Machine Learning

• In the early 90s a shift began in NLP.
• Toy systems were replaced with systems that really worked.
• This was heavily influenced by the Message Understanding Competitions (MUCs).
• The idea was that large data sets could be marked up, and then machine learning could be used to learn essential parts.
• This could then be translated to unmarked text, producing a useful system.
• This also took advantage of machine assisted learning where an expert would mark up text.
• This work inspired other systems that learned from large corpora.
• For example, you can learn to parse by examining large copora of syntax trees.
• You can even learn grammars.