FlowNet

Data and code for "Linguistic Versus Latent Relations for Modeling a Flow in Paragraphs" by Dongyeop Kang, Hiroaki Hayashi, Alan W Black, and Eduard Hovy, EMNLP 2019

Dependency

• tensorflow
• nltk
• gensim
• nlgeval

Notes

• (Aug 2019) This repository is not heavily tested due to the major refactoring of flownet with pytorch and transfomers. Please stay tuned for the new pytorch version of flownet.

Dataset preparation

Once you download the PeerRead and BookCorpus datasets, preprocess them using the scripts under prepare_{arxiv,book}.py. For discourse relations, we parse each paragraph using DPLP. The final dataset looks like below: one from BookCorpus's SciFi and other from PeerRead's paper.

[Category] \t [ID] \t [Paragraph Length] \t [Avg Sentence Length] \t [Paragraph #1] \t ... [Paragraph #7] \t [Flattened RST tree]
Science_fiction	u3633	7	13.86	" Don ' t talk to him , " Baasen snapped .	Dark blooms appeared on his yellow-green cheeks .	" If you talk to anyone , you talk to me , and you don ' t talk to me .	You just march yourself to the ship .	I ' ve been kind not to break you .	Jabba don ' t care if you still got knees when you reach him .	So don ' t- "	( NS-elaboration ( EDU " Don ' t talk to him , " Baasen snapped . )  ( NN-list ( EDU Dark blooms appeared on his yellow-green cheeks . )  ( NS-elaboration ( EDU " If you talk to anyone , you talk to me , and you don ' t talk to me . )  ( NS-elaboration ( EDU You just march yourself to the ship . )  ( NN-list ( NS-elaboration ( EDU I ' ve )  ( EDU been kind not to break you . )  )  ( NN-list ( NS-elaboration ( EDU Jabba don ' t )  ( NS-condition ( EDU care )  ( SN-condition ( EDU if you still got knees )  ( EDU when you reach him . )  )  )  )  ( NS-elaboration ( EDU So don ' t- )  ( EDU " )  )  )  )  )  )  )  )
1. Introduction	1405.7908	4	15.75	Comp and Decomp are variations on the unsupervised learning algorithm of Turney ( 2012 ) .	Super is based on the supervised algorithm of Turney ( 2013 ) .	These algorithms were originally designed for the recognition task .	The main contribution of this paper is to show that , working together , the algorithms can scale up from recognition to generation .	( NS-elaboration ( EDU Comp and Decomp are variations on the unsupervised learning algorithm of Turney ( 2012 ) . )  ( NS-elaboration ( NS-elaboration ( EDU Super is based on the supervised algorithm of Turney ( 2013 ) . )  ( EDU These algorithms were originally designed for the recognition task . )  )  ( SN-attribution ( EDU The main contribution of this paper is to show )  ( EDU that , working together , the algorithms can scale up from recognition to generation . )  )  )  )

Models

In order to experiment with (and hopefully improve) our models, you can run following commands:

To run flownet with delta relations, run

    python train_lm.py \
        --data_path ${DATA_PATH} \
        --dataset ${DATASET_NAME} \
        --model_path ${MODEL_PATH} \
        --model_name drnn

To run flownet with discourse relations, run

    python train_clm.py \
        --data_path ${DATA_PATH} \
        --dataset ${DATASET_NAME} \
        --model_path ${MODEL_PATH} \
        --model_name rstseq2seq

You can add --reload or --test options to reload the pre-trained model or test on testing data, respectively.

Citation

@inproceedings{kang19emnlp_flownet,
  title = {Linguistic Versus Latent Relations for Modeling a Flow in Paragraphs},
  author = {Dongyeop Kang and Hiroaki Hayashi and Alan W Black and Eduard Hovy},
  booktitle = {Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  address = {Hong Kong},
  month = {November},
  url = {https://arxiv.org/abs/1908.11790},
  year = {2019}
}

Acknowledgement

• We also thank Jason Weston, Dan Jurafsky, and anonymous reviewers for their helpful comments.