Subtopic-oriented biomedical extractive summarization

Work done for my Bachelor's thesis.

Summarization framework based on subtopics and extractive summarization using pretrained language models.

Framework evaluation

Installation

Install the dependencies by running:

pip install -r framework/requirements.txt

Dataset preprocessing

Refer to biomed-ext-summ for converting an abstractive dataset into an extractive one.

Evaluation

To evaluate a configuration of the framework, run:

python eval/eval.py                             \
    --dataset=<dataset name>                    \
    --dataset-path=<path to extractive dataset> \
    --embedding=<embedding>

Available <embedding> are bow, word2vec, glove, fasttext, biowordvec, minilm, biobert and pubmedbert.

Instead of --embedding, the following options are available:

• --plain to evaluate using only a pretrained model (without the framework).
• --oracle to evaluate using the reference extractive summary.

Web app prototype

This is a prototype to present the result of the framework to the end user.

To start the application, run in the repository root:

docker compose up

By default, the application will be served at http://localhost:8001.

Note that nvidia-container-toolkit is required in order to use the GPU in a container.