Constructing set-compositional and negated representations for first-stage ranking

Code for reproducing the CIKM 2025 paper, by Antonios Minas Krasakis, Andrew Yates, Evangelos Kanoulas

Reproducing experiments

Zero-shot experiments (RQ1): QUEST: zeroshot.sh

Finetuning scripts, models with vanilla splade and SNRelu/negative term weights (RQ2, RQ3): Quest: quest_ft.sh NevIR: nevir_ft.sh

Figure 3: analysis_negation_proximity.py

Notes:

• (bash) scripts are written to run on a slurm environment. In case slurm is not detected if [ -z "$SLURM_ARRAY_TASK_ID" ], a python debugger (debugpy) client is starting. To execute without debugging change to PYTHON_OR_DEBUGPY=python '
• queries are expected to be JSONL files under /home/akrasakis/data/quest_domain_split/queries_jsonl/questTRAIN.jsonl or questTEST.jsonl. Note that for this paper, we re-split the QUEST train/test set keeping all “Film and “Book” category queries in the training set and “Plants” and “Animals” in the test set (reasoning and details can be found on Section 4.1 of the paper). Unfortunately, I have lost access to the cluster/data where experiments were performed, and I am unable to provide the models or data files. I can guide you to reproduce them if needed though, so please drop me an email!

For questions, or collaborations feel free to contact me

If you find this research or code useful, please use the following citation:

@inproceedings{krasakis2025constructing,
title={Constructing set-compositional and negated representations for first-stage ranking},
author={Krasakis, Antonios Minas and Yates, Andrew and Kanoulas, Evangelos},
booktitle={Proceedings of the 34th ACM International Conference on Information and Knowledge Management},
pages={1406--1416},
year={2025}
}