Repository files navigation

CS224n is awesome, the original code of assignments is based on Python 2, I'm a fan of Python 3 though, for those who prefer 3 to 2, feel free to clone or fork this repo, have fun with it!

bug fix(early 2018 version)

Assignment 3:

• q1_window.py

  WindowModel(NERModel).create_feed_dict(), the defaultdropout should be 0(i.e.keep_prob=1 inTensorFlow by default, we would like to disable droupout when predicting). However if you treatedropout askeep_prob, then there's no tr�ouble.

• q2_rnn.py

  InRNNModel(NERModel).preprocess_sequence_data(), should passwindow_size=self.config.window_size() callingfeaturize_windows(), or your model will crash if you changewindow_size inConfig.

Progress

• Assignment 1:

  1. Softmax ✔️
  2. Neural Network Basics ✔️
  3. word2vec ✔️
  4. Sentiment Analysis(I can not fix encoding error in Python3, this was done by Python2) ✔️

• Assignment 2:

  1. Tensorflow Softmax✔️
  2. Neural Transition-Based Dependency Parsing ✔️

  model	training loss(debug)	dev UAS(debug)	training loss(full)	dev UAS(full)
  baseline	0.1203	69.97	0.0703	86.68
  + L2 reg	0.2402	66.38	0.1212	85.81

  adding L2 regularization hurts the model, for the model is simple(low capacity), regularization actually reduce the capacity of baseline model.

• Assignment 3:

  1. A window into NER ✔️

     • best score(Entity level P/R/F1): 0.85/0.87/0.86

     • Token-level confusion matrix

       gold\guess	PER	ORG	LOC	MISC	O
       PER	2968	45	53	18	65
       ORG	92	1738	67	88	107
       LOC	35	87	1931	16	25
       MISC	32	53	32	1056	95
       O	40	43	22	37	42617

     • Token-level scores:

       label	acc	prec	rec	f1
       PER	0.99	0.94	0.94	0.94
       ORG	0.99	0.88	0.83	0.86
       LOC	0.99	0.92	0.92	0.92
       MISC	0.99	0.87	0.83	0.85
       O	0.99	0.99	1	0.99
       micro	0.99	0.98	0.98	0.98
       macro	0.99	0.92	0.9	0.91
       not-O	0.99	0.91	0.89	0.9

  2. Recurrent neural nets for NER ✔️

  3. Grooving with GRUs (30 points) ✔️

     • best score(Entity level) P/R/F1: 0.87/0.86/0.86