icetea233/ampersand-emnlp2019

Code and Data for EMNLP 2019 paper titled AMPERSAND: Argument Mining for PERSuAsive oNline Discussions

Please leave a star if you use our code

AMPERSAND-EMNLP2019

Code and Data for EMNLP 2019 paper titled AMPERSAND: Argument Mining for PERSuAsive oNline Discussions These models were fintuned using older version of hugging face and not transformers package . used pytorch-pretrained-bert=0.4.0

There are two steps of fine tuning involved here.

                    Intermediate fine-tuning on Distantly labeled data for improved representation learning
                    Task Specific Fine-tuning on labeled data

I have provided Intermediate fine-tuned model on Distantly labeled data below, however you have to train on task specific data If you want to train on Hidey et al (2017) the dataset used in this paper, get data here https://github.com/chridey/change-my-view-modes

You can also email me [email protected] for a preprocessed version.

To load finetuned models on distantly labeled data IMHO (intra relation / claim - premise ) classification and QR for inter relation classification

load respective models in this line https://github.com/tuhinjubcse/AMPERSAND-EMNLP2019/blob/master/argmining/examples/run_classifier.py#L498

For Argumentative Component Prediction and Relation Prediction: Link to FineTuned Pytorch Model using IMHO+Context as Intermediate Pretraining over BERT: https://drive.google.com/uc?id=11U_kLNn6ngPltWQ1raN16SSy8tQ9fpL5&export=download

Link to QR fine-tuned model https://drive.google.com/file/d/1wWs_0pb2N9dmXz6RjnW7TiJkV-b1m9Np/view?usp=sharing

Link to IMHO+Context dataset: You can choose to keep the IMO/IMHO keywords. We removed them based on Chakrabarty et al (2019) https://drive.google.com/file/d/1HGInaDp6nlAZUfqU5V1BM8j4su3DKKsc/view?usp=sharing

Link to QR dataset https://drive.google.com/file/d/10l96wL1VQlApC1h0LPOjUpGAtyRZvMPO/view?usp=sharing

To reproduce our results in paper follow the details in AMPERSAND_Supplementary.pdf uploaded

Load fine-tuned models instead of pretrained BERT and use the data mentioned below to further fine-tune on task specific data

If you want to know more see this issue tuhinjubcse#2

For RST :

Use https://www.aclweb.org/anthology/D18-1116.pdf for EDU segmentation Use for getting RST parse trees https://github.com/jiyfeng/DPLP Once you have parse trees you can get features from them

For Summarization

https://github.com/nlpyang/BertSum https://drive.google.com/file/d/1iyPb_z775V7qVRD8_kGXCFxGuSoku3Hr/view?usp=sharing (doc-summary pairs)

If you use any of these , please cite us

    @article{chakrabarty2020ampersand,
    title={AMPERSAND: Argument Mining for PERSuAsive oNline Discussions},
    author={Chakrabarty, Tuhin and Hidey, Christopher and Muresan, Smaranda and Mckeown, Kathy and Hwang, Alyssa},
    journal={arXiv preprint arXiv:2004.14677},
    year={2020}
  }