Code for our ACL2019 paper Reliability-aware Dynamic Feature Composition for Name Tagging.

• <input_dir>
  • embed.vocab.tsv (embedding vocab file, 1st column: token, 2nd column: index)
  • embed.count.tsv (embedding token frequency file, 1st column: token, 2nd column: frequency)
  • bc
    • train.tsv (training set)
    • dev.tsv (development set)
    • test.tsv (test set)
    • token.vocab.tsv (token vocab file, 1st column: token, 2nd column: index)
    • char.vocab.tsv (character vocab file: 1st column: character, 2nd column: index)
    • label.vocab.tsv (label vocab file: 1st column: label, 2nd column: index)
  • bn
  • mz
  • nw
  • tc
  • wb

Note:

• Other subsets have train.tsv, dev.tsv, test.tsv, token.vocab.tsv, char.vocab.tsv, and label.vocab.tsv in their directories.
• In our experiments, we generated *.vocab.tsv from a merged data set of all subsets.
• In our experiments, we use CoNLL format files generated from OntoNotes 5.0 with Pradhan et al.'s scripts, which can be found at https://cemantix.org/data/ontonotes.html.

The following functions in proprocess.py can be used to create vocab and frequency files.

• build_all_vocabs takes as input a list of CoNLL format files, and generate {token,char,label}.vocab.tsv in output_dir.
• build_embed_vocab takes a pre-trained embedding file as input and return the embedding vocab.
• build_embed_token_count takes a pre-trained embedding file as input and generate an embedding token frequency file.

python train_lstmcnn_all.py -d 0 -i <input_dir> -o <output_dir> -e <embedding_file>
  --embed_vocab <embedding_vocab_file> --char_dim 50 --seed <random_seed>

This script train a model for each subset (which can be specified with the --datasets argument) and report within-subset (within-genre) and cross-subset (cross-genre) performance.

python train_lstmcnn_dfc_all.py -d 0 -i <input_dir> -o <output_dir> -e <embedding_file>
  --embed_vocab <embedding_vocab_file> --embed_count <embedding_freq_file> --char_dim 50 --seed <random_seed>

• Python 3.5+
• Pytorch 1.0

• We use the 100d case-sensitive word embedding in Pre-trained Word Embeddings

Lin, Y., Liu, L., Ji, H., Yu, D., Han, J. (2019) Reliability-aware Dynamic Feature Composition for Name Tagging. Proceedings of The 57th Annual Meeting of the Association for Computational Linguistics.

@article{lin2019reliability,
  title={Reliability-aware Dynamic Feature Composition for Name Tagging},
  author={Lin, Ying and Liu, Liyuan and Ji, Heng and Yu, Dong and Han, Jiawei},
  booktitle={Proceedings of The 57th Annual Meeting of the Association for Computational Linguistics (ACL2019)},
  year={2019}
}