Automatic Clinical Relation Extraction

Task - Identifying test results and measurements within clinical documents in Italian, Spanish and Basque.

Dataset for Italian - https://e3c.fbk.eu/clinkart Dataset for Spanish and Basque - https://e3c.fbk.eu/testlinkiberlef

Preliminary Instructions

• Download the zip file of the datasets and keep them in the repo main folder.

Fine-tuning Transformer models

Done in 2 stages

Flowchart of the Relation Extraction training process

3 models used for both stages - mBERT, XLM-Roberta and BERT; fine-tuned with monolingual and multilingual data.

In Fine-Tuning-Transformers/

• sh prepare_script.sh - to prepare required data
• create_multilingual_data.py - to create multilingual data from each language dataset.
• train_ner*.py - to train on the NER task.
• infer_ner*.py - to infer on the NER task.
• rc_dataset*.py - to create the Relation Classification dataset.
• train_rc*.py - to train on the RC task.
• infer_rc*.py - to infer on the RC task.
• sweep scripts are for hyperparameter tuning using wandb sweeps.

LLM Few shot learning

Few shot learning with gpt-4-turbo, gpt-4 and gpt-3.5-turbo

Files inside LLMs\

Error Analysis

In Error_Analysis/

Quantitative and Qualiatative Analysis of NER and Relation Extraction output.