Cropping outperforms dropout as an augmentation strategy for training self-supervised text embeddings

Abstract

Text embeddings, i.e. vector representations of entire texts, play an important role in many NLP applications, such as retrieval-augmented generation, sentiment analysis, clustering, or visualizing collections of texts for data exploration. Currently, top-performing embedding models are derived from pre-trained language models via extensive supervised fine-tuning using curated text pairs. This contrasts with computer vision, where self-supervised training based on data augmentations has demonstrated remarkable success. Here we systematically compare the two most well-known augmentation strategies for positive pair generation in contrastive learning of text embeddings. We assess embedding quality on MTEB and additional in-domain evaluations and show that cropping augmentation strongly outperforms the dropout-based approach. We find that on out-of-domain data, the quality of resulting embeddings is below the supervised SOTA models, but for in-domain data, self-supervised fine-tuning produces high-quality text embeddings after very short fine-tuning, sometimes only marginally below the supervised SOTA. Finally, we show that representation quality increases towards the last transformer layers, which undergo the largest change during fine-tuning; and that fine-tuning only those last layers is sufficient to reach similar embedding quality.