Reproduce “LogiCoL: Logically-Informed Contrastive Learning for Set-based Dense Retrieval” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yanzhen Shen, Sihao Chen, Xueqiang Xu, Yunyi Zhang, Chaitanya Malaviya, Dan Roth
LOGICOL introduces a logically-informed contrastive learning framework that enhances dual- and bi-encoder dense retrievers by explicitly encoding logical relations from queries. The method significantly reduces logical consistency violations by over 20% and improves retrieval recall for intersection and negation queries by up to 5% on augmented datasets, allowing models to better distinguish between logically distinct queries.
Key insights
- Standard dense retrievers struggle to distinguish between logically distinct queries (e.g., 'A AND B' vs. 'A NOT B') due to their focus on semantic similarity, leading to high correlation in their similarity distributions.
- Strategically grouping logically related queries within training batches (e.g., variants of A and B) is crucial for models to learn and encode logical distinctions in their embedding space.
- Explicitly incorporating logical consistency constraints (exclusion and subset relations) into the contrastive learning objective is effective for ensuring retrieval results align with logical intent.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.