Dual-Objective Contrastive Learning for Novel Class Discovery in Document-Level Relationship Extraction

Document-level relationship extraction (DocRE) expands upon traditional relationship extraction (RE) by allowing for multiple entity mentions and multiple relationships between entity pairs across a document. While most DocRE methods focus on classifying a fixed dataset of predefined relationships, few have investigated novel class discovery applied to this domain where previously unknown and unlabeled relationships are extracted. To address this, we propose a multi-step filter and train approach utilizing a mix of supervised cross-entropy and contrastive representation learning to reduce the prevalence of “true negatives” and discover novel relationships at the document level. Additionally, we incorporate an alternative, entity-focused pre-trained encoder, LUKE, into an existing DocRE model, slightly improving the Fl score on the DocRED [1] dataset from 0.6132 to 0.6189. Our filtering methods ensure clustering is applied to only the most likely candidates for novel classes and increase the representation of novel class instances by an average of 9 times compared to naive random sampling. At test time, we retrieve both novel and known classes through simulated human-labeled cluster propositions, achieving a macro F1 score of 0.30 for novel classes. Furthermore, our approach does not negatively impact the classification performance on known classes.