Axiomatic Thinking

Synopsis

Axiomatic thinking in information retrieval formalizes what a good ranking should look like as a set of constraints, called axioms. Each axiom usually describes a pairwise preference that any reasonable retrieval model should satisfy, for example, that of two otherwise similar documents, the one containing the query terms more often should be ranked higher. Axioms thus make the behavior of retrieval systems explicit and testable.

Our research has used axiomatic constraints to explain the decisions of neural ranking models [ICTIR 2021] or which retrieval principles they might have learned [ICTIR 2026]. Axioms are also used to directly improve rankings, by aggregating their pairwise preferences for re-ranking [CIKM 2016]. We also apply axiomatic re-ranking in specialized retrieval domains such as argument retrieval [Touché 2022] [ SIGIR 2025]. We also transfer axiomatic thinking to retrieval-augmented generation, with axioms that capture desirable properties of generated answers; this work received a best paper honorable mention [ICTIR 2025] [award].

Our open-source Python framework ir_axioms implements many common retrieval axioms as well as axioms for retrieval-augmented generation, and integrates with retrieval frameworks like PyTerrier and Pyserini [code] [pypi] [publication]. The ir_axioms documentation includes extensive references to further related work [documentation].

People

Publications