EECS 496: Advanced Topics in Deep Learning
Final Project: Chinese Question Answering with BERT (Baidu DuReader Dataset)
The DuReader dataset is a machine reading comprehension (MRC) dataset in Chinese. It is the rough equivalent to the popular Stanford Question Answering Dataset (SQuAD) in English. Because DuReader comprises additional question types beyond SQuAD, namely "yes/no" and entity detection, we ignored those components and only predicted answer spans within the input paragraph, as with regular MRC models.
To Download DuReader dataset:
git clone git@github.com:baidu/DuReader.git
cd DuReader/data && bash download.sh
We format the DuReader dataset in the format identical to what BERT uses for the SQuAD dataset:
python3 src/preprocessing/dr_to_squad.py [path/to/dureader.processed.json]
Additional flags can be found in dr_to_squad.py or run python3 src/preprocessing/dr_to_squad.py --help.
BERT is a new method of pre-training transformers for a variety of NLP tasks, including QA-IR. It achieved state of the art results on the SQuAD dataset so we wanted to apply it to DuReader.
First install the PyTorch implementation of BERT with Huggingface's python package (https://github.com/huggingface/pytorch-pretrained-BERT):
pip install pytorch-pretrained-bert
Using our preprocessed training files or larger training sets obtained with our preprocessing script, run the training script with the command:
python run_dureader.py --bert_model bert-base-chinese --do_train --do_lower_case --train_file data/20000_search.train.json --train_batch_size 12 --gradient_accumulation_steps 3 --learning_rate 3e-5 --num_train_epochs 2.0 --max_seq_length 384 --doc_stride 128 --output_dir ../duoutput
The hyperparameter setting has been tested on GTX1080 8GB. Generating features from the training file for the Chinese IR task can take a long time with the current scripts and it is CPU-only. For a training set of size 10000, it takes about 8-10 hours on our setup. Training itself is 1-2 hours depending on the hyperparameters. After training is complete, use the following command to generate predictions for the 1000-example preprocessed development set included in this repository:
python run_dureader.py --bert_model ../duoutput --do_predict --predict_file data/1000_search.dev.json --max_seq_length 384 --doc_stride 128 --output_dir ../duprediction
BLEU scoring is an algorithm used to evaluate the quality of text. It has a fairly high correlation to human judgement and is significantly better than accuracy. DuReader evaluates results based off of the BLEU scoring metric.
The BERT model will output predictions.json.
python3 src/bleu/bert_bleu.py [path/to/predictions.json] [path/to/preprocessed.json]