PCFG vs SR Parser

问题

It looks like stanfordnlp has these SR models for some time. I am really new to NLP but we are currently using PCFG parser and we are having serious performance issues( that we cut down the parse length to 35)

1. I was thinking if we could try using SR. I tried it with POS tagger from stanford(english-left3words-distsim.tagger) Would you know how SR is on accuracy vs PCFG? I also find sentence root detection issues with SR and dep parse: Example:
   Michael Jeffrey Jordan, also known by his initials, MJ, is an American former professional basketball player, entrepreneur, and current majority owner and chairman of the Charlotte Bobcats The PCFG is really accurate with the root and detects player as the root.
2. Would also appreciate a little insight on the NN people use e.g.(https://mailman.stanford.edu/pipermail/java-nlp-user/2014-November/006513.html) in above post. Do I need to use another tagger like - left3words with this? I am sorry if this sounds a little naive. But all I want is a correct sentence root and its dependencies. Does POS tagging upfront make it fast?

Thanks a lot in advance.

回答1:

1. The English shift-reduce parser shipped with CoreNLP is actually slightly better than the PCFG parser on our test data. You can see performance metrics at the bottom of the shift-reduce parser homepage.
2. I've asked for clarification in a comment above.

来源：https://stackoverflow.com/questions/29241123/pcfg-vs-sr-parser