NMN-MultiHopQA-dev

This repo contains the source code for the following paper

• Yichen Jiang and Mohit Bansal, "Self-Assembling Modular Networks for Interpretable Multi-Hop Reasoning" in EMNLP, 2019. (paper).
• The basic code structure was adapted from BiDAF.
• The NMN code was adapted from SNMN.

1. Preparation

Dependencies

• We tested our code on TF1.3, TF1.8, TF1.11 and TF1.13.
• See requirements.txt.

Data

• Run download_data.sh to pull HotpotQA data and GloVe vectors.

Preprocessing

• Run the following command to preprocess HotpotQA data:

python3 -m hotpotqa.prepro --find_bridge_entity

The preprocessed data will be save at data/hotpotqa/.

Model checkpoint

• Download our trained model here and put it under out/hotpotqa/NMN/.

2. Evaluate and Train the Model

Evaluation

• Run the following command to evaluate the NMN on HotpotQA dev set:

python3 -m basic.cli --mode=test --batch_size=32 --dataset=hotpotqa --len_opt --run_id=00 --reasoning_layer=snmn --nmn_attention_type=biattn --nmn_cfg --supervise_bridge_entity --self_att --nmn_span_pred_hop=second --nmn_separate_controllers

• The predicted answer will be saved to out/hotpotqa/NMN/[RUN_ID]/answer/test-[GLOBAL_STEPS].json.
• To help visualizing our NMN, the predicted module probability and question decomposition (attention) probability at every hop will be saved to out/hotpotqa/NMN/[RUN_ID]/answer/test-[GLOBAL_STEPS]-module_prob_ques_attn.json.

Training

• Run the following command to train a NMN for HotpotQA distractor setting:

python3 -m basic.cli --mode=train --batch_size=24 --dataset=hotpotqa --len_opt --run_id=01 --reasoning_layer=snmn --nmn_attention_type=biattn --nmn_cfg --supervise_bridge_entity --self_att --nmn_span_pred_hop=second --nmn_separate_controllers --train_nmn_ctrl_separately --occasional_train_nmn_ctrl --train_nmn_ctrl_period=3768 --num_steps=80000 --noload

The model can be trained with a single Nvidia 1080Ti GPU.

Citation

@inproceedings{Jiang2019Self-Assembling, 
  author={Yichen Jiang and Mohit Bansal}, 
  booktitle={Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing}, 
  title={Self-Assembling Modular Networks for Interpretable Multi-Hop Reasoning}, 
  year={2019}, 
}