Search papers, labs, and topics across Lattice.
This paper introduces CoDeR, a local constraint-compatible retrieval method that addresses the limitations of traditional information retrieval systems which rely solely on semantic similarity. By separating topical relevance from constraint compatibility through a bi-encoder trained with lexical-polarity supervision, CoDeR effectively mitigates the issue of constraint-violating evidence exposure. Evaluation results demonstrate significant improvements over existing methods, with reductions in V@2 scores and enhanced ranking of violating documents across multiple benchmarks.
CoDeR achieves up to a 23.53 point reduction in V@2 scores, revolutionizing how we handle constraint-sensitive queries in information retrieval.
Information retrieval systems have long treated semantic similarity as a proxy for relevance. For constraint-sensitive queries, this proxy can fail when a document is topically close to the query but supports the opposite constraint direction, such as satisfying an attribute that should be excluded or affirming a relation that should be negated. We study this failure as constraint-violating evidence exposure and propose CoDeR, a local constraint-compatible dense retrieval method that separates topical relevance from constraint compatibility. CoDeR keeps a standard topical encoder for candidate coverage and adds a compatibility scorer, implemented as a bi-encoder, trained with lexical-polarity supervision over contrastive satisfying and violating evidences. The compatibility signal can be used to rescore topical candidates or to retrieve an auxiliary compatibility-oriented candidate set, producing a ranked document list without external Large Language Model~(LLM) calls at inference time. We evaluate CoDeR on controlled diagnostics and public negative-constraint retrieval benchmarks. Across three controlled diagnostic sets targeting antonymy, negation, and exclusion, CoDeR reduces V@2 by 20.59, 23.53, and 5.77 points relative to the strongest non-CoDeR baselines, and improves FVR by pushing the first violating document deeper in the ranking.
