Topological Sort for Sentence Ordering

Code accompanying the paper: Topological Sort for Sentence Ordering

Pre-requisites

• Python 3.6
• Pytorch 1.0.1
• transformers 2.3.0
• requirements for using the transformers code
• Glove embeddings (840B tokens, 2.2M vocab, cased, 300d vectors, 2.03 GB download) for the L-TSort model

Data

For the AAN, NIPS and NSF data please contact the authors of Sentence Ordering and Coherence Modeling using Recurrent Neural Networks.

The SIND dataset can be downloaded from the Visual Storytelling website.

Quickstart

Refer to example.sh file to see the commands.

• Use prepare_data.py script to create train, valid and test data to train classifier for AAN, NIPS, SIND and NSF datasets.

python prepare_data.py --data_dir nips/ --out_dir nips_data/ --task_name nips

BERT-based representation (B-TSort Classifier)

The model.py script in bert_classifier should be used to train and test the B-TSort model.

• Train the B-TSort models using the following command:

python model.py --data_dir ../nips_data/ --output_dir ../trained_models/nips_bert/ --do_train --do_eval --evaluate_during_training --per_gpu_train_batch_size 32 --per_gpu_eval_batch_size 16

• Run the trained B-TSort model on the test set. This script creates a test_results.tsv file and saves it in the data_sir path provided. This is later useful to get the sentence orders for the document using topological sort script.

python model.py --data_dir ../nips_data/ --output_dir ../trained_models/nips_bert/checkpoint-X/ --do_test --per_gpu_eval_batch_size 16

LSTM-based representation (L-TSort Classifier)

The code base for this classifier is in lstm_classifier folder.

• Create a data dir in the lstm_classifier folder and run the following preprocess command:

python preprocess.py -train_src ../nips_data/train.tsv -valid_src ../nips_data/dev.tsv -save_data data/nips_data -src_vocab_size 16721 -src_embedding PATH_TO/glove.840B.300d.txt -seq_length 100

• Train the L-TSort models using the following command:

python train.py -data data/nips_data.train.pt -brnn -gpus 1 -sequence_length 105 -pre_word_vecs_enc data/nips_data.embed.pt -save_model nips_lstm/model

• Test the L-TSort models:

python translate.py -model MODEL_NAME.pt -src ../nips_data/test.tsv -max_sent_length 100 -gpu 0 -output nips_lstm_results.tsv

Topological Sort

Run the topological sort script on the outputs of the B-TSort and L-TSort models to calculate results for various metrics.

python topological_sort.py --file_path nips_data/test_results.tsv

Trained Models

• Download the trained B-TSort models for each of the four datasets from the below links:

http://tts.speech.cs.cmu.edu/sentence_order/nips_bert.tar
http://tts.speech.cs.cmu.edu/sentence_order/aan_bert.tar
http://tts.speech.cs.cmu.edu/sentence_order/nsf_bert.tar
http://tts.speech.cs.cmu.edu/sentence_order/sind_bert.tar

• Download the trained L-TSort models for NIPS and SIND datasets from the below links:

http://tts.speech.cs.cmu.edu/sentence_order/nips_lstm.pt
http://tts.speech.cs.cmu.edu/sentence_order/sind_lstm.pt

Baselines

For the code of the model denoted as AON in the paper, please email the authors of Deep Attentive Sentence Ordering Network.

Contributors

If you use this code please cite the following:

Prabhumoye, Shrimai, Ruslan Salakhutdinov, and Alan W. Black. "Topological Sort for Sentence Ordering." In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics.

@inproceedings{topSort2020,
  title={Topological Sort for Sentence Ordering},
  author={Prabhumoye, Shrimai and Salakhutdinov, Ruslan and Black, Alan W},
  booktitle={Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
  publisher={Association for Computational Linguistics},
  year={2020},
  }