Methods

• One thing I really like about NLP is that there are a lot of standard Experiments.
• For example, the brown corpus has over a million words of text, and they've been marked up as parse trees. So, if I want to compare my parser to other parser, I just use Parseval on this corpus.
• Another example is the MUCs where different systems complete on an unseen data set.
• This prevents training on the test set.
• Message Understanding Competitions, Summeval, Parseval, Semeval
• Standard Information Retrieval Metrics of Precision, Recall and F-Measures
• Precision is number-correct/number-guessed.
• Recall is number-correct-guessed/total-correct
• An F-Measure combines these two by, for instance, multiplying thus giving a good score.