A novel bundling learning paradigm for named entity recognition

Abstract

Multi-task learning (MTL) takes advantage of the information gained from multiple related NLP tasks in order to improve performance across these tasks. MTL-based models for named entity recognition (NER) have traditionally included relation extraction and (or) coreference resolution, which requires additional data annotations in NER corpora, whereas these annotations are often unavailable. Indeed, we generally model the NER task using either a sequence labeling-based or span-based approach. Motivated by MTL, we propose a novel Bundling Learning (BL) paradigm for the NER task, which is achieved by bundling sequence labeling-based and span-based NER models together, thus allowing us to model the task from both token- and span-level perspectives. In addition, BL does not require additional data annotations compared to MTL. In experiments on NER and RE tasks, it is shown that BL consistently improves the performance of the two tasks across several benchmark datasets. Detailed analyses further confirm the effectiveness of BL.