ELEVANT is a tool that helps you evaluate, analyse and compare entity linking systems in detail. You can explore a demo instance of the ELEVANT web app at https://elevant.cs.uni-freiburg.de/. If you are using ELEVANT for your research please cite our paper "ELEVANT: A Fully Automatic Fine-Grained Entity Linking Evaluation and Analysis Tool".
For the ELEVANT instance of the EMNLP 2023 paper "A Fair and In-Depth Evaluation of Existing End-to-End Entity Linking Systems" see https://elevant.cs.uni-freiburg.de/emnlp2023.
Get the code, and build and start the docker container:
git clone https://github.com/ad-freiburg/elevant.git .
docker build -t elevant .
docker run -it -p 8000:8000 -v <data_directory>:/data -v $(pwd)/evaluation-results/:/home/evaluation-results -v $(pwd)/benchmarks/:/home/benchmarks elevant
where <data_directory> is the directory in which the required data files will be stored. What these data files are
and how they are generated is explained in section Get the Data. Make sure you can read from and
write to all directories that are being mounted as volumes from within the docker container (i.e. your
<data_directory>, evaluation-results and benchmarks), for example (if security is not an issue) by giving all
users read and write permissions to the directories in question with:
chmod a+rw -R <data_directory> evaluation-results/ benchmarks/
Unless otherwise noted, all the following commands should be run inside the docker container. If you want to use the system without docker, follow the instructions in Setup without Docker before continuing with the next section.
(Note: If you want to use a custom knowledge base instead of Wikidata/Wikipedia/DBpedia you can skip this step and instead follow the instructions in Use a Custom Knowledge Base.)
For linking entities in text or evaluating the output of a linker, our system needs information about entities and mention texts, e.g. entity names, aliases, popularity scores, types, the frequency with which a mention is linked to a certain article in Wikipedia, etc. This information is stored in and read from several files. Since these files are too large to upload them on GitHub, you can either download them from our servers (fast) or build them yourself (slow, RAM intensive, but the resulting files might be based on a more recent Wikidata/Wikipedia dump).
To download the files, simply run
make download_all
This will automatically run make download_wikidata_mappings, make download_wikipedia_mappings and
make download_entity_types_mapping which will download the compressed files, extract them and move them to the
correct location. See Mapping Files for a description of files downloaded in this steps.
NOTE: This will overwrite existing Wikidata and Wikipedia mappings in your <data_directory> so make sure this is what
you want to do.
If you rather want to build the mappings yourself, you can replace each download command (except download_all) by a
generate command. See
Data Generation for more details.
To start the evaluation web app, run
make start_webapp
You can then access the webapp at http://0.0.0.0:8000/.
The evaluation results table contains one row for each experiment. In ELEVANT, an experiment is a run of a particular entity linker with particular linker settings on a particular benchmark. We already added a few experiments, including oracle predictions (perfect linking results generated from the ground truth), so you can start exploring the web app right away. The section Add a Benchmark explains how you can add more benchmarks and the section Add an Experiment explains how you can add more experiments yourself.
See Evaluation Web App for a detailed overview of the web app's features.
You can easily add a benchmark if you have a benchmark file that is in the JSONL format we use, in the common NLP Interchange Format (NIF), in the IOB-based format used by Hoffart et al. for their AIDA/CoNLL benchmark or in a very simple JSONL format. Benchmarks in other formats have to be converted into one of these formats first.
To add a benchmark, simply run
python3 add_benchmark.py <benchmark_name> -bfile <benchmark_file> -bformat <ours|nif|aida-conll|simple-jsonl>
This converts the <benchmark_file> into our JSONL format (if it is not in this format already), annotates ground
truth labels with their Wikidata label and Wikidata types and writes the result to the file
benchmarks/<benchmark_name>.benchmark.jsonl.
The benchmark can now be linked with a linker of your choice using the link_benchmark_entities.py script with the
parameter -b <benchmark_name>. See section Add an Experiment for details on how to link a
benchmark and the supported formats.
See Add A Benchmark for more details on adding a benchmark including a description of the supported file formats.
Many popular entity linking benchmarks are already included in ELEVANT and can be used with ELEVANT's scripts out of the box. See Benchmarks for a list of these benchmarks.
You can add an experiment, i.e. a row in the table for a particular benchmark, in two steps: 1) link the benchmark articles and 2) evaluate the linking results. Both steps are explained in the following two sections.
To link the articles of a benchmark with a single linker configuration, use the script link_benchmark_entities.py:
python3 link_benchmark_entities.py <experiment_name> -l <linker_name> -b <benchmark_name>
The linking results will be written to
evaluation-results/<linker_name>/<adjusted_experiment_name>.<benchmark_name>.linked_articles.jsonl where
<adjusted_experiment_name> is <experiment_name> in lowercase and characters other than [a-z0-9-] replaced by
_.
For example
python3 link_benchmark_entities.py Baseline -l baseline -b kore50
will create the file evaluation-results/baseline/baseline.kore50.linked_articles.jsonl. The result file contains
one article as JSON object per line. Each JSON object contains benchmark article information such as the article
title, text, and ground truth labels, as well as the entity mentions produced by the specified linker.
<experiment_name> is the name that will be displayed in the first column of the evaluation results table in the
web app.
See Link Benchmark Articles for information on how you can transform your existing linking result files into our format, and instructions for how to link multiple benchmarks using multiple linkers with a single command.
See Included Linkers for a list of linkers that can be used out of the box with ELEVANT. These are for example ReFinED, OpenAI's GPT (you'll need an OpenAI API key for that), REL, TagMe (you'll need an access token for that which can be obtained easily and free of cost) and DBpediaSpotlight.
To evaluate a linker's predictions use the script evaluate_linking_results.py:
python3 evaluate_linking_results.py <path_to_linking_result_file>
This will print precision, recall and F1 scores and create two new files where the .linked_articles.jsonl file
extension is replaced by .eval_cases.jsonl and .eval_results.json respectively. For example
python3 evaluate_linking_results.py evaluation-results/baseline/baseline.kore50.linked_articles.jsonl
will create the files evaluation-results/baseline/baseline.kore50.eval_cases.jsonl and
evaluation-results/baseline/baseline.kore50.eval_results.json. The eval_cases file contains information about
each true positive, false positive and false negative case. The eval_results file contains the scores that are shown
in the web app's evaluation results table.
In the web app, simply reload the page (you might have to disable caching) and the experiment will show up as a row in the evaluation results table for the corresponding benchmark.
See Evaluate Linking Results for instructions on how to evaluate multiple linking results with a single command.
If you want to remove an experiment from the web app, simply (re)move the corresponding .linked_articles.jsonl,
.eval_cases.jsonl and .eval_results.json files from the evaluation-results/<linker_name>/ directory and reload
the web app (again disabling caching).