Multi-BiDAF: Multiple Sentences Bi-directional Attention Flow

Multi Sentences Bi-directional Attention Flow (Multi-BiDAF) network is a model designed to fit the BiDAF model of Seo et al. (2017) for the Multi-RC dataset. This implementation is built on the AllenNLP library.

Installation

To install Multi-BiDAF, start by cloning our git repository:

$ git clone https://github.com/eitanhaimashiah/multibidaf.git

Create a Python 3.6 virtual environment, and install the necessary requirements by running:

$ ./scripts/install_requirements.sh

(The above is assuming CUDA 9 installed on a linux machine; use a different pytorch version as necessary.)

Training Multi-BiDAF

Once you've installed Multi-BiDAF, you can train our model fully by running:

$ ./scripts/train_fully.sh

When you run this it will compute an unified vocabulary for the SQuAD and MultiRC datasets, pretrain the Multi-BiDAF model on SQuAD, and eventually train the model on MultiRC. Each of these tasks can be accomplished by running separate scripts (scripts/make_unified_vocab.sh, scripts/pretrain_on_squad.sh, scripts/train_on_multirc.sh, respectively).

Moreover, you can create a prediction file (adapted to the official MultiRC evaluation script) of the development set by running:

$ ./scripts/predict_multirc_dev.sh