Attribute structuring(AS)

Attribute structuring based GPT scoring of generated text summaries. Given a ground truth discharge summary and a summary generated by a model, AS uses an ontology to extract important variables, and scores the generated summary using GPT4.

1. Access to MIMIC III dataset (https://physionet.org/content/mimiciii/)
2. Manual annotation of selected MIMIC Discharge Summaries. data/annotations.json contains annotations for 30 documents by 3 annotators.
3. Access to GPT APIs. Here we used GPT4 via Microsoft Azure API to structure and score the summaries.

1. install the libraries in the requirements.txt file
2. add your api_base and api_key to your system environment
3. modify/set the gpt parameters inside src/variables [OPTIONAL]
4. replace the [MIMIC_DIR] inside src/main with your mimic path
5. -- run the src/main.py script