a-little-story / ner-1

BiLSTM-ELMo-CNN-CRF for CoNLL 2003

ELMo for NER

This project is a Tensorflow implementation of "mainstream" neural tagging scheme based on works of Deep contextualized word representations, Peters, et. al., 2018.

Requirements

• python 3.6
• tensorflow 1.10.0
• numpy 1.14.3
• gensim 3.6.0
• tqdm 4.26.0

Evaluation

Model	Dataset	Test F1
Peters et. al	CoNLL 2003	92.22(+/-0.10)
Ours	CoNLL 2003	92.23

Prepare Data

1. Download pre-trained word vector from http://nlp.stanford.edu/data/glove.6B.zip, unzip glove.6B.50d.txt to resources/pretrained/glove.
2. Download pre-trained elmo models elmo_2x1024_128_2048cnn_1xhighway_weights.hdf5 and elmo_2x1024_128_2048cnn_1xhighway_options.json, put them in resources/elmo.

Change glove to word2vec format

Open glove.6B.50d.txt with your favorite text editor, add 400000 50 to the first line like this:

400000 50
the 0.418 0.24968 ...
of 0.70853 0.57088 ...
...

Train

python elmo_train.py

Training Log

Epoch	Loss	Dev F1	Test F1
1	32237	90.81	87.77
2	12320	93.21	90.16
3	8823	94.19	91.75
4	6900	94.80	91.74
5	5821	94.30	91.03
6	4996	94.92	91.26
7	4467	95.18	92.06
8	3869	95.06	91.54
9	3483	95.13	91.88
10	3500	95.42	91.66
11	2989	95.01	91.82
12	2770	95.39	91.70
13	2649	95.39	91.68
14	2529	95.44	92.23
15	2407	95.02	91.77
16	2140	95.13	91.80
17	2149	95.09	92.06
18	1935	95.22	91.58
19	1946	94.88	91.91
20	1767	95.35	92.13