sighsmile / conlleval Goto Github PK

请问一下您，true_tags和pred_tags的输入是否类似[[,,,,,,,,,],[,,,,,,,],[,,,,,,,,,,]]，即二维嵌套list类型？

#这样会报错
from conlleval import evaluate
true_tags =[ ['O', 'B-Part', 'I-Part'],[ 'O', 'O', 'O']]
pred_tags = [ ['O', 'B-Part', 'I-Part'],[ 'O', 'O', 'O']]
evaluate(true_tags,pred_tags)

#这样不报错
from conlleval import evaluate
true_tags =[ 'O', 'B-Part', 'I-Part', 'O', 'O', 'O']
pred_tags = [ 'O', 'B-Part', 'I-Part', 'O', 'O', 'O']
evaluate(true_tags,pred_tags)

这是否意味着每次只能对一个序列进行性能评估？

I have ran the script on the following data:

(978) B-Phone I-Phone
934-3623 I-Phone I-Phone

In the IOB2 mode, every entity tag should starts with B. So the precision of the above should be 0. But the script you provided shows a result of 100% precision. However, I tried the original perl version conlleval script, same result.

Getting this error when using arbitrary tags like , , etc.

count_chunks(true_seqs, pred_seqs)
129
130 _, true_type = split_tag(true_tag)
--> 131 _, pred_type = split_tag(pred_tag)
132
133 if correct_chunk is not None:

ValueError: not enough values to unpack (expected 2, got 1)

I believe this happens if tags don't have a "-" in them or aren't an "O".

My output.txt file is the following format:
Sao NC B-LOC B-LOC
Paulo VMI I-LOC I-LOC
( Fpa O O
Brasil NC B-LOC B-LOC
) Fpt O O
, Fc O O
23 Z O O
may NC O O
( Fpa O O
EFECOM NP B-ORG B-ORG
) Fpt O O
. Fp O O

Which should be okay according to the conll format. But after running the conll.py I am getting this error:

Do you have any idea why? @sighsmile
Thanks in advance!

my output file is like:

a  B-LOC  B-LOC
b  I-LOC  E-LOC
c  E-LOC  S-LOC

and the result is different,
for conlleval.py

processed 3 tokens with 1 phrases; found: 2 phrases; correct: 0.
accuracy:  33.33%; (non-O)
accuracy:  33.33%; precision:   0.00%; recall:   0.00%; FB1:   0.00
              LOC: precision:   0.00%; recall:   0.00%; FB1:   0.00  2

for connlleval_perl.py

processed 3 tokens with 1 phrases; found: 1 phrases; correct: 1.
accuracy:  33.33%; precision: 100.00%; recall: 100.00%; FB1: 100.00
              LOC: precision: 100.00%; recall: 100.00%; FB1: 100.00  1

which one is reliable?

IOB2 is supported as input as you told in the readme, but the evaluation metrics is not in iob2 way. It's in iob1 way. So, you can either provide option for people to choose evaluation scheme for iob1 or iob2, or you can stop the support of iob2 format input.