MultiQA

MultiQA is a project whose goal is to facilitate training and evaluating reading comprehension models over arbitrary sets of datasets. All datasets are in a single format, and it is accompanied by an AllenNLP DatasetReader and model that enable easy training and evaluation on multiple subsets of datasets.

This repository contains the code for our paper MultiQA: An Empirical Investigation of Generalization and Transfer in Reading Comprehension.

This work was performed at The Allen Institute of Artificial Intelligence.

This project is constantly being improved. Contributions, comments and suggestions are welcome!

Datasets

Link to the single format datasets are provided in to formats: MultiQA, and SQuAD2.0. The SQuAD2.0 are GZipped JSONs that are the results of applying convert_multiqa_to_squad_format to the MultiQA dataset. To used them with Pytorch-Transformers code simply unzip them. (see models Readme for an example)

Dataset	MultiQA format	SQuAD2.0 format (GZipped)
SQuAD-1.1	train , dev	train , dev
SQuAD-2.0	train , dev	train , dev
NewsQA	train , dev	train , dev
HotpotQA	train , dev	train , dev
TriviaQA-unfiltered	dev	dev
TriviaQA-wiki	train , dev	train , dev
SearchQA	train , dev	train , dev
BoolQ	train , dev	train , dev
ComplexWebQuestions	train , dev	train , dev
DROP	train , dev	train , dev
WikiHop	train , dev	train , dev
DuoRC Paraphrase	train , dev	train , dev
DuoRC Self	train , dev	train , dev
ComplexQuestions	train , dev	train , dev
ComQA	train , dev	train , dev

Datasets will be addeed weekly, so please stay tuned!

Models and Results

Trained Models for download and results will be posted in this table. The BERT-Base column contains evaluation results (EM/F1) as well as a link to the trained model. The MultiQA-5Base column contain the link to the model (in the header) and evalution results for this model. This model is BERT-Base that has been trained on 5 datasets. The pytorch-transformers columns contain .bin models trained with pytorch-transformers code added.

Dataset	BERT-Base AllenNLP	BERT Base uncased Pytorch-Transformers	MultiQA-5Base AllenNLP (model)	BERT-Large AllenNLP
SQuAD-1.1	80.1 / 87.5 (model)	80.2 / 87.7 (model)	81.7 / 88.8	83.3 / 90.3 (model)
NewsQA	47.5 / 62.9 (model)	47.5 / 62.3 (model)	48.3 / 64.7	50.3 / 66.0 (model)
HotpotQA	50.1 / 63.2 (model)		-	54.0 / 67.0 (model)
TriviaQA-unfiltered	59.4 / 65.2 (model)		59.0 / 64.7	60.7 / 66.5 (model)
TriviaQA-wiki	57.5 / 62.3 (model)		-	-
SearchQA	58.7 / 65.2 (model)		58.8 / 65.3	60.5 / 67.3 (model)

multiqa commands

In order to simply train BERT on multiple datasets (with AllenNLP) please use:

python multiqa.py train --datasets SQuAD1-1,NewsQA,SearchQA --cuda_device 0,1,2,3
python multiqa.py evaluate --model SQuAD1-1 --datasets SQuAD1-1,NewsQA,SearchQA --cuda_device 0

By default the output will be stored in models/datatset1_dataset2_... to change this please change --serialization_dir

Type python multiqa.py for additional options.

Note, this version uses the default multiqa format datasets stored in s3, to use your own dataset please see Readme for using allennlp core commands.

MultiQA format to SQuAD2.0 format

If you prefer using SQuAD2.0 format, or run the Pytorch-Trasformers models, please use:

python convert_multiqa_to_squad_format.py --datasets https://multiqa.s3.amazonaws.com/data/HotpotQA_dev.jsonl.gz --output_file data/squad_format/HotpotQA_dev.json

Setup

Setting up a virtual environment

First, clone the repository:

git clone https://github.com/alontalmor/MultiQA.git

Change your directory to where you cloned the files:
```
cd MultiQA
```

Create a virtual environment with Python 3.6 or above:

virtualenv venv --python=python3.7 (or python3.7 -m venv venv or conda create -n multiqa python=3.7)

Activate the virtual environment. You will need to activate the venv environment in each terminal in which you want to use MultiQA.
```
source venv/bin/activate (or source venv/bin/activate.csh or conda activate multiqa)
```
Install the required dependencies:
```
pip install -r requirements.txt
```

Tests

You can test all challenges using pytest, or using pycharm tests directory (pytest-pycharm added):

pytest pytests

Data

The allennlp caching infra is used, so be sure to have enough disk space, and control the cache directory using ALLENNLP_CACHE_ROOT env variable.

Build Dataset

This will take a dataset from it's original URL and output the same dataset in the MultiQA format.

python build_dataset.py --dataset_name HotpotQA --split train --output_file path/to/output.jsonl.gz --n_processes 10

Predict

first argument is the allennlp model, second is the preprocessed evalutaion file ( path/to/output.jsonl.gz in preprocess), then the dataset name (in order to create the official predictions format)

python predict.py --model https://multiqa.s3.amazonaws.com/models/BERTBase/SQuAD1-1.tar.gz --dataset https://multiqa.s3.amazonaws.com/data/SQuAD1-1_dev.jsonl.gz --dataset_name SQuAD

To predict only a the first N examples use --sample_size N

To add a GPU device simply append: --cuda_device 0

By default the output will be saved at results/DATASET_NAME/... You may also change the output filename and path using --prediction_filepath path/to/my/output

Multiqa Data Format

see Readme in the datasets folder. A json-schema for a single context in multiqa is available here.

Training using AlleNLP

see Readme in the models folder.

Training using Pytorch-Trasformers