Implementation of Nested Named Entity Recognition

Some files are part of NeuroNLP2.

Requirements

We tested this library with the following libraries:

• Python (3.7)
• PyTorch (1.10.0)
• Numpy (1.17.3)
• StanfordNLP (0.2.0) for accessing the Java Stanford CoreNLP Server (3.9.2)
• Flair (0.9)

Running experiments

Testing this library with a sample data

1. Put the embedding file PubMed-shuffle-win-2.bin into the "./embeddings/" directory
2. Run the gen_data.py to generate the processed data files for training, and they will be placed at the "./data/" directory

   python gen_data.py
   

3. Run the train.py to start training

   python train.py
   

Reproducing our experiment on the ACE-2004 dataset

1. Put the corpus ACE-2004 into the "../ACE2004/" directory
2. Put this .tgz file into the "../" and extract it
3. Run the parse_ace2004.py to extract sentences for training, and they will be placed at the "./data/ace2004/"

   python parse_ace2004.py
   

4. Put the embedding file GoogleNews-vectors-negative300.bin.gz into the "./embeddings/" directory
5. Decompress the embedding file GoogleNews-vectors-negative300.bin.gz

   gzip -d embeddings/GoogleNews-vectors-negative300.bin.gz
   

6. Run the gen_data_for_ace2004.py to prepare the processed data files for training, and they will be placed at the "./data/" directory

   python gen_data_for_ace2004.py
   

7. Run the train.py to start training

   python train.py
   

Reproducing our experiment on the ACE-2005 dataset

1. Put the corpus ACE-2005 into the "../ACE2005/" directory
2. Put this .tgz file into the "../" and extract it
3. Run the parse_ace2005.py to extract sentences for training, and they will be placed at the "./data/ace2005/"

   python parse_ace2005.py
   

4. Put the embedding file GoogleNews-vectors-negative300.bin.gz into the "./embeddings/" directory
5. Decompress the embedding file GoogleNews-vectors-negative300.bin.gz

   gzip -d embeddings/GoogleNews-vectors-negative300.bin.gz
   

6. Run the gen_data_for_ace2005.py to prepare the processed data files for training, and they will be placed at the "./data/" directory

   python gen_data_for_ace2005.py
   

7. Run the train.py to start training

   python train.py
   

Reproducing our experiment on the GENIA dataset

1. Put the corpus GENIA into the "../GENIA/" directory
2. Run the parse_genia.py to extract sentences for training, and they will be placed at the "./data/genia/"

   python parse_genia.py
   

3. Put the embedding file PubMed-shuffle-win-2.bin into the "./embeddings/" directory
4. Run the gen_data_for_genia.py to prepare the processed data files for training, and they will be placed at the "./data/" directory

   python gen_data_for_genia.py
   

5. Run the train.py to start training

   python train.py
   

Configuration

Configurations of the model and training are in config.py

Citation

Please cite our paper:

@article{shibuya-hovy-2020-nested,
  title = "Nested Named Entity Recognition via Second-best Sequence Learning and Decoding",
  author = "Shibuya, Takashi and Hovy, Eduard",
  journal = "Transactions of the Association for Computational Linguistics",
  volume = "8",
  year = "2020",
  doi = "10.1162/tacl_a_00334",
  pages = "605--620",
}