NIL_EL

This repository maintains code, data and for ACL Findings 2023 paper

Learn to Not Link: Exploring \nil{} Prediction in Entity Linking

Get Started

Environment

git clone https://github.com/solitaryzero/NIL_EL.git
cd NIL_EL/src
pip install -r requirements.txt

Build the NEL Dataset

The scripts/preprocess folder contains shell scripts for building the NEL dataset.

vectorize.sh builds the standard NEL dataset, vectorize_no_nil.sh builds the dataset without NIL entries, vectorize_partial.sh and vectorize_partial_nep.sh build the dataset with only a portion of NIL or NEP entries.

Training and Evaluation

In the scripts/nel, scripts/nel_nonil, scripts/nel_partial and scripts/nel_partial_nep folders, train_[model].sh works for training and evaluate_[model].sh works for evaluation.

The model names refer to:

blink: the standard bi-encoder model
clink: bi-encoder with type prediction
cross: the standard cross-encoder model
cross_clink: cross-encoder with type prediction

Evaluation on Standard Datasets

The scripts/standard folder contains the evaluation scripts for standard datasets like AIDA, MSNBC and WNED.

The standard dataset benchmarks can be downloaded at [Google Drive]

solitaryzero / NIL_EL