This repository contains the practical implementations corresponding to the graduate thesis "Explaining Relationships Between Academic Documents Using Generative Transformer Models" as part of the MSc. Data Science in Business and Economics at the University of Tübingen. In this work, amongst others, a novel conditional context representation for the two papers is proposed and evaluated across several state-of-the-art Transformer language models. A schematic overview of this new method is shown below.
For more details, please contact me to receive the entire thesis.
The demo notebook demo_notebook.ipynb contains an exemplary application of the methods discussed and developed in the corresponding paper and invites for further experiments. To use it and experiment with the fine-tuned models, please download the corresponding Python module relationship_explanation from Google Drive.
Generally, to use the provided scripts, please create a new Python 3 environment and execute the following command to to install all packages and dependencies.
pip install -r requirements.txt
Apart from the provided scripts, larger data files are made available in a corresponding Google Drive folder. This includes the pre-trained dygie model scierc.tar.gz (source), the pre-trained joint tagger joint_tagger_train_scibert_final.model (source) from the CORWA pipeline as well as the scigen and scigpt2 folders (source). For more details, please refer to the README files in the separate folders.