Environment

• Python 2.7.11
• torch
• CoNLL-2003 corpora

Preparation

1. Put CoNLL-2003 corpora into data/conll2003 folder. Your CoNLL 2003 files should be renamed and placed as below:

• data/conll2003/eng.testa.dev;
• data/conll2003/eng.testb.test;
• data/conll2003/eng.train.

2. Run script bash utils/prepare-senna-data.sh. It downloads senna embeddings, gazeteers and puts them into data/.
3. Install torch and python libs from requirements.txt (pip install -r requirements.txt).

Experiments

All experiments are done by using AWS g2.2xlarge with GPU.

Run bash experiments/convolution-net.sh. After about 5 hours model with 87.5% F1 will be learnt. In snapshots directory will be saved the model with the best F1 (each 2 epochs). Learning logs also saving there.

For example, learning log for 74 epoch:

processed 46666 tokens with 5648 phrases; found: 5778 phrases; correct: 4990.
accuracy:  97.65%; precision:  86.36%; recall:  88.35%; FB1:  87.34
              LOC: precision:  90.35%; recall:  90.89%; FB1:  90.62  1678
             MISC: precision:  73.82%; recall:  75.50%; FB1:  74.65  718
              ORG: precision:  80.69%; recall:  85.01%; FB1:  82.79  1750
              PER: precision:  93.87%; recall:  94.74%; FB1:  94.31  1632

References

• Named entity recognition using syntactic and semantic features and neural networks
• Natural Language Processing (almost) from Scratch
• http://ml.nec-labs.com/senna/
• https://github.com/patverga/torch-ner-nlp-from-scratch
• https://github.com/attardi/deepnl