kyubyong / bert_ner

Ner with Bert

PyTorch Implementation of NER with pretrained Bert

I know that you know BERT. In the great paper, the authors claim that the pretrained models do great in NER. It's even impressive, allowing for the fact that they don't use any prediction-conditioned algorithms like CRFs. We try to reproduce the result in a simple manner.

Requirements

• python>=3.6 (Let's move on to python 3 if you still use python 2)
• pytorch==1.0
• pytorch_pretrained_bert==0.6.1
• numpy>=1.15.4

Training & Evaluating

• STEP 1. Run the command below to download conll 2003 NER dataset.

bash download.sh

It should be extracted to conll2003/ folder automatically.

• STEP 2a. Run the command if you want to do the feature-based approach.

python train.py --logdir checkpoints/feature --batch_size 128 --top_rnns --lr 1e-4 --n_epochs 30

• STEP 2b. Run the command if you want to do the fine-tuning approach.

python train.py --logdir checkpoints/finetuning --finetuning --batch_size 32 --lr 5e-5 --n_epochs 3

Results in the paper

• Feature-based approach

• Fine-tuning

Results

• F1 scores on conll2003 valid dataset are reported.
• You can check the classification outputs in checkpoints.

epoch	feature-based	fine-tuning
1	0.2	0.95
2	0.75	0.95
3	0.84	0.96
4	0.88
5	0.89
6	0.90
7	0.90
8	0.91
9	0.91
10	0.92
11	0.92
12	0.93
13	0.93
14	0.93
15	0.93
16	0.92
17	0.93
18	0.93
19	0.93
20	0.93
21	0.94
22	0.94
23	0.93
24	0.93
25	0.93
26	0.93
27	0.93
28	0.93
29	0.94
30	0.93

bert_ner's Issues

F1, recall and precision calculation

Hi,
I was wondering how you are actually calculating your scores.

y_true = np.array([hp.tag2idx[line.split()[1]] for line in open(f, 'r').read().splitlines() if len(line) > 0])
y_pred = np.array([hp.tag2idx[line.split()[2]] for line in open(f, 'r').read().splitlines() if len(line) > 0])

num_proposed = len(y_pred[y_pred>1])
num_correct = (np.logical_and(y_true==y_pred, y_true>1)).astype(np.int).sum()
num_gold = len(y_true[y_true>1])

precision = num_correct / num_proposed
recall = num_correct / num_gold

Can you explain what the above code means?
How does this translate to say recall = TP / TP + FN? Don't you have to use some multi-class method?

Also, why are you only taking the index where y_true>1? Is it because you do not want the Other tag to skew your results? Thanks!

The ##word should not be predicted

In bert paper, it seems that the words start with '##' should not be predicted. And you did compute is_head variable, but why this variable is not used when computing loss ?

datasets

Can I use other datasets on this model?

Data download link invalid

Data download link invalid.
thank you！！！

I don't see you do truncate when the sequence longer than 512 lengths?

We know that when loading pretrained bert, max_seqence=512, but when you are dealing with the sentences from data set, each sentence is padding based on the longest sentence of the batch, but I don't see you do truncate when the sequence longer than 512 lengths?

Thanks

Cuda out of memory error for finetuning

I'm getting the following error when i try to run the finetuning example:

RuntimeError: CUDA out of memory. Tried to allocate 85.00 MiB (GPU 0; 4.00 GiB total capacity; 3.04 GiB already allocated; 9.21 MiB free; 15.31 MiB cached)

Reducing the batch_size didn't help.

How to use BERT for ENTITY extraction from a Sequence without classification in the NER task?

My requirement here is given a sentence(sequence), I would like to just extract the entities present in the sequence without classifying them to a type in the NER task. I see that BertForTokenClassification for NER does the classification. Can this be adapted for just the extraction?

Can you give me an idea of how to do entity extraction/identification using BERT?

can not find the file pytorch_pretrained_bert.py

hi,
Thank you for sharing the code. But I can not find the file 'pytorch_pretrained_bert.py', hence I could not run the code.
Please tell how can I find the file 'pytorch_pretrained_bert.py'. Thanks

Data download link invalid

The data download link is invalid. can you upload another connection? tank you very much