SPECTER: Document-level Representation Learning using Citation-informed Transformers

Representation learning is a critical ingredient for natural language systems. Recent Transformer language models like BERT learn powerful representations, but these models are targeted towards token- and-level training objectives and do not leverage information on-document relatedness, which limits their document-level representation. For applications on scientific documents, such as classification and, the embeddings power strong performance on end tasks. We SPECTER, a new method to generate document-level embedding of documents based on pretraining a Transformer language model on a signal of document-level relatedness: the citation graph. Unlike pretrained language models, SPECTER can be easily applied to applications without task-specific fine-tuning. Additionally, to further research on document-level models, we introduce SciDocs, a evaluation benchmark consisting of seven document-level tasks ranging from prediction, to document classification and recommendation. We show SPECTER outperforms a variety of competitive baselines on the benchmark.