acceptodds
Desk rejected at ICLR 2027

Ravel: Separating Romanization Collisions in Multilingual Dense Retrieval

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Romanization erases script cues that separate names, allowing a faithful variant of one entity to share its surface form with another identity. We formulate script robustness as two coupled obligations: reconnect relevant variants and keep identity-colliding passages outside the retrieved neighborhood. The Fracture–Collapse Protocol (FCP) evaluates benchmark-wide noisy Recall@100 together with candidate-micro collision intrusion over a model-independent population, and connects retrieval behavior to neighborhood geometry, source-disjoint noise, real input-method errors, generated answers, and serving cost. Ravel realizes this constraint with BGE-M3 semantics, a 12M-parameter byte/character form channel, and identity-labeled collision negatives, while retaining one vector per query and passage. Across MKQA, mMARCO, and XOR-Retrieve, it gains 6.9–8.3 Recall@100 points over BGE-M3 and 3.5–3.7 points over the same-backbone invariance control, with 0.3–0.4 points of clean regression and 32 ms p95 latency. On MKQA, Ravel reaches 69.7 noisy recall at 1.8% candidate intrusion, compared with 66.2 and 4.1% for invariance; its 6.9-point source-disjoint advantage persists, and fixed-generator wrong-entity answers fall from 4.8% to 3.3%. These results establish identity separation as a co-equal criterion for training and selecting script-robust single-vector retrievers.

Primary area:
foundation or frontier models, including LLMs
Keywords:
retrieval rag and knowledge dynamics, Ravel

Related papers

Discussion (0)

Sign in to comment.