Less is More: Pre-train a Strong Text Encoder for Dense Retrieval Using a Weak Decoder
This addresses a bottleneck in dense retrieval for search and recommendation systems, offering a novel pre-training approach with demonstrated gains.
The paper tackles the problem that autoencoder-based language models for dense retrieval may produce poor text embeddings due to decoder shortcuts, and proposes pre-training with a weak decoder, resulting in significant effectiveness and few-shot ability improvements across web search, news recommendation, and open domain question answering tasks.
Dense retrieval requires high-quality text sequence embeddings to support effective search in the representation space. Autoencoder-based language models are appealing in dense retrieval as they train the encoder to output high-quality embedding that can reconstruct the input texts. However, in this paper, we provide theoretical analyses and show empirically that an autoencoder language model with a low reconstruction loss may not provide good sequence representations because the decoder may take shortcuts by exploiting language patterns. To address this, we propose a new self-learning method that pre-trains the autoencoder using a \textit{weak} decoder, with restricted capacity and attention flexibility to push the encoder to provide better text representations. Our experiments on web search, news recommendation, and open domain question answering show that our pre-trained model significantly boosts the effectiveness and few-shot ability of dense retrieval models. Our code is available at https://github.com/microsoft/SEED-Encoder/.