Condenser: a Pre-training Architecture for Dense Retrieval
Luyu Gao, Jamie Callan
TL;DR
This paper identifies why standard pre-trained LMs struggle as dense bi-encoders and proposes Condenser, a Transformer-based pre-training architecture that actively conditions LM predictions on dense representations to establish structural readiness for dense retrieval. By separating early and late backbone processing and introducing a Condenser head, the method guides the model to produce informative dense representations that can be fine-tuned as a standard encoder. Across sentence similarity, open-domain QA retrieval, and web-search retrieval, Condenser yields strong gains, particularly in low-data settings, and approaches or surpasses more complex pipelines in many full-data scenarios. Attention analyses corroborate that Condenser maintains a more task-friendly internal structure, suggesting structural readiness is a key factor in efficient dense retrieval pre-training and deployment.
Abstract
Pre-trained Transformer language models (LM) have become go-to text representation encoders. Prior research fine-tunes deep LMs to encode text sequences such as sentences and passages into single dense vector representations for efficient text comparison and retrieval. However, dense encoders require a lot of data and sophisticated techniques to effectively train and suffer in low data situations. This paper finds a key reason is that standard LMs' internal attention structure is not ready-to-use for dense encoders, which needs to aggregate text information into the dense representation. We propose to pre-train towards dense encoder with a novel Transformer architecture, Condenser, where LM prediction CONditions on DENSE Representation. Our experiments show Condenser improves over standard LM by large margins on various text retrieval and similarity tasks.