Official repository of "Interlocking-free Selective Rationalization Through Genetic-based Learning".
We split GenSPP and selective rationalization baselines in two standalone projects: genetic, and baselines.
Please, see the README.md file in genetic folder.
Please, see the README.md file in baseline folder.
Don't hesitate to file an issue if you find some bugs!
@inproceedings{ruggeri-signorelli-2025-interlocking,
title = "Interlocking-free Selective Rationalization Through Genetic-based Learning",
author = "Ruggeri, Federico and
Signorelli, Gaetano",
editor = "Che, Wanxiang and
Nabende, Joyce and
Shutova, Ekaterina and
Pilehvar, Mohammad Taher",
booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
month = jul,
year = "2025",
address = "Vienna, Austria",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.acl-long.59/",
doi = "10.18653/v1/2025.acl-long.59",
pages = "1175--1191",
ISBN = "979-8-89176-251-0",
abstract = "A popular end-to-end architecture for selective rationalization is the select-then-predict pipeline, comprising a generator to extract highlights fed to a predictor. Such a cooperative system suffers from suboptimal equilibrium minima due to the dominance of one of the two modules, a phenomenon known as interlocking. While several contributions aimed at addressing interlocking, they only mitigate its effect, often by introducing feature-based heuristics, sampling, and ad-hoc regularizations. We present GenSPP, the first interlocking-free architecture for selective rationalization that does not require any learning overhead, as the above-mentioned. GenSPP avoids interlocking by performing disjoint training of the generator and predictor via genetic global search. Experiments on a synthetic and a real-world benchmark show that our model outperforms several state-of-the-art competitors."
}
Eleonora Misino, Paolo Torroni
This publication is supported by the project European Commission’s NextGeneration EU programme, PNRR – M4C2 – Investimento 1.3, Partenariato Esteso, PE00000013 - “FAIR - Future Artificial Intelligence Research” – Spoke 8 “Pervasive AI” and by the European Union’s Justice Programme under Grant Agreement No. 101087342 for the project “Principles Of Law In National and European VAT”.