我们为什么仍然要扁平化嵌入空间?
大多数密集检索系统依赖于余弦相似度或点积,这隐含地假设了一个平坦的嵌入空间。然而,嵌入空间往往存在于具有非均匀结构的曲面流形上——密集区域、语义间隙、不对称路径。
我一直在探索以下内容:
- 将Ricci曲率作为重排序信号
- 使用软图来保持局部密度
- 在训练过程中使用测地线感知损失
我很好奇是否有其他人尝试过类似的做法?特别是在信息检索、问答或可解释性方面。如果有兴趣,我很乐意分享一些实验(FiQA/BEIR)的结果。
查看原文
ost dense retrieval systems rely on cosine similarity or dot-product, which implicitly assumes a flat embedding space. But embedding spaces often live on curved manifolds with non-uniform structure—dense regions, semantic gaps, asymmetric paths.<p>I’ve been exploring the use of:<p>- Ricci curvature as a reranking signal<p>- Soft-graphs to preserve local density<p>- Geodesic-aware losses during training<p>Curious if others have tried anything similar? Especially in information retrieval, QA, or explainability. Happy to share some experiments (FiQA/BEIR) if there's interest.