Unsupervised Entity Disambiguation using Pretrained Contextualized Word Embeddings

This repository contains the code for the equally named master thesis at UHH's LT Group.

Requirements

All code in this repository is written in Python (python 3.9.6 to be precise). Python dependencies are managed with Pipenv, which needs to be installed. Then, to install dependencies run:

pipenv install

To build the models and run the experiments, we need datasets. In the thesis, we use two models:

• AIDA YAGO
• Wikipedia Dump

AIDA YAGO

For license issues, you need to provide the AIDA dataset yourself. Place the full AIDA TSV in aida/aida-yago2-dataset/AIDA-YAGO2-dataset.tsv, the Makefile will split it into training and test files.

Wikipedia

The dump, that we use, consists of one JSON object per line. Each line is a JSON object with the keys id, title and text corresponding to one Wikipedia page.

Run

The Makefile holds all commands used to build the models and run the experiments. Our pipeline consists of multiple Python files that all load and save pickles.