Entity Spaces for Tolerant Entity Linking: Code and Data
Date of Posting: July 18, 2019
This package contains the software and data necessary for running our experiments for applying the concept of entity spaces to the task of entity linking. We also include the results of our experiments.
This repo is easier to understand if you are reading the associated paper, which is currently under review.
The taggers here run on python3.
We assume you're running in a virtualenv
pip install -r requirements.txt
You'll need to have installed:
In ./amstel - you'll find the taggers used in the pipeline.
All data and evaluation results and scripts are in /data
- All gold standard data is included in this repo under (data/evaulation_datasets) except for AIDA Yago, which we are not allowed to redistribute.
Experimental results are in this includes intermediate output.
Shell scripts to run experiments are in:
- Note, you'll need to uncomment different parts of the scripts to run el and ed variants for the neural based entity linker. This is because the end2end_neural_el runs on the same port for both services.
- This will generate results in the associated file.
Paper Pipelines to Scripts mapping:
- [neural el] - Run end2end_neural_el in its entity linking setting. This uses tagger_neural_fulldoc2.py.
- [neural ed] - Run end2end_neural_el in it entity disambiguation setting. This uses tagger_neural_fulldoc_disambig.py
- [neural d1.0] - parameters are shown in the scripts
- [neural d0.8] - parameters are shown in the scripts
- [mag ed] - Run AGDISTIS. Assumes you are running locally. You can modify tagger_agdistis_fulldoc.py to use the remote services.
- [mag d1.0] - parameters are shown in the scripts
- [mag d0.8] - arameters are shown in the scripts
To generate the evaluation data for non-linkable entities:
- run the script conll_conversion/convertSysOutput2CoNLLEvalInput_NMEsKeptIn.py on the data in data/experiments-results-agdistis and data/experiments-results/neural
- results can be found in data/evaluate_with-non-linkable-entities_agdistis and data/evaluate_with-non-linkable-entities_neural
Please contact Marieke van Erp for any questions.