I am implementing the Inside-outside algorithm to estimate the parameters of a PCFG based on the train corpus. One observation (on existing implementations) is that, the grammar is converted into Chomsky normal form (binarization) prior to the Expectation-Maximization (EM) step.

However, checking the EM formula, I feel the algorithm can work perfectly without the binarization. Thus, I would like to ask if the binarization step is compulsory, or it is to enhance the performance, or it is just to make the formula clean.


  • Will you be doing plain PCFG where the P depends on just the current nodes or will you also be taking some hierarchy into account? If it's the former, I think not converting it to CNF will be to your advantage, if it's the latter, I would like to know how it turns out for you. – prash Apr 13 '15 at 19:13

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.