Skip to content
No description or website provided.
Python Perl Shell
Branch: master
Clone or download
Latest commit 4a1aad1 Aug 29, 2019
Type Name Latest commit message Commit time
Failed to load latest commit information.
commands Update Aug 28, 2019
dataloaders Updating readme and other minor Aug 28, 2019
eval adding conll eval Aug 29, 2019
helper_scripts adding the code Aug 28, 2019
models adding the entropy Aug 29, 2019
utils Updating readme and other minor Aug 28, 2019 adding conll eval Aug 29, 2019 adding the code Aug 28, 2019 Updating readme and other minor Aug 28, 2019

Active Learning for Entity Recognition


python 2.7
DynetVersion commit 284838815ece9297a7100cc43035e1ea1b133a5


In the data/, create a directory per language as shown for data/Spanish. Download the CoNLL train/dev/test NER datasets for that language here. To acquire LDC datasets, please get the required access.

For storing the trained models, create directory saved_models in the parent folder.


Combine monolingual data acquired from Wikipedia with the plain text extracted from the labeled data. Train 100-d Glove embeddings

Active Learning Simulation

The best NER performance was obtained using fine-tuning training scheme. The scripts below runs simulation active learning runs for different active learning strategies: cd commands

  • ETAL + Partial-CRF + CT (Proposed recipe)
  • ETAL + Full-CRF + CT
  • CFEAL + Full-CRF + CT
  • SAL + CT
    Things to note:

We load the vocabulary from the following path--aug_lang_train_path. Therefore, create a conll formatted file with dummy labels from the unlabeled text. For our experiments, we concatenated the transferred data with the unlabeled data (which was the entire training dataset) into a single conll formatted file. The conll format is a tab separated two-column format as shown below:

El O
grupo O

The LDC NER label set differ from the CoNLL label set by one tag. Therefore, add --misc to the argument set when running any experiments on CoNLL data. The label set has been hard-coded in the data_loaders/ file.

Cross-Lingual Transferred Data

We used the model proposed by (Xie et al. 2018) to get the cross-lingually transferred data from English. Please refer to their code here.

For the Fine-Tune training scheme, train a base NER model on the transferred model as follows:

python -u ../ \
    --dynet-seed 3278657 \
    --word_emb_dim 100 \
    --batch_size 10 \
    --model_name ${MODEL_NAME} \
    --lang es \
    --fixedVocab \
    --test_conll \
    --tot_epochs 1000 \
--aug_lang_train_path $DATA/vocab.conll \
    --init_lr 0.015 \
    --valid_freq 1300 \
    --misc \
    --pretrain_emb_path $DATA/esp.vec \
    --dev_path $DATA/ \
    --test_path $DATA/esp.test \
    --train_path $DIR/transferred_data.conll  2>&1 | tee ${MODEL_NAME}.log 


If you make use of this software for research purposes, we will appreciate citing the following:

    title = {A Little Annotation does a Lot of Good: A Study in Bootstrapping Low-resource Named Entity Recognizers},
    author = {Aditi Chaudhary and Jiateng Xie and Zaid Sheikh and Graham Neubig and Jaime Carbonell},
    booktitle = {Conference on Empirical Methods in Natural Language Processing (EMNLP)},
    address = {Hong Kong},
    month = {November},
    url = {},
    year = {2019}


For any issues, please feel free to reach out to

You can’t perform that action at this time.