GRASP entity linking results

Linking results, GRASP logs and configs for the GRASP entity-linking experiments, plus baseline linkers for comparison, visualised with ELEVANT.

🔗 Browse the results online in the ELEVANT evaluation webapp.

Experiments

GRASP was benchmarked on: Wiki-Fair v2.0 (no coref), News-Fair v2.0 (no coref), KORE50, MSNBC (updated), AIDA-CoNLL (test).

Using the LLMs: Gemma 4 31B IT, GPT-5.4 mini, Qwen3.5-397B, Qwen3.6-27B, Mistral Small 4, Kimi K2.7.

• 9 additional experiments with Qwen3.6-27B on KORE50 using different random seeds for generation.
• 1 additional experiment with a reduced knowledge-graph function set (base) using Qwen3.6-27B on KORE50.

Contents

• evaluation_results — linking results of GRASP, plus the ReFinED and GENRE linkers and an oracle for comparison, and Claude Code entity-linking baselines (Claude Haiku 4.5 and Claude Sonnet, one headless subagent per article querying only a Wikidata QLever endpoint) evaluated on KORE50, MSNBC (updated), and a 20-article subsample of Wiki-Fair v2.0 (no coref) (wiki-fair-v2-no-coref-sample20)
• benchmarks — benchmark files (ELEVANT JSONL) with metadata and statistics
• grasp_logs — GRASP logs of the linking process for the different experiments
• configs — GRASP config files used for the different experiments

ELEVANT setup

We use ELEVANT to visualise the results.

Clone the repository and build the image

git clone https://github.com/ad-freiburg/elevant
cd elevant
docker build -t elevant .

Run the container

# create a dir for caching elevant data
mkdir data
docker run -it -p 8617:8000 \
  -v $(pwd)/data/:/data \
  -v $(pwd)/../evaluation_results/:/home/evaluation-results \
  -v $(pwd)/benchmarks/:/home/benchmarks \
  elevant

# inside the container
make download-all   # downloads data into /data, only needs to be run once
make start-webapp

You can now access port 8617 to browse the evaluation webapp.

← Entity-linking data release