Code for the paper Multi-step Retriever-Reader Interaction for Scalable Open-domain Question Answering
Acknowledgement: This codebase started from the awesome Dr.QA repository created and maintained by Adam Fisch. Thanks Adam!
The requirements are in the requirements file. In my env, I also needed to set PYTHONPATH (as in the setup.sh)
pip install -r requirements.txt
source setup.sh
We are making the pre-processed data and paragraph vectors available so that is is easier to get started. They can downloaded from here. (41GB compressed, 56GB decompressed; user/pass: guest/guest). If you need the pretrained paragraph encoder used to generate the vectors, feel free to get in touch with me. After un-taring, you will find a directory corresponding to each dataset. Each directory further contains:
data/ -- Processed data (*.pkl files)
paragraph_vectors/ -- Saved paragraph vectors of context for each dataset used for nearest-neighbor search
vocab/ -- int2str mapping
embeddings/ -- Saved lookup table for faster initialization. The embeddings are essentially saved fast-text embeddings.
If you want to train new paragraph embeddings instead of using the ones we used, please refer to this readme
python scripts/reader/train.py --data_dir <path-to-downloaded-data> --model_dir <path-to-downloaded-model> --dataset_name searchqa|triviaqa\quasart --saved_para_vectors_dir <path-to-downloaded-data>/dataset_name/paragraph_vectors/web-open
Some important command line args
dataset_name -- searchqa|triviaqa|quasart
data_dir -- path to dataset that you downloaded
model_dir -- path where model would be checkpointed
saved_para_vectors_dir -- path to cached paragraph and query representations in disk. It should be in the data you have downloaded
multi_step_reasoning_steps -- Number of steps of interaction between retriever and reader
num_positive_paras -- (Relevant during training) -- Number of "positive" (wrt distant supervision) paragraphs fed to train to the reader model.
num_paras_test -- (Relevant during inference time) -- Number of paragraphs to be sent to the reader by the retriever.
freeze_reader -- when set to 1, the reader parameters are fixed and only the parameters of the GRU (multi-step-reasoner) is trained.
fine_tune_RL -- fune tune the GRU (multi-step-reasoner) with reward (F1) from the fixed reader
Training details:
- During training, we first train the reader model by setting
multi_step_reasoning_steps = 1 - After the reader has been trained, we fix the reader and just pretrain the
multi-step-reasoner(freeze_reader 1) - Next, we fine tune the reasoner with reinforcement learning (
freeze_reader = 1, fine_tune_RL = 1)
In our experiments for searchqa and quasart, we found step 2 (pretraining the GRU was not important) and the reasoner was directly able to learn via RL. However, pretraining never hurt the performance as well.
We are also providing pretrained models for download and scripts to run them directly. Download the pretrained models from here.
Usage: /bin/bash run_pretrained_models.sh dataset_name data_dir model_dir out_dir
dataset_name -- searchqa|triviaqa|quasart
data_dir -- path to dataset that you downloaded
model_dir -- path to pretrained model that you downloaded
out_dir -- directory for logging
- Integrate with code for SGTree
@inproceedings{
das2018multistep,
title={Multi-step Retriever-Reader Interaction for Scalable Open-domain Question Answering},
author={Rajarshi Das and Shehzaad Dhuliawala and Manzil Zaheer and Andrew McCallum},
booktitle={ICLR},
year={2019},
}