Sequence Labeling
Sequence labeling is a task that assigns categorial label to each element in an input sequence. Many problems can be formalized as sequence labeling task, including speech recognition, video analysis and various problems in NLP (e.g. POS tagging, NER, Chunking, etc.). Traditionally sequence labeling requires large amount of hand-engineered features and domain-specific knowledge, but recently neural approaches have achieved state-of-the-art performance on several sequence labeling benchmarks. A common data format for sequence labeling task is IOB (Inside-Outside-Beginning), although other alternative formats (e.g. IO, IOBES, BMEWO, BMEWO+, BILOU, etc.) might be used.
Figure 1: An NER example in IOB format
Setting
- Python 3.6.6
- Tensorflow 1.12
- NumPy 1.15.4
DataSet
- CoNLL2003 is a multi-task dataset, which contains 3 sub-tasks, POS tagging, syntactic chunking and NER. For NER sub-task, it contains 4 types of named entities: persons, locations, organizations and names of miscellaneous entities that do not belong to the previous three groups.
- OntoNotes5 is a multi-task dataset, which contains several sub-tasks, including POS tagging, word sense disambiguation, coreference, NER and others. For NER sub-task, it contains 18 types of named entities: PERSON, LOC, ORG, DATE, MONEY and others. This dataset can be converted into CoNLL format using common tool.
- Treebank3 is a distributed release of Penn Treebank (PTB) project, which selected 2,499 stories from a three year Wall Street Journal (WSJ) collection of 98,732 stories for syntactic annotation, including POS tagging and constituency parsing.
- GloVe is an unsupervised learning algorithm for obtaining vector representations for words. Training is performed on aggregated global word-word co-occurrence statistics from a corpus, and the resulting representations showcase interesting linear substructures of the word vector space.
Usage
- Preprocess data
# preprocess train data
python conll/preprocess.py --format json --input_file data/conll2003/eng.train --output_file data/ner/train-conll2003/train-conll2003.ner.json
# preprocess dev data
python conll/preprocess.py --format json --input_file data/conll2003/eng.testa --output_file data/ner/dev-conll2003/dev-conll2003.ner.json
# preprocess test data
python conll/preprocess.py --format json --input_file data/conll2003/eng.testb --output_file data/ner/test-conll2003/test-conll2003.ner.json- Run experiment
# run experiment in train + eval mode
python sequence_labeling_run.py --mode train_eval --config config/config_sequence_template.xxx.json
# run experiment in train only mode
python sequence_labeling_run.py --mode train --config config/config_sequence_template.xxx.json
# run experiment in eval only mode
python sequence_labeling_run.py --mode eval --config config/config_sequence_template.xxx.json- Search hyper-parameter
# random search hyper-parameters
python hparam_search.py --base-config config/config_sequence_template.xxx.json --search-config config/config_search_template.xxx.json --num-group 10 --random-seed 100 --output-dir config/search- Visualize summary
# visualize summary via tensorboard
tensorboard --logdir=output- Export model
# export frozen model
python sequence_labeling_run.py --mode export --config config/config_sequence_template.xxx.json- Setup service
# setup tensorflow serving
docker run -p 8500:8500 -v output/xxx/model:models/ner -e MODEL_NAME=ner -t tensorflow/servingExperiment
Bi-LSTM + Char-CNN + Softmax
Figure 1: Bi-LSTM + Char-CNN + Softmax architecture
| CoNLL2003 - NER | F1 Score | Precision | Recall |
|---|---|---|---|
| Dev | 94.92 | 94.97 | 94.87 |
| Test | 91.29 | 90.41 | 92.18 |
Table 1: The performance of Bi-LSTM + Char-CNN + Softmax on CoNLL2003 NER sub-task with setting: num layers = 2, unit dim = 200, window size = [3]
| OntoNotes5 - NER | F1 Score | Precision | Recall |
|---|---|---|---|
| Dev | 86.22 | 84.21 | 88.32 |
| Test | 85.09 | 82.66 | 87.67 |
Table 2: The performance of Bi-LSTM + Char-CNN + Softmax on OntoNotes5 NER sub-task with setting: num layers = 2, unit dim = 200, window size = [3,5]
| Treebank3 - POS | Accuracy |
|---|---|
| Dev | 97.36 |
| Test | 97.58 |
Table 3: The performance of Bi-LSTM + Char-CNN + Softmax on Treebank3 POS tagging sub-task with setting: num layers = 2, unit dim = 200, window size = [3]
Bi-LSTM + Char-CNN + CRF
Figure 2: Bi-LSTM + Char-CNN + CRF architecture
| CoNLL2003 - NER | F1 Score | Precision | Recall |
|---|---|---|---|
| Dev | 94.93 | 94.92 | 94.93 |
| Test | 91.30 | 90.47 | 92.15 |
Table 4: The performance of Bi-LSTM + Char-CNN + CRF on CoNLL2003 NER sub-task with setting: num layers = 2, unit dim = 200, window size = [3]
| OntoNotes5 - NER | F1 Score | Precision | Recall |
|---|---|---|---|
| Dev | 86.45 | 84.11 | 88.93 |
| Test | 85.25 | 82.57 | 88.11 |
Table 5: The performance of Bi-LSTM + Char-CNN + CRF on OntoNotes5 NER sub-task with setting: num layers = 2, unit dim = 200, window size = [3,5]
| Treebank3 - POS | Accuracy |
|---|---|
| Dev | 97.27 |
| Test | 97.51 |
Table 6: The performance of Bi-LSTM + Char-CNN + CRF on Treebank3 POS tagging sub-task with setting: num layers = 2, unit dim = 200, window size = [3]
Reference
- Zhiheng Huang, Wei Xu, and Kai Yu. Bidirectional LSTM-CRF models for sequence tagging [2015]
- Jason PC Chiu and Eric Nichols. Named entity recognition with bidirectional lstm-cnns [2015]
- Xuezhe Ma and Eduard Hovy. End-to-end Sequence Labeling via Bi-directional LSTM-CNNs CRF [2016]
- Guillaume Lample, Miguel Ballesteros, Sandeep Subramanian, KazuyaKawakami, and ChrisDyer. Neural architectures for named entity recognition [2016]