Distribution-Aligned Fine-Tuning for Efficient Neural Retrieval (2211.04942v1)

Abstract: Dual-encoder-based neural retrieval models achieve appreciable performance and complement traditional lexical retrievers well due to their semantic matching capabilities, which makes them a common choice for hybrid IR systems. However, these models exhibit a performance bottleneck in the online query encoding step, as the corresponding query encoders are usually large and complex Transformer models. In this paper we investigate heterogeneous dual-encoder models, where the two encoders are separate models that do not share parameters or initializations. We empirically show that heterogeneous dual-encoders are susceptible to collapsing representations, causing them to output constant trivial representations when they are fine-tuned using a standard contrastive loss due to a distribution mismatch. We propose DAFT, a simple two-stage fine-tuning approach that aligns the two encoders in order to prevent them from collapsing. We further demonstrate how DAFT can be used to train efficient heterogeneous dual-encoder models using lightweight query encoders.