lividwo / perturbed-masking Goto Github PK

Hi,

Could you share the visualization code to reproduce Figure 1 in the paper?

Thank you!

Best,
Wenting

Thanks for the code publishing. Here comes a issue when I am trying to repro the ABSC part in paper.
For the ABSC part, the paper use PWCN model and compare the induced tree with the left/right chain additionally. And the left/right get different performance results. However, based on the data pre-processing in PWCN, the left and the right chain should be same. So I am wondering why they two have different output results? or there is some trick missing in the paper?
Expect the reply for this issue, really Thanks.

This error occurs for constituency probe:

parsing.py: error: argument --decoder: invalid choice: 'mart' (choose from 'eisner', 'cle', 'right_chain', 'top_down', 'right_branching', 'left_branching')

after running:

python parsing.py --probe constituency \
    --matrix [the generated impact matrix file] \
    --decoder mart

Thank you for your interesting and impressive work on interpreting BERT in unsupervised linguistic tasks. I have a small question about Figure 1 in the paper, just as below:

What is the meaning of each cell, does it mean the distance or the impact? I understand it should be impact since the following sections address that the higher impact represents the closer correlation. However, the right-side color bar ranging from 0 to 5 makes me very confused. So, I'm wondering how to compute the cell using Dist or Prob?

Looking forward to your response, thanks :)

I see the WSJ data is empty except WSJ_0001.MRG：

and in your code, the test id is WSJ_23：

but I cant see it in your data file

Hi, @LividWo , thanks for your great work.
I visualized some samples' impact matrix on PUD and found that usually, the "CLS" row has very high values. Any ideas about this phenomenon? Thanks again!

Reading the paper (section 2.2), I see the two difference functions

• "Dist" (Euclidean distance between the embeddings) and
• "Prob" (difference in probability of target token)

However, looking at the code, I see the options for the --metric flag are

• dist (which subtracts and takes norm to get Euclidean distance),

  Perturbed-Masking/dependency/get_matrix_for_dep_probe.py

  Lines 78 to 79 in f578592

  	if args.metric == 'dist':
  	init_matrix[i][j] = np.linalg.norm(base_state - state)

  and
• cos (which gets cosine similarity)

  Perturbed-Masking/dependency/get_matrix_for_dep_probe.py

  Lines 80 to 82 in f578592

  	if args.metric == 'cos':
  	init_matrix[i][j] = np.dot(base_state, state) / (
  	np.linalg.norm(base_state) * np.linalg.norm(state))

Is the metric cos the "Prob" of the paper?

Thanks for publishing your code.
I am trying to reproduce the results of the ABSC tasks. I used bert-base-uncased to generate the distance of Laptops, but got different distances from your Laptops_Test_Gold.xlm.seg.dist. Since the distance generation does not include random factors (BERT in evaluation mode should always give the same prediction), could you help to achieve the distance as you got in the ABSC folder ? Which layer was you using and metric, etc. ?

when calculating the impact of x_i+1 to x_i, is there a must to mask x_i? I have done a test on sentence segment without masking x_i , and it achieved a comparatively better result with less matrix calculation.

I tried NLTK to plot the tree, but it looks different.