NeuralRST-TopDown

Paper

Fajri Koto, Jey Han Lau, and Timothy Baldwin. Top-down Discourse Parsing via Sequence Labelling. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2021): Main Volume.

About the code

This code uses LSTM, if you want to use the Transformer version please find it here. The encoder is designed based on Yu et al., 2018 where we use three main embeddings:

Word embedding, initialized by glove.6B.200d.txt.gz.
POS Tags embedding, initialized randomly.
Syntax Embedding from BiAffine Dependency Parser. Please refer to RSTExtractor to see how we extract it

Dependencies

Python 3.6
Run pip install -r requirements.txt

Data and Resource

We use English RST Tree Bank. Please make sure you have a right to access this data. Our code uses the input of the binarized discourse tree, provided by Yu et al., 2018.

In this repository, we do not provide you with the raw RST Tree Bank, but the binarized version split in train/dev/test based on Yu et al., 2018. We also provide the extracted syntax feature for each data split. Please download them here.

Running the code

In the experiment we use 1 GPU V100 (16GB).

For training the LSTM with static oracle (normal training)

CUDA_VISIBLE_DEVICES=0 python train_rst_parser.py --experiment=exp_static \
                        --word_embedding_file=[path_to_glove] \
                        --train=[path_to_train_data] --test=[path_to_test_data] --dev=[path_to_dev_data] \
                        --train_syn_feat=[path_to_syntax_feature_of_train] \
                        --test_syn_feat=[path_to_syntax_feature_of_test] \
                        --dev_syn_feat=[path_to_syntax_feature_of_dev] \
                        --max_sent_size=100 --hidden_size=256 --hidden_size_tagger=128 --batch_size=4 \
                        --grad_accum=2 --lr=0.001 --ada_eps=1e-6 --gamma=1e-6 \
                        --loss_seg=1.0 --loss_nuc_rel=1.0 --depth_alpha=0 \
                        --elem_alpha=0.35

For training the LSTM with the dynamic oracle:

CUDA_VISIBLE_DEVICES=0 python train_rst_parser.py --experiment=exp_dynamic \
                        --word_embedding_file=[path_to_glove] \
                        --train=[path_to_train_data] --test=[path_to_test_data] --dev=[path_to_dev_data] \
                        --train_syn_feat=[path_to_syntax_feature_of_train] \
                        --test_syn_feat=[path_to_syntax_feature_of_test] \
                        --dev_syn_feat=[path_to_syntax_feature_of_dev] \
                        --max_sent_size=100 --hidden_size=256 --hidden_size_tagger=128 --batch_size=4 \
                        --grad_accum=2 --lr=0.001 --ada_eps=1e-6 --gamma=1e-6 \
                        --loss_seg=1.0 --loss_nuc=1.0 --beam_search=1 --depth_alpha=0 \
                        --elem_alpha=0.35 --use_dynamic_oracle=1 --start_dynamic_oracle=50 --oracle_prob=0.65

Models

We also provide the result of static and dynamic training of our LSTM model. Please download them here. To run this model, please download all data for the input, and adjust the config.cfg accordingly. This includes word_embedding_file, train_path, test_path, dev_path, train_syn_feat_path, dev_syn_feat_path, test_syn_feat_path, model_path, model_name, alphabet_path.

You can run the model by

CUDA_VISIBLE_DEVICES=0 python run_rst_parser.py --config_path=path_to_config.cfg

Output for static training, (rst = RST Parseval, ori = original Parseval):

Reading dev instance, and predict...
S (rst): Recall: R=3338/3886=0.859, Precision: P=3338/3886=0.859, Fmeasure: 0.859
N (rst): Recall: R=2840/3886=0.7308, Precision: P=2840/3886=0.7308, Fmeasure: 0.7308
R (rst): Recall: R=2389/3886=0.6148, Precision: P=2389/3886=0.6148, Fmeasure: 0.6148
F (rst): Recall: R=2372/3886=0.6104, Precision: P=2372/3886=0.6104, Fmeasure: 0.6104
-----------------------------------------------------------------------------------
S (ori): Recall: R=1395/1943=0.718, Precision: P=1395/1943=0.718, Fmeasure: 0.718
N (ori): Recall: R=1203/1943=0.6191, Precision: P=1203/1943=0.6191, Fmeasure: 0.6191
R (ori): Recall: R=1022/1943=0.526, Precision: P=1022/1943=0.526, Fmeasure: 0.526
F (ori): Recall: R=1004/1943=0.5167, Precision: P=1004/1943=0.5167, Fmeasure: 0.5167

Reading test instance, and predict...
S (rst): Recall: R=3998/4616=0.8661, Precision: P=3998/4616=0.8661, Fmeasure: 0.8661
N (rst): Recall: R=3419/4616=0.7407, Precision: P=3419/4616=0.7407, Fmeasure: 0.7407
R (rst): Recall: R=2827/4616=0.6124, Precision: P=2827/4616=0.6124, Fmeasure: 0.6124
F (rst): Recall: R=2811/4616=0.609, Precision: P=2811/4616=0.609, Fmeasure: 0.609
-----------------------------------------------------------------------------------
S (ori): Recall: R=1690/2308=0.7322, Precision: P=1690/2308=0.7322, Fmeasure: 0.7322
N (ori): Recall: R=1440/2308=0.6239, Precision: P=1440/2308=0.6239, Fmeasure: 0.6239
R (ori): Recall: R=1167/2308=0.5056, Precision: P=1167/2308=0.5056, Fmeasure: 0.5056
F (ori): Recall: R=1148/2308=0.4974, Precision: P=1148/2308=0.4974, Fmeasure: 0.4974

Output for dynamic training, (rst = RST Parseval, ori = original Parseval):

Reading dev instance, and predict...
S (rst): Recall: R=3349/3886=0.8618, Precision: P=3349/3886=0.8618, Fmeasure: 0.8618
N (rst): Recall: R=2858/3886=0.7355, Precision: P=2858/3886=0.7355, Fmeasure: 0.7355
R (rst): Recall: R=2422/3886=0.6233, Precision: P=2422/3886=0.6233, Fmeasure: 0.6233
F (rst): Recall: R=2409/3886=0.6199, Precision: P=2409/3886=0.6199, Fmeasure: 0.6199
-----------------------------------------------------------------------------------
S (ori): Recall: R=1406/1943=0.7236, Precision: P=1406/1943=0.7236, Fmeasure: 0.7236
N (ori): Recall: R=1228/1943=0.632, Precision: P=1228/1943=0.632, Fmeasure: 0.632
R (ori): Recall: R=1035/1943=0.5327, Precision: P=1035/1943=0.5327, Fmeasure: 0.5327
F (ori): Recall: R=1019/1943=0.5244, Precision: P=1019/1943=0.5244, Fmeasure: 0.5244

Reading test instance, and predict...
S (rst): Recall: R=4005/4616=0.8676, Precision: P=4005/4616=0.8676, Fmeasure: 0.8676
N (rst): Recall: R=3399/4616=0.7364, Precision: P=3399/4616=0.7364, Fmeasure: 0.7364
R (rst): Recall: R=2849/4616=0.6172, Precision: P=2849/4616=0.6172, Fmeasure: 0.6172
F (rst): Recall: R=2818/4616=0.6105, Precision: P=2818/4616=0.6105, Fmeasure: 0.6105
-----------------------------------------------------------------------------------
S (ori): Recall: R=1697/2308=0.7353, Precision: P=1697/2308=0.7353, Fmeasure: 0.7353
N (ori): Recall: R=1440/2308=0.6239, Precision: P=1440/2308=0.6239, Fmeasure: 0.6239
R (ori): Recall: R=1206/2308=0.5225, Precision: P=1206/2308=0.5225, Fmeasure: 0.5225
F (ori): Recall: R=1173/2308=0.5082, Precision: P=1173/2308=0.5082, Fmeasure: 0.5082

fajri91 / neuralrst-topdown Goto Github PK