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

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

PDF Abstract ACL 2020 PDF ACL 2020 Abstract

Datasets


Introduced in the Paper:

SciDocs
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Representation Learning SciDocs SciBERT Avg. 59.6 # 5
Representation Learning SciDocs Citeomatic Avg. 76.0 # 3
Representation Learning SciDocs SPECTER Avg. 80.0 # 2
Citation Prediction SciDocs (Citation Prediction) SPECTER MAP 88.3 # 2
Document Classification SciDocs (MAG) SPECTER F1 (micro) 82.0 # 1
Document Classification SciDocs (MeSH) SPECTER F1 (micro) 86.4 # 2

Methods