1

I am developing a parser which is not a dependency parser, but it still is a transition-based system. Instead of using 3 actions SHIFT, REDUCE-LEFT and REDUCE-RIGHT like arc-transition, I proposed my own actions for my parser. As my understanding through the paper of Andor et al. (2016), I can train SyntaxNet on any transition-based system as long as the system follows some rules defined in the paper. But I do not know exactly where I should start. I looked at the code but it seems that they are mainly for training a dependency parser (with 3 actions SHIFT, REDUCE-LEFT and REDUCE-RIGHT). I really appreciate if you can highlight some parts in the code which can help me achieve my goal.

Also, I do not quite understand how SyntaxNet can transform the ConNLL data format into gold training data, can you show me which source file is responsible for this task?

Thank you very much in advance

hippietrail
  • 15,848
  • 18
  • 99
  • 158
The Lazy Log
  • 3,564
  • 2
  • 20
  • 27

1 Answers1

0

If you want to change the transition system with your own one, you will need to make several changes to the code. Most notably, you will need to implement an oracle for mapping the gold parse trees to a sequence of transitions. Nivre (2008) is a good introduction to the topic: http://www.mitpressjournals.org/doi/pdf/10.1162/coli.07-056-R1-07-027

Slav Petrov
  • 1
  • Yes, but my question is related to SyntaxNet as I want to put my own ideas into this framework instead of starting from scratch. – The Lazy Log Aug 14 '16 at 23:23