Interpret and Control Dense Retrieval with Sparse Latent Features

TL;DR

This paper proposes a method using sparse autoencoders to interpret and control dense retrieval embeddings, maintaining performance while enabling meaningful manipulation of retrieval outcomes.

Contribution

It introduces a contrastive loss for sparse autoencoders that preserves retrieval effectiveness and enhances interpretability and controllability of dense embeddings.

Findings

Sparse latent features retain high retrieval accuracy.

Latent space manipulation allows control over retrieval results.

Reconstructed embeddings are faithful to original dense vectors.

Abstract

Dense embeddings deliver strong retrieval performance but often lack interpretability and controllability. This paper introduces a novel approach using sparse autoencoders (SAE) to interpret and control dense embeddings via the learned latent sparse features. Our key contribution is the development of a retrieval-oriented contrastive loss, which ensures the sparse latent features remain effective for retrieval tasks and thus meaningful to interpret. Experimental results demonstrate that both the learned latent sparse features and their reconstructed embeddings retain nearly the same retrieval accuracy as the original dense vectors, affirming their faithfulness. Our further examination of the sparse latent space reveals interesting features underlying the dense embeddings and we can control the retrieval behaviors via manipulating the latent sparse features, for example, prioritizing documents from specific perspectives in the retrieval results.