Source codes for the paper Cognitive Graph for Multi-Hop Reading Comprehension at Scale. (ACL 2019 Oral)
We also have a Chinese blog about CogQA on Zhihu (η₯δΉ) besides the paper.
CogQA is a novel framework for multi-hop question answering in web-scale documents. Founded on the dual process theory in cognitive science, CogQA gradually builds a cognitive graph in an iterative process by coordinating an implicit extraction module (System 1) and an explicit reasoning module (System 2). While giving accurate answers, our framework further provides explainable reasoning paths.
- Download and setup Redis database following https://redis.io/download
- Download the dataset, evalute script and fullwiki data (enwiki-20171001-pages-meta-current-withlinks-abstracts) from https://hotpotqa.github.io. Unzip
improved_retrieval.zip
in this repo. pip install -r requirements.txt
- Run
python read_fullwiki.py
to load wikipedia documents to redis (check the size ofdump.pkl
is about 2.4GB). - Run
python process_train.py
to generatehotpot_train_v1.1_refined.json
, which contains edges in gold-only cognitive graphs. mkdir models
The codes automatic assign tasks on all available devices, each handling batch_size / num_gpu
samples. We recommend that each gpu has at least 11GB memory to hold 2 batch.
- Run
python train.py
to train Task #1(span extraction). - Run
python train.py --load=True --mode='bundle'
to train Task #2(answer prediction).
The cogqa.py
is the algorithm to answer questions with a trained model. We split the 1-hop nodes found by another similar model into fullwiki_input_improved_by_cogqa1hop.zip
for reuse in other algorithm. It can directly improve your result on fullwiki setting by just replacing the original input.
-
unzip
fullwiki_input_improved_by_cogqa1hop.zip
. -
python cogqa.py --data_file='hotpot_dev_fullwiki_v1_merge.json'
-
python hotpot_evaluate_v1.py hotpot_dev_fullwiki_v1_merge_pred.json hotpot_dev_fullwiki_merge_v1.json
-
You can check the cognitive graph (reasoning process) in the
cg
part of the predicted json file.
- The changes of this version from the preview version is mainly about detailed comments.
- The relatively sensetive hyperparameters includes the number of negative samples, top K, learning rate of task #2, scale factors between different parts...
- If our work is useful to you, please cite our paper or star π our repo~~