Axiomatic Re-Ranking for Argument Retrieval

In this paper, four new argumentative axioms are introduced. These axioms consider the similarity of document sentences—specifically argument units—with respect to the query.

Experiments

The experiments involve a re-ranking of the Dirichlet-LM baseline from the 2020 and 2021 editions of the Touché Shared Task on Argument Retrieval. Additionally, we re-rank the participant submissions for Touché 2021 to demonstrate that the axioms can improve systems that already outperform the Dirichlet-LM baseline.

Code Structure

Due to difficulties with conflicting Python packages, the code is divided into two parts:

1. Socket Server
   Acts as a database for the embeddings. In additition to the requirement stated in requirements_socket.txt, you have to download the spacy corpus with python -m spacy download en_core_web_lg We used python 3.10
2. Argument Re-Ranking
   Uses ir_axioms to perform the argument re-ranking. We used python 3.8

Touché Results

Scores for the Touché participants can be found here:

https://github.com/touche-webis-de/TOUCHE-21/tree/master/touche21-data/runs

For access, please send a message to webis@listserv.uni-weimar.de.