SLAHAN: Syntactically Look-A-Head Attention Network for Sentence Compression

SLAHAN is an implementation of Kamigaito et al., 2020, "Syntactically Look-A-Head Attention Network for Sentence Compression", In Proc. of AAAI2020.

Citation

@article{Kamigaito2020SyntacticallyLA,
  title={Syntactically Look-Ahead Attention Network for Sentence Compression},
  author={Hidetaka Kamigaito and Manabu Okumura},
  journal={ArXiv},
  year={2020},
  volume={abs/2002.01145}}

Prerequisites

• GCC 7.5.0 (supporting the C++11 language standard)
• CMake 3.14.1
• Boost Libraries version 1.63
• CUDA 10.1 and 10.0
• HDF5 Library 1.10.4
• Python 3.69
• TensorFlow 1.15.0 (Requires CUDA 10.0)
• AllenNLP 0.9.0
• h5py 2.8.0
• Perl 5.26.2
• Java
• Bash

Build Instruction

Set the environment value BOOST_ROOT:

export BOOST_ROOT="/path/to/boost"

and run:

./scripts/setup.sh

Prediction & Evaluation

You can predict the test sets of Google dataset and BNC Corpus by the following commands:

### Prediction for the Google dataset ###
./scripts/predict/google/predict.sh
### Prediction for the BNC Corpus ###
./scripts/predict/bcn/predict.sh

Precisions, recalls, f-measures, and compression ratios for each model are stored in the file results/{dev or test}_{dataset type}_{evaluation metric}.csv, respectively.

You can also calculate rouge scores by using the ROUGE-1.5.5 script. After the setup of ROUGE-1.5.5 and the execution of predict.sh, you can ran the following command for obtaining rouge scores for each model:

### Calculate rouge for the Google dataset ###
./scripts/predict/google/rouge.sh
### Calculate rouge for the BNC Corpus ###
./scripts/predict/bcn/rouge.sh

Rouge scores are also stored in the file results/test_{dataset type}_{evaluation metric}.csv, respectively.

You can also calcute compression rates in characters:

### Calculate compression rates in characters for the Google dataset ###
./scripts/predict/google/char_len.sh
### Calculate compression rates in characters for the BNC Corpus ###
./scripts/predict/bcn/char_len.sh

Compression rates in characters are also stored in the file results/test_{dataset type}_char_cr.csv, respectively.

Note that the reference compression is located on dataset/, and the system compression is located on models/{dataset size}/{model name}_{process id}/comp.txt.

Compressed Sentences

The compressed sentences used for the evaluation in our paper are included in the directory share.

Retrain the Models

Before retraining the models, you should extract features of the training data set.

./scripts/train/google/extract_features.sh

Note that this step includes feature extractions of BERT, ELMo and Glove. It takes almost 1 day and 300GB of disk space. After that, you can retrain each model by the below command.

./scripts/train/google/{model name}.sh {process_id}

All trained models are saved in the directory models/{dataset name}/{model name}_{process id}/save_{epoch}. To reproduce our results, run the following commands:

### Tagger ###
./scripts/train/google/tagger.sh 0
./scripts/train/google/tagger.sh 1
./scripts/train/google/tagger.sh 2

### LSTM ###
./scripts/train/google/lstm.sh 0
./scripts/train/google/lstm.sh 1
./scripts/train/google/lstm.sh 2

### LSTM-Dep ###
./scripts/train/google/lstm-dep.sh 0
./scripts/train/google/lstm-dep.sh 1
./scripts/train/google/lstm-dep.sh 2

### Attn ###
./scripts/train/google/attn.sh 0
./scripts/train/google/attn.sh 1
./scripts/train/google/attn.sh 2

### Base ###
./scripts/train/google/base.sh 0
./scripts/train/google/base.sh 1
./scripts/train/google/base.sh 2

### Parent w syn ###
./scripts/train/google/parent_w_syn.sh 0
./scripts/train/google/parent_w_syn.sh 1
./scripts/train/google/parent_w_syn.sh 2

### Parent w/o syn ###
./scripts/train/google/parent_wo_syn.sh 0
./scripts/train/google/parent_wo_syn.sh 1
./scripts/train/google/parent_wo_syn.sh 2

### Child w syn ###
./scripts/train/google/child_w_syn.sh 0
./scripts/train/google/child_w_syn.sh 1
./scripts/train/google/child_w_syn.sh 2

### Child w/o syn ###
./scripts/train/google/child_wo_syn.sh 0
./scripts/train/google/child_wo_syn.sh 1
./scripts/train/google/child_wo_syn.sh 2

### SLAHAN w syn ###
./scripts/train/google/slahan_w_syn.sh 0
./scripts/train/google/slahan_w_syn.sh 1
./scripts/train/google/slahan_w_syn.sh 2

### SLAHAN w/o syn ###
./scripts/train/google/slahan_wo_syn.sh 0
./scripts/train/google/slahan_wo_syn.sh 1
./scripts/train/google/slahan_wo_syn.sh 2

After these processes, you can run the following prediction and evaluation scripts:

./scripts/predict/google/predict.sh
./scripts/predict/google/rouge.sh
./scripts/predict/google/char_len.sh
./scripts/predict/bcn/predict.sh
./scripts/predict/bcn/rouge.sh
./scripts/predict/bcn/char_len.sh

Finally, you can obtain results of the models in the directory ./results.

LICENSE

MIT License