This is the code repository for the paper "Modeling Non-Cooperative Dialogue: Theoretical and Empirical Insights" to appear in TACL.
The code to reproduce experiments is given in this repository and the data is available soon on OSF.
This repository is highly dependent on existing implementations for GuessWhat?! in the Cooperative Setting. Specifically, see this repository provided by Strub et al. and De Vries et al.. Before running the code as described below, one should complete the steps required of this dependency and establish an identical experimental environment (i.e., see README-dependency.md or the original README provided by the dependency). We have already copied over code and added needed submodules, so these steps can be followed within our repository without making changes to the code (e.g., besides resolving any version issues - we used all the same versions as the dependency).
You can skip RL training for the question-generator's conversation policy (aka QGen), since our code modifies this part. We have also included some redundant data in our OSF storage, so some other steps may be skipped too.
In general, our code contributions are demarcated with _dim standing for "deception identification modification". We use deception as an identifier for components designed for the non-cooperative setting, so this can be used to help navigate the code to make modifications.
To train the non-cooperative oracle and train QGen using RL in the non-cooperative setting, respectively, we use the following files:
train_oracle_dim.pytrain_qgen_reinforce_2oracle_w_guesser.py
Arguments for train_oracle_dim.py are identical to the counterpart in the Cooperative Setting train_oracle.py besides the data directory. We provide some example scripts in oracle-examples.sh. The script qgen-rl-examples.sh provides commands to run train_qgen_reinforce_2oracle_w_guesser.py as well.
Non-Cooperative Data: https://osf.io/2k6vu/
Code for Cooperative Setting: https://github.com/GuessWhatGame/guesswhat
Cooperative Setting Data Proposal: https://openaccess.thecvf.com/content_cvpr_2017/papers/de_Vries_GuessWhat_Visual_Object_CVPR_2017_paper.pdf
Cooperative Setting Model Proposal: https://www.ijcai.org/proceedings/2017/0385.pdf
PyTorch Implementation of Cooperative Setting: https://github.com/shekharRavi/Beyond-Task-Success-NAACL2019
If you use our implementation or ideas for the non-cooperative setting, please cite out paper (forthcoming).
Please also provide appropriate citation for the original cooperative setting.
If you run into issues with the code or data, don't hesitate to contact us directly using the emails provided on our publication. You can also raise issues here.