Filtered Semantic Search via Vector Arithmetic
Keywords: passage retrieval, dense retrieval, feature representation
TL;DR: We describe a lightweight modification to vector-based semantic search that efficiently restricts results to those matching a specified, semantically-meaningful filter.
Abstract: How can we retrieve search results that are both semantically relevant and satisfy certain filter criteria? Modern day semantic search engines are increasingly reliant on vector-based search, yet the ability to restrict vector search to a fixed set of filter criteria remains an interesting problem with no known satisfactory solution. In this note, we leverage the rich emergent structure of vector embeddings of pre-trained search transformers to offer a simple solution. Our method involves learning, for each filter, a vector direction in the space of vector embeddings, and adding it to the query vector at run-time to perform a search constrained by that filter criteria. Our technique is broadly applicable to any finite set of semantically meaningful filters, compute-efficient in that it does not require modifying or rebuilding an existing $k$-NN index over document vector embeddings, lightweight in that it adds negligible latency, and widely compatible in that it can be utilized with any transformer model and $k$-NN algorithm. We also establish, subject to mild assumptions, an upper bound on the probability that our method errantly retrieves irrelevant results, and reveal new empirical insights about the geometry of transformer embeddings. In experiments, we find that our method, on average, yields more than a 21% boost over the baseline (measured in terms of nDCG@10) across three different transformer models and datasets.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 7731
Loading