SympTEMIST

Preparation

1. Clone xMEN repository (https://github.com/hpi-dhc/xmen) to obtain dataloaders and benchmark script:

   git clone https://github.com/hpi-dhc/xmen
   cd xmen
   poetry install
   

2. Get latest version of SympTEMIST gazetteer from Zenodo (https://zenodo.org/records/10635215) and put the TSV into xmen/local_files

3. Prepare KB and indicies for candidate generation:

   xmen dict benchmarks/benchmark/symptemist.yaml --code examples/dicts/bsc_gazetteer.py
   xmen dict benchmarks/benchmark/symptemist.yaml --all
   

Notebooks

Notebook	Description
0_Dataset.ipynb	Statistics for SympTEMIST shared task and comparable datasets
1_LLM_Simplification.ipynb	Applying LLM-based text simplification

Experiments

Executing 1_LLM_Simplification.ipynb results in a dataset of candidates based on simplified mentions (symptemist_candidates_simplified_cutoff) in the current folder. It can be used as a candidate set for running the full SympTEMIST entity linking pipeline with a trainable re-ranker. The BERT checkpoint for initializing the cross-encoder can also be adapted.

cd xmen/benchmarks

Without Simplification

python run_benchmark.py benchmark=symptemist output=./training

With Simplification (Pre-Processed Candidates)

python run_benchmark.py benchmark=symptemist output=./training +candidates_path=../../symptemist/symptemist_candidates_simplified_cutoff

With Simplification and Alternative BERT Checkpoint

(e.g. for PlanTL-GOB-ES/roberta-base-bne)

python run_benchmark.py benchmark=symptemist output=./training +candidates_path=../../symptemist/symptemist_candidates_simplified_cutoff linker.reranking.training.model_name=PlanTL-GOB-ES/roberta-base-bne