Paper: Argument Undermining: Counter-Argument generation by Attacking Weak Premises

This is the code for the paper Argument Undermining: Counter-Argument generation by Attacking Weak Premises.

Milad Alshomary, Shahbaz Syed, Martin Potthast, and Henning Wachsmuth

  @InProceedings{alshomary:2021b,
    author =              {Milad Alshomary, Shahbaz Syed, Martin Potthast, and Henning Wachsmuth},
    booktitle =           {The Joint Conference of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (ACL-IJCNLP 2021)},
    month =               August,
    publisher =           {ACL-IJCNLP},
    title =               {{Argument Undermining: Counter-Argument generation by Attacking Weak Premises}},
    year =                2021
  }

---

Data:

The data folder consists of the following:

• predictions: contains all counter-generation predictions of our approach and the baselines.
• evaluations: contains annotations results for all our manual evaluation studies.
• cmv_dataset: would be made available soon.

Code:

• prepare_ds.ipynb : Contains the code for processing Jo et al. 2020 dataset and collecting counter-arguments. For this notebook to work, you need the data provided in Jo et al. 2020. The code folder contains all python file and notebooks necessary to reproduce our results:

• premise_attackability: Contains the code for training our LTR-bert model (ltr_identify_vunerability.ipynb) and the commands needed to re-train BERT-classifier of Jo et al. 2020. For these notebooks to work, you need the tensorflow ranking library (https://github.com/tensorflow/ranking) and the code of Jo et al (https://github.com/yohanjo/emnlp20_arg_attack)

• attack_generation: Contains the code for fine-tuning GPT model and the baseline, as well as generting attacks on argument given the weak-premises.

• overall_approach.ipynb: Contains our code for the overall approach, that is identifying weak premises using ltr-bert and then generate attacks.

• evaluation: Contains all notebooks for evaluating all steps of our approach