Disentangled Contrastive Learning Enhances Zero-Shot Multilingual Dense Retrieval
_reachsumit · x · 2026-08-04
The Challenge: Multilingual dense retrieval struggles with robust transfer to low-resource languages lacking annotated data. Existing methods often entangle semantic and linguistic features, interfering with semantic relevance optimization.
The Solution: The paper proposes a Disentangled Contrastive Learning (DCL) method that separates multilingual sentence representations into semantic and linguistic subspaces.
Technical Details:
- Designs disentangled optimization objectives based on hierarchical semantic alignment and language debiasing contrastive learning.
- Aligns retrieval-relevant semantics across languages at both sentence and token levels while capturing language-specific variations in the linguistic subspace.
Results: By jointly optimizing with the retrieval objective, DCL reduces language-induced interference in semantic matching, facilitating stable zero-shot transfer from English supervision to multiple languages.
More from Research
- Treating Models as Learned Optimizers: Multi-start Strategies in Training — YouJiacheng · 2026-08-04
- Atomic-scale logic circuits built from silicon dangling bonds showcased in new paper — teortaxesTex · 2026-08-04
- EviSD: Evidence-Conditioned Self-Distillation for Search Agents — _reachsumit · 2026-08-04
- Samsung Proposes PROGRESS: Coverage-Guided RL to Train Search-Augmented LLM Agents — _reachsumit · 2026-08-04
- Study Reveals Agentic RAG Flaw: Agents Often Skip Reading Evidence Before Answering — _reachsumit · 2026-08-04
- A Brief Look at Kimi K3's MoE and Attention Architecture — hsu_byron · 2026-08-04