alignment errors

[xiulinyang]

Hi Austin,

When I ran align_with_pretrained_model.py, I got some errors. One seems caused by postprocess_graph (https://github.com/ablodge/leamr/blob/e1c2f8e4e46f22519e8ae4a9329359565dd5cbb2/models/subgraph_model.py#L178) so I used try except to print out alignments when postprocess_graph fails (they are shown in the error message). The other one is get_alignment. I put the error message below. I have no clue why because all the AMRs can be read by the penman package. Thanks!

 python align_with_pretrained_model.py -t  out_ewt/post_orc.txt --subgraph-model ldc+little_prince.subgraph_params.pkl --relation-model ldc+little_prince.relation_params.pkl --reentrancy-model ldc+little_prince.reentrancy_params.pkl
[amr] Loading AMRs from file: out_ewt/post_orc.txt
Loading model: ldc+little_prince.subgraph_params.pkl
Apply Rules = True
Preprocessing: 199 / 340<AMR_Alignment: subgraph>: tokens [7] nodes ['1.1.1'] edges [] (subgraph : see => see-01)
Preprocessing: 276 / 340<AMR_Alignment: subgraph>: tokens [9] nodes ['1.1.2'] edges [] (subgraph : experience => experience-01)
<AMR_Alignment: subgraph>: tokens [12] nodes ['1.1.2.2'] edges [] (subgraph : forget => forget-01)
Preprocessing coverage: 74.94%
 80%|████████████████████████████████▊        | 272/340 [00:08<00:01, 43.43it/s]0
0
0
0
100%|█████████████████████████████████████████| 340/340 [00:11<00:00, 28.35it/s]
Writing subgraph alignments to: out_ewt/post_orc.subgraph_alignments.json
Loading model: ldc+little_prince.relation_params.pkl
Preprocessing coverage: 86.34%
100%|████████████████████████████████████████| 340/340 [00:00<00:00, 777.55it/s]
Writing relation alignments to: out_ewt/post_orc.relation_alignments.json
199
276
Loading model: ldc+little_prince.reentrancy_params.pkl
276 / 340 preprocessedTraceback (most recent call last):
  File "align_with_pretrained_model.py", line 64, in <module>
    main()
  File "align_with_pretrained_model.py", line 57, in main
    reent_alignments = reent_model.align_all(eval_amrs)
  File "/local/xiulyang/leamr/models/base_model.py", line 92, in align_all
    alignments = self.get_initial_alignments(amrs, preprocess)
  File "/local/xiulyang/leamr/models/reentrancy_model.py", line 201, in get_initial_alignments
    self.align_primary_edges(amr, reentrancy_alignments)
  File "/local/xiulyang/leamr/models/reentrancy_model.py", line 174, in align_primary_edges
    rel_align = amr.get_alignment(self.relation_alignments, token_id=talign.tokens[0])
IndexError: list index out of range

[nschneid]

@xiulinyang Could you please provide the sentence/AMR for which it fails? (Is it from one of the AMR corpora or a parser output?)

[xiulinyang]

This is one example that caused errors from a parser output. Thanks!!

  File "align_with_pretrained_model.py", line 64, in <module>
    main()
  File "align_with_pretrained_model.py", line 46, in main
    rel_alignments = rel_model.align_all(eval_amrs)
  File "/local/xiulyang/leamr/models/relation_model.py", line 309, in align_all
    alignments = super().align_all(amrs, alignments, preprocess, debug)
  File "/local/xiulyang/leamr/models/base_model.py", line 102, in align_all
    aligns, scores = self.align(amr, alignments, n, unaligned, return_all=True)
  File "/local/xiulyang/leamr/models/relation_model.py", line 248, in align
    candidate_spans = [span for span in candidate_spans if (parent.tokens[0]<span[0]<child.tokens[0])
  File "/local/xiulyang/leamr/models/relation_model.py", line 249, in <listcomp>
    or (child.tokens[0]<span[0]<parent.tokens[0])
IndexError: list index out of range

# ::snt If you know someone who might fit the bill ask them to contact me and/or send me their resume .
(p0 / ask-02
    :mode imperative
    :ARG0 (p1 / you)
    :ARG1 (p2 / and-or
              :op1 (p3 / contact-01
                       :ARG0 p12
                       :ARG1 (p4 / i))
              :op2 (p5 / send-01
                       :ARG0 p12
                       :ARG1 (p6 / resume
                                 :poss p12)
                       :ARG2 p4))
    :ARG2 p12
    :condition (p7 / know-02
                   :ARG0 p1
                   :ARG1 (p8 / someone
                             :ARG1-of (p9 / fit-06
                                          :ARG2 (p10 / bill)
                                          :ARG1-of (p11 / possible-01)))))

[nschneid]

(BTW the sentence is from EWT, not the AMR dataset.)

[ablodge]

Sorry for the delay. It looks like some of the nodes aren't getting aligned. Can you please add the following lines before the line that produces the error (leamr/models/relation_model.py, line 249)? Then tell me what the output says.

if ' '.join(amr.tokens) == 'If you know someone who might fit the bill ask them to contact me and/or send me their resume .':
    print(parent.readable(amr))
    print(child.readable(amr))

[xiulinyang]

No problem! Here is the output:

subgraph : ask => ask-02
subgraph : you => you, imperative
parent+child
subgraph : ask => ask-02
subgraph : know => know-02
parent+child
subgraph : resume => resume
 => 
  0%|                                                                                                                                  | 0/1 [00:00<?, ?it/s]
Traceback (most recent call last):
  File "align_with_pretrained_model.py", line 64, in <module>
    main()
  File "align_with_pretrained_model.py", line 46, in main
    rel_alignments = rel_model.align_all(eval_amrs)
  File "/local/xiulyang/leamr/models/relation_model.py", line 312, in align_all
    alignments = super().align_all(amrs, alignments, preprocess, debug)
  File "/local/xiulyang/leamr/models/base_model.py", line 102, in align_all
    aligns, scores = self.align(amr, alignments, n, unaligned, return_all=True)
  File "/local/xiulyang/leamr/models/relation_model.py", line 251, in align
    candidate_spans = [span for span in candidate_spans if (parent.tokens[0]<span[0]<child.tokens[0])
  File "/local/xiulyang/leamr/models/relation_model.py", line 252, in <listcomp>
    or (child.tokens[0]<span[0]<parent.tokens[0])
IndexError: list index out of range

Thanks a lot!