Download presentation
Presentation is loading. Please wait.
1
Dependency parsing spaCy and Stanford nndep
2
Universal Dependency (UD)
nsubj: nominal subject (more agentive, the do-er) obj: object csubj: clause subject aux: auxiliary verb
3
Dependency parsers spaCy Stanford corenlp nndep
open-source library for industrial-strength NLP (Cpython) Transition-based parser, averaged perceptron MIT license accurate and fast (~2 sec for a Wikipedia article) An Improved Non-monotonic Transition System for Dependency Parsing, CoNLL 2013 Stanford corenlp nndep Open-source java library Transition-based parser, 1 hidden layer neural network full GPL license accurate but slow (~15 sec for a Wikipedia article) A Fast and Accurate Dependency Parser using Neural Networks, EMNLP 2014
4
Transition-based parser
Arc-eager (spaCy) Actions: Shift, left-arc, right-arc, reduce Reduce: remove the top token from Stack Shift: takes the first incoming word from the buffer and push it onto the stack. Left-arc: creates a dependency between the top of the stack and the first word of the buffer, and reduce Right-arc: creates a dependency between the top of the stack and the first word of the buffer, and shift Arc-standard (Stanford corenlp nndep) Actions: Shift, left-arc, right-arc Output (They, subj, told) (told, iobj, him) (a, det, story) (told, dobj, story)
5
Features used to determine actions (Stanford nndep)
s1.w: word of the top token in stack b1.t: part-of-speech of the first token in buffer o: concatenate (s1.t = VB, s2.t= NN, s1.t o s2.t =VBNN) 1 hidden layer feedforward neural network is used
6
Features used to determine actions (spaCy)
averaged perceptron is used
8
The End Thank you very much. http://universaldependencies.org/u/dep/
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.