Presentation is loading. Please wait.

Presentation is loading. Please wait.

University of Edinburgh27/10/20151 Lexical Dependency Parsing Chris Brew OhioState University.

Similar presentations


Presentation on theme: "University of Edinburgh27/10/20151 Lexical Dependency Parsing Chris Brew OhioState University."— Presentation transcript:

1 University of Edinburgh27/10/20151 Lexical Dependency Parsing Chris Brew OhioState University

2 University of Edinburgh27/10/20152 Using Lexical Dependencies  Lexical information crucial to parser success –Original version is Magerman’s SPATTER –Each is simpler than the last –Often also with improved performance

3 University of Edinburgh27/10/20153 The Task  Generate trees as in Wall Street Journal part of Penn Treebank  Collins provides new statistical model for P(T|S)  PCFGs used rules, DOP used fragments, LR used parser states  This uses Bigram Lexical Dependencies plus a few extras

4 University of Edinburgh27/10/20154 The components of the model  A model of base NPs P(B|S) –Obtained using bigram statistics and POS tags  A model of dependencies P(D|S,B)  A bijective mapping which can interconvert between –Trees –Pairings of base NP structure and dependencies

5 University of Edinburgh27/10/20155 A parse tree –Base NPs [John Smith][the president][IBM] [his resignation] [yesterday] – Treebank is linguistically odd here

6 University of Edinburgh27/10/20156 Propagating head words –Small set of rules propagate heads S(announced) NP(Smith) NNP John NNP Smith NP(president) NP DT the NN president PP(of) IN of NP NNP IBM VP(announced) VBD announced NP(resignation) PRP$ his NN resignation NP NN yesterday

7 University of Edinburgh27/10/20157 Extracted structure  Base NPs plus Dependencies  Dependencies labelled with triplesof nonterminals NB. Not all dependencies shown here NP [JohnSmith] NP [thepresident]of[IBM] SNPVP announced[hisResignation][yesterday] VPVBDNP VP VBD

8 University of Edinburgh27/10/20158 Statistical model  Gives probabilities to dependencies  So the probability of a rule like VP -> VBD NP NP, which involves two dependencies, is made from the probabilities of the components. announced[hisResignation][yesterday] VPVBDNP VPVBD


Download ppt "University of Edinburgh27/10/20151 Lexical Dependency Parsing Chris Brew OhioState University."

Similar presentations


Ads by Google