Today's lecture will cover some of the key models that have been applied to Machine Translation in the past 20 years and explain some of the challenges posed by this task.
Building on the EM algorithm from last week, we will look at how word alignments can be inferred from parallel text, how early statistical approaches built on these to make simple statistical machine translation systems in the 'noisy channel' paradigm, how data driven approaches then found ad hoc solutions to many of the challenges of MT during the 2000s and finally how Neural Machine Translation has changed the types of problems we work on today.
- (Intro ) Hands on task
- (Slides) Word alignments
- (Videos) part1, part2
- (Notes ) Word alignment notes
- (Slides) Machine Translation (part 1)
- (Slides) Machine Translation (part 2)
Todays homework assignment involves implementing some word alignment models.