LADA

This repo contains codes for the following paper:

Jiaao Chen*, Zhenghui Wang*, Ran Tian, Zichao Yang, Diyi Yang: Local Additivity Based Data Augmentation for Semi-supervised NER. In Proceedings of The 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP'2020)

If you would like to refer to it, please cite the paper mentioned above.

Getting Started

These instructions will get you running the codes of LADA.

Requirements

• Python 3.6 or higher
• Pytorch >= 1.4.0
• Pytorch_transformers (also known as transformers)
• Pandas, Numpy, Pickle, faiss, sentence-transformers

Code Structure

├── code/
│   ├── BERT/
│   │   ├── back_translate.ipynb --> Jupyter Notebook for back translating the dataset
│   │   ├── bert_models.py --> Codes for LADA-based BERT models
│   │   ├── eval_utils.py --> Codes for evaluations
│   │   ├── knn.ipynb --> Jupyter Notebook for building the knn index file
│   │   ├── read_data.py --> Codes for data pre-processing
│   │   ├── train.py --> Codes for trianing BERT model
│   │   └── ...
│   ├── flair/
│   │   ├── train.py --> Codes for trianing flair model
│   │   ├── knn.ipynb --> Jupyter Notebook for building the knn index file
│   │   ├── flair/ --> the flair library
│   │   │   └── ...
│   │   ├── resources/
│   │   │   ├── docs/ --> flair library docs
│   │   │   ├── taggers/ --> save evaluation results for flair model
│   │   │   └── tasks/
│   │   │       └── conll_03/
│   │   │           ├── sent_id_knn_749.pkl --> knn index file
│   │   │           └── ... -> CoNLL-2003 dataset
│   │   └── ...
├── data/
│   └── conll2003/
│       ├── de.pkl -->Back translated training dataset with German as middle language
│       ├── labels.txt --> label index file
│       ├── sent_id_knn_700.pkl
│       └── ...  -> CoNLL-2003 dataset
├── eval/
│   └── conll2003/ --> save evaluation results for BERT model
└── README.md

BERT models

Downloading the data

Please download the CoNLL-2003 dataset and save under ./data/conll2003/ as train.txt, dev.txt, and test.txt.

Pre-processing the data

We utilize Fairseq to perform back translation on the training dataset. Please refer to ./code/BERT/back_translate.ipynb for details.

Here, we have put one example of back translated data, de.pkl, in ./data/conll2003/ . You can directly use it for CoNLL-2003 or generate your own back translated data following ./code/BERT/back_translate.ipynb.

We also provide the kNN index file for the first 700 training sentences (5%) ./data/conll2003/sent_id_knn_700.pkl. You can directly use it for CoNLL-2003 or generate your own kNN index file following ./code/BERT/knn.ipynb

Training models

These section contains instructions for training models on CoNLL-2003 using 5% training data.

Training BERT+Intra-LADA model

python ./code/BERT/train.py --data-dir 'data/conll2003' --model-type 'bert' \
--model-name 'bert-base-multilingual-cased' --output-dir 'eval/conll2003' --gpu '0,1' \
--labels 'data/conll2003/labels.txt' --max-seq-length 164 --overwrite-output-dir \
--do-train --do-eval --do-predict --evaluate-during-training --batch-size 16 \
--num-train-epochs 20 --save-steps 750 --seed 1 --train-examples 700  --eval-batch-size 128 \
--pad-subtoken-with-real-label --eval-pad-subtoken-with-first-subtoken-only --label-sep-cls \
--mix-layers-set 8 9 10  --beta 1.5 --alpha 60  --mix-option --use-knn-train-data \
--num-knn-k 5 --knn-mix-ratio 0.5 --intra-mix-ratio 1 

Training BERT+Inter-LADA model

python ./code/BERT/train.py --data-dir 'data/conll2003' --model-type 'bert' \
--model-name 'bert-base-multilingual-cased' --output-dir 'eval/conll2003' --gpu '0,1' \
--labels 'data/conll2003/labels.txt' --max-seq-length 164 --overwrite-output-dir \
--do-train --do-eval --do-predict --evaluate-during-training --batch-size 16 \
--num-train-epochs 20 --save-steps 750 --seed 1 --train-examples 700  --eval-batch-size 128 \ 
--pad-subtoken-with-real-label --eval-pad-subtoken-with-first-subtoken-only --label-sep-cls \ 
--mix-layers-set 8 9 10  --beta 1.5 --alpha 60  --mix-option --use-knn-train-data \
--num-knn-k 5 --knn-mix-ratio 0.5 --intra-mix-ratio -1  

Training BERT+Semi-Intra-LADA model

python ./code/BERT/train.py --data-dir 'data/conll2003' --model-type 'bert' \
--model-name 'bert-base-multilingual-cased' --output-dir 'eval/conll2003' --gpu '0,1' \
--labels 'data/conll2003/labels.txt' --max-seq-length 164 --overwrite-output-dir \
--do-train --do-eval --do-predict --evaluate-during-training --batch-size 16 \
--num-train-epochs 20 --save-steps 750 --seed 1 --train-examples 700  --eval-batch-size 128 \
--pad-subtoken-with-real-label --eval-pad-subtoken-with-first-subtoken-only --label-sep-cls \
--mix-layers-set 8 9 10  --beta 1.5 --alpha 60  --mix-option --use-knn-train-data \
--num-knn-k 5 --knn-mix-ratio 0.5 --intra-mix-ratio 1 \
--u-batch-size 32 --semi --T 0.6 --sharp --weight 0.05 --semi-pkl-file 'de.pkl' \
--semi-num 10000 --semi-loss 'mse' --ignore-last-n-label 4  --warmup-semi --num-semi-iter 1 \
--semi-loss-method 'origin' 

Training BERT+Semi-Inter-LADA model

python ./code/BERT/train.py --data-dir 'data/conll2003' --model-type 'bert' \
--model-name 'bert-base-multilingual-cased' --output-dir 'eval/conll2003' --gpu '0,1' \
--labels 'data/conll2003/labels.txt' --max-seq-length 164 --overwrite-output-dir \
--do-train --do-eval --do-predict --evaluate-during-training --batch-size 16 \
--num-train-epochs 20 --save-steps 750 --seed 1 --train-examples 700  --eval-batch-size 128 \ 
--pad-subtoken-with-real-label --eval-pad-subtoken-with-first-subtoken-only --label-sep-cls \
--mix-layers-set 8 9 10  --beta 1.5 --alpha 60  --mix-option --use-knn-train-data \
--num-knn-k 5 --knn-mix-ratio 0.5 --intra-mix-ratio -1 \
--u-batch-size 32 --semi --T 0.6 --sharp --weight 0.05 --semi-pkl-file 'de.pkl' \
--semi-num 10000 --semi-loss 'mse' --ignore-last-n-label 4  --warmup-semi --num-semi-iter 1 \
--semi-loss-method 'origin' 

flair models

flair is a BiLSTM-CRF sequence labeling model, and we provide code for flair+Inter-LADA

Downloading the data

Please download the CoNLL-2003 dataset and save under ./code/flair/resources/tasks/conll_03/ as eng.train, eng.testa (dev), and eng.testb (test).

Pre-processing the data

We also provide the kNN index file for the first 749 training sentences (5%, including the -DOCSTART- seperator) ./code/flair/resources/tasks/conll_03/sent_id_knn_749.pkl. You can directly use it for CoNLL-2003 or generate your own kNN index file following ./code/flair/knn.ipynb

Training models

These section contains instructions for training models on CoNLL-2003 using 5% training data.

Training flair+Inter-LADA model

CUDA_VISIBLE_DEVICES=1 python ./code/flair/train.py --use-knn-train-data --num-knn-k 5 \
--knn-mix-ratio 0.6 --train-examples 749 --mix-layer 2  --mix-option --alpha 60 --beta 1.5 \
--exp-save-name 'mix'  --mini-batch-size 64  --patience 10 --use-crf