AESINER

This is the implementation of Improving Named Entity Recognition with Attentive Ensemble of Syntactic Information at Findings of EMNLP-2020.

Citations

If you use or extend our work, please cite our paper at EMNLP2020.

@inproceedings{nie-emnlp-2020-aesiner,
    title = "Improving Named Entity Recognition with Attentive Ensemble of Syntactic Information",
    author = "Nie, Yuyang and
      Tian, Yuanhe  and
      Song, Yan  and
      Ao, Xiang and
      Wan, Xiang",
    booktitle = "Findings of the 2020 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2020",
}

Requirements

• torch==1.1.0
• spacy==2.2.4
• tqdm==4.38.0
• fastNLP==0.5

Download Stanford CoreNLP

You can get the Stanford CoreNLP toolkits from the script stanford.sh

Download Pre-trained Embeddings

For English NER, we use three types of word embeddings, namely GloVe, ELMo and BERT. Among them, GloVe and ELMo can be automatically downloaded by running the script run_en.py; bert can be downloaded pre-trained BERT-large-cased from Google or from HuggingFace. If you download it from Google, you need to convert the model from TensorFlow version to PyTorch version.

For Chinese NER, we also use three types of word embeddings, namely Giga, Tencent Embedding and ZEN. Among them, Giga can be downloaded from here, Tencent Embedding can be downloaded from here, ZEN can be downloaded from here

All pretrained embeddings should be placed in ./data/

Download AESINER

You can download the models we trained for each dataset from here.

Run on sample data

Run run_en.py to train a model on the small sample data under the sample_data directory.

Datasets

We use three English datasets (ON5e, WN16, WN17) and three Chinese datasets (ON4c, RE, WE) in our paper.

For ON5e and ON4c, you need to obtain the official data first, and then put the data in data/ON5e and data/ON4c, respectively.

For WN16, you can download the dataset from here and then put the train, dev and test files in data/W16.

For WN17, you can download the dataset from here and then put the emerging.train.conll, emerging.dev.conll and emerging.test.conll files in data/W17.

For RE, you can download the dataest from here and then put the train.char.bmes, dev.char.bmes and test.char.bmes files in data/RE.

For WE, you can download the dataset from here and then put the weiboNER_2nd_conll.train, weiboNER_2nd_conll.dev and weiboNER_2nd_conll.test files in data/WE.

All the data files should be named as train.txt, test.txt and dev.txt in corresponding dictionaries.

For all datasets, you can run the script data_preprocess.py with python data_preprocess.py --dataset ${dataset}$ --data_dir ${data_dir}$ to preprocess the aforementioned datasets automatically.

Data Preprocess

After downloading Tencent Embedding, you need to extract the unigrams according to python data_process.py --file_path=${PATH_TO_TENCENT_EMBEDDING}$

For all syntactic information, you can run the script data_helper.py to get the syntactic information needed to run our model, you can run the script by python data_helper.py --dataset $DATASET_NAME$

Training

You can find the command lines to train models on a specific dataset in run_en.py for English datasets and run_cn.py for Chinese datasets.