Search code examples
nlpstanford-nlpcontext-free-grammar

what are the methods to estimate probabilities of production rules?

I know that n-gram is useful for finding the probability of words,I want to know, How to estimate probabilities of production rules? How many methods or rules to calculate probabilities of production rules?

I could not find any good blog or something on this topic.Now I am studying on probabilistic context free grammar & CKY parsing algorithm.

Solution

  • As I understand your question, you are asking how to estimate the parameters of a PCFG model from data.

    In short, it's easy to make empirical production-rule probability estimates when you have ground-truth parses in your training data. If you want to estimate the probability that S -> NP VP, this is something like Count(S -> NP VP) / Count(S -> *), where * is any possible subtree.

    You can find a much more formal statement in lots of places on the web (search for "PCFG estimation" or "PCFG learning"). Here's a nice one from Michael Collins' lecture notes: http://www.cs.columbia.edu/~mcollins/courses/nlp2011/notes/pcfgs.pdf#page=9