Retrieval Disagreement as a Signal: Diagnosing and Steering Multilingual Dense Embedders
What it is
This paper analyzes why multilingual dense retrievers produce inconsistent results when the same query is translated into different languages, identifying that correct answers are buried by same-language or English documents in the shared index. It proposes a 'disagreement-guided steering' method that uses the difference between consensus and idiosyncratic retrieval results to create language-specific steering vectors added to queries at inference time.
The problem it solves
Multilingual dense retrievers exhibit 'retrieval disagreement,' where the same semantic query retrieves different documents depending on the language used, often because correct answers are buried under non-relevant documents in the shared index.
What you could use it for
- Improving cross-lingual retrieval consistency in multilingual search engines
- Reducing language-specific bias in dense embedding models
- Steering vector representations to mitigate 'crowding' effects in shared embedding spaces
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.