Neural Biomedical Question Answering

This repository collects the source code of the work published in Neural Domain Adaptation for Biomedical Question Answering. It has been written by Dirk Weissenborn and Georg Wiese at the German Research Center for Artificial Intelligence (DFKI).

Inference Server via Docker

To merely try out our pre-trained models, the simplest way is to use the pre-built Docker image. To start server with single model on port 5000:

$ docker run -it -p 127.0.0.1:5000:5000 \
  georgwiese/biomedical-qa \
  ./start_server.sh single

The server accepts POST requests to localhost:5000/answer which contain a payload in BioASQ JSON format.

Training a Model

The code structure in this repository is as follows:

• notebooks: Contains Jupyter Notebooks for analysis.
• biomedical_qa: Root directory for all python code.
• biomedical_qa/tools: All command line tools.

Data Sources

Download the datasets from SQuAD and BioASQ (Task B training dataset). The BioASQ dataset needs to be split into dev and train datasets, e.g. via biomedical_qa/tools/split_bioasq.py.

Build Embedders

Download the GloVe and PubMed Embeddings. Then, use biomedical_qa/tools/embedder_from_glove_and_pubmed.py to create transfer models that are used by the QA model.

Example Training

Here is an example to train a model from scratch on SQuAD:

python3 biomedical_qa/training/train_qa.py \
    --with_question_type_features \
    --size 100 \
    --max_epochs 40 \
    --model_type simple_pointer \
    --data data/squad \
    --trainset_prefix train \
    --validset_prefix dev \
    --start_output_unit sigmoid \
    --ckpt_its 1000 \
    --batch_size 64 \
    --save_dir $SAVE_DIR \
    --dropout 0.5 \
    --dataset squad \
    --transfer_model_config model_checkpoints/glove_pubmed_embedder/config.pickle \
    --transfer_model_path model_checkpoints/glove_pubmed_embedder/model.tf \
    --composition LSTM

It assumes the SQuAD train and dev JSONs are located at data/SQuAD and the embedder is stored at model_checkpoints/glove_pubmed_embedder.

Fine-tuning on BioASQ works analogously by passing the --model_config and --init_model_path flags.