From Tokens to Concepts: Leveraging SAE for SPLADE

TL;DR

This paper introduces SAE-SPLADE, a semantic concept-based IR model that maintains SPLADE's effectiveness while enhancing efficiency and addressing vocabulary limitations.

Contribution

It proposes replacing the backbone vocabulary with a learned semantic concept space using Sparse Auto-Encoders, improving multilingual and multi-modal IR performance.

Findings

SAE-SPLADE achieves comparable retrieval performance to SPLADE.

SAE-SPLADE offers improved efficiency over traditional SPLADE.

The model effectively addresses vocabulary limitations in IR.

Abstract

Learned Sparse IR models, such as SPLADE, offer an excellent efficiency-effectiveness tradeoff. However, they rely on the underlying backbone vocabulary, which might hinder performance (polysemicity and synonymy) and pose a challenge for multi-lingual and multi-modal usages. To solve this limitation, we propose to replace the backbone vocabulary with a latent space of semantic concepts learned using Sparse Auto-Encoders (SAE). Throughout this paper, we study the compatibility of these 2 concepts, explore training approaches, and analyze the differences between our SAE-SPLADE model and traditional SPLADE models. Our experiments demonstrate that SAE-SPLADE achieves retrieval performance comparable to SPLADE on both in-domain and out-of-domain tasks while offering improved efficiency.