This repo hosts the code for the following paper:
Back to the Future: Unsupervised Backprop-based Decoding for Counterfactual and Abductive Commonsense Reasoning
Lianhui Qin, Vered Shwartz, Peter West, Chandra Bhagavatula, Jena Hwang, Ronan Le Bras, Antoine Bosselut, Yejin Choi
EMNLP 2020
- The code consists of the implementations for the two tasks, namely counterfactual reasoning and abductive reasoning, respectively.
- Example small data is included in
data/ - Ranking code is included in
ranking/ - Decoding and ranking results will be put in
output/
Note: it seems important to use the same versions of pytorch and transformers as listed in requirements.txt. Using different versions may produce different generations.
Run the following cmd to do DeLorean decoding for counterfactual reasoning, on the example data in data/counterfactual/small_data.json
sh run_counterfactual_main.sh
Results are written to output/counterfactual/, which include the generated hypotheses using different hyperparameters (#forward-backward passes and #backward iterations). These results are then to be ranked in the following.
Run the following cmds to rank the hypotheses
cd ranking/
sh run_counterfactual_ranking.sh
Ranked results are written to output/counterfactual/ranking
The code and usage are largely the same as those of counterfactual reasoning. We write different code files for different data processing, loss functions, etc.
Run the following cmd to do DeLorean decoding for abductive reasoning, on the example data in data/abductive/small_data.json
sh run_abductive_main.sh
Results are written to output/abductive/, which include the generated hypotheses using different hyperparameters (#forward-backward passes and #backward iterations). These results are then to be ranked in the following.
Run the following cmds to rank the hypotheses
cd ranking/
sh run_abductive_ranking.sh
Ranked results are written to output/abductive/ranking
Acknowledgement: the decoding and ranking code uses Huggingface Transformers. The decoding code is adapted (though with large changes) from the Plug-and-Play LM code.