zhiwei delivered a lecture on machine translation, entitled “Machine Translation & Its Types: PBMT, SMT and NMT,” on May 11th, 2020. The lecture was hosted by Professor Han Ziman.
Professor Feng first reviewed the history of translation, pointing out that translation technology would be an effective way for overcoming language barriers, in view of the multilingualism we encounter today in the Internet age. Professor Feng also looked back on the history of machine translation, introduced Peter Troyanski’s theoretical thoughts and demonstrated with many pictures the first experiment in the field, showing why machine translation is considered as the pearl on the crown of artificial intelligence. Professor Feng then provided a detailed introduction to the three types of machine translation: PBMT (Phrase-Based Machine Translation), SMT (Statistical Machine Translation) and NMT (Neural Machine Translation), and with examples in multiple languages, including English, Chinese, Japanese, German and Russian, explained how these machine translation systems work.
With the help of Vauquois’s MT Pyramid, Professor Feng focused on PBMT and its two approaches: Interlingual Approach and Transfer Approach. He pointed out that rule-based machine translation relies on large-scale bilingual dictionaries and man-programmed grammatical and transformation rules. SMT treats translation as a process of decoding, in which machine translation becomes a noisy channel that can be decoded by different channel models. Thus, statistics-based machine translation requires language models based on monolingual corpuses and translation models built on bilingual corpuses. As an emerging method of machine translation based on neutral networks, NMT also requires bilingual corpus training as SMT does. However, NMT presents better translation performance without phrase and rule tables, which are necessary in SMT, because all its calculations depend on unsigned true value.
At the end of the lecture, Professor Feng reminded the audience of the the problems in machine translation, for example, inaccuracy in the translation of polysemy and acronyms, word omissions, low accuracy of non-English translations as well as the scarcity of corpus resources, despite the achievements in the field.
During the Q&A session, Professor Feng answered questions on machine translation and natural language processing and learning. Though delivered in simple language, Professor Feng’s profound lecture is considered to be very instructive and informative by the more than 500 attendants.