Neighborhood Contrastive Learning for Scientific Document Representations with Citation Embeddings
This addresses the core problem of finding related research papers in scientific literature, particularly for papers lacking direct citations, though it is an incremental improvement over existing contrastive learning approaches.
The paper tackles the problem of learning scientific document representations by proposing a contrastive learning method that uses nearest neighbor sampling over citation graph embeddings instead of discrete citation relations, which allows for continuous similarity learning and better handling of papers without direct citations. The resulting SciNCL method outperforms state-of-the-art on the SciDocs benchmark and shows sample-efficient training capabilities.
Learning scientific document representations can be substantially improved through contrastive learning objectives, where the challenge lies in creating positive and negative training samples that encode the desired similarity semantics. Prior work relies on discrete citation relations to generate contrast samples. However, discrete citations enforce a hard cut-off to similarity. This is counter-intuitive to similarity-based learning, and ignores that scientific papers can be very similar despite lacking a direct citation - a core problem of finding related research. Instead, we use controlled nearest neighbor sampling over citation graph embeddings for contrastive learning. This control allows us to learn continuous similarity, to sample hard-to-learn negatives and positives, and also to avoid collisions between negative and positive samples by controlling the sampling margin between them. The resulting method SciNCL outperforms the state-of-the-art on the SciDocs benchmark. Furthermore, we demonstrate that it can train (or tune) models sample-efficiently, and that it can be combined with recent training-efficient methods. Perhaps surprisingly, even training a general-domain language model this way outperforms baselines pretrained in-domain.