gxulmh / mner

PyTorch code for the paper "Adaptive Co-Attention Network for Named Entity Recognition in Tweets"

MNER

PyTorch code for the paper "Adaptive Co-Attention Network for Named Entity Recognition in Tweets" published in AAAI 2018 The authors released their Keras implementation (https://github.com/jlfu/NERmultimodal)

Dependencies:

pytorch (tested on version 0.3.0.post4, should work fine on newer 1.0.0 version as well)
torchcrf
tqdm

This model is slightly different from the model explained in the paper. Instead of coAttn, we use stackedAttn.

The perforamcne is similar to the CNN+BiLSTM+CRF (Ma and Hovy 2016) model as mentioned in the paper.

Data can be downloaded from http://pan.baidu.com/s/1boSlljL

Train a model using the following command:

python main.py --split_file /data/extDisk2/vvaibhav/vner/NERmultimodal/data --word2vec_model /data/extDisk2/vvaibhav/vner/glove.6B.200d.word2vec --image_features_dir /data/extDisk2/vvaibhav/vner/feat_bu/ --mode 0 --visual_feature_dimension 2048 --regions_in_image 36 --hidden_dimension 256 --hidden_dimension_char 50 --embedding_dimension 200 --embedding_dimension_char 30 --use_char_embedding 1 --batch_size 10  --use_filter_gate 1 --gamma 0.05 --dropout 0.2 --use_only_text 0 --lr 0.015

To evaluate a trained model on test set:

python main.py --split_file /data/extDisk2/vvaibhav/vner/NERmultimodal/data --image_features_dir /data/extDisk2/vvaibhav/vner/feat_bu/ --hidden_dimension 256 --hidden_dimension_char 50 --embedding_dimension 200 --embedding_dimension_char 30 --use_char_embedding 1 --batch_size 10 --visual_feature_dimension 2048 --regions_in_image 36 --use_filter_gate 1 --gamma 0.05 --dropout 0.2 --use_only_text 0 --lr 0.015 --mode 1 --model_file_name best_model_weights_13_0.934.t7

If you found this code useful in your research, please consider acknowledging the repo.