DocSplit: Simple Contrastive Pretraining for Large Document Embeddings

摘要

Existing model pretraining methods only consider local information. For example, in the popular token masking strategy, the words closer to the masked token are more important for prediction than words far away. This results in pretrained models that generate high-quality sentence embeddings, but low-quality embeddings for large documents. We propose a new pretraining method called DocSplit which forces models to consider the entire global context of a large document. Our method uses a contrastive loss where the positive examples are randomly sampled sections of the input document, and negative examples are randomly sampled sections of unrelated documents. Like previous pretraining methods, DocSplit is fully unsupervised, easy to implement, and can be used to pretrain any model architecture. Our experiments show that DocSplit outperforms other pretraining methods for document classification, few shot learning, and information retrieval tasks.