LAMAR: An Open Language-Aware Multilingual Alignment Reranker
The paper’s claim is that multilingual rerankers should rank same-language documents higher when relevance is otherwise equivalent.
The authors say current multilingual rerankers do not reliably account for the query-document language match. LAMAR is trained for both semantic relevance and language coherence, using English-anchored distillation and preference alignment. In their controlled test, it performs best overall and for each examined language, while staying competitive on standard multilingual reranking benchmarks. HF Daily Papers' note
The authors say current multilingual rerankers do not reliably account for the query-document language match. LAMAR is trained for both semantic relevance and language coherence, using English-anchored distillation and preference alignment. In their controlled test, it performs best overall and for each examined language, while staying competitive on standard multilingual reranking benchmarks. HF Daily Papers' note
score 4