Do Multi-hop Readers Dream of Reasoning Chains?

Notes:

The codes are modified based on the original BERT repo. Some unrelated modules from the original repo have been deleted in order to make it easy to understand.

Data:

We provide the processed data mentioned in the paper under data/.

Training

Please refer to run.sh for the entry point

Inference

Please refer to run.sh for the entry point

Results

random: 75.27%
single_oracle: 77.86%
ordered_oracle: 79.15%
ordered_oracle_with_token: 79.49%
oredered_oracle_co_matching: 80.77%

to evaluate using the script: python evaluate-v1.1.py data/hotpot_dev_distractor_v1.ordered_support_only.json <model_folder>/prediction.json