Label-Centric Curriculum Contrastive Learning for Zero-shot Extreme Multi-label Biomedical Document Classification

Anonymous

Abstract: Extreme multi-label text classification (XMC) aims to assign relevant labels to a document from a large set of candidate labels. Prior XMC research has typically concentrated on supervised learning methods. However, real-world scenarios frequently present situations where complete supervision signals, in the form of labeled and balanced datasets, are not available, highlighting the importance and relevance of zero-shot learning settings in XMC. In this paper, we study the XMC task on biomedical documents under the zero-shot setting which does not require any annotated documents in the training phase. We propose a novel label-centric curriculum contrastive learning framework for the training phase, which effectively utilizes hierarchical label information and label-metadata co-occurrence. For the inference phase, we employ a multi-stage retrieve and re-rank framework to make more accurate predictions by ruling out the irrelevant labels before ranking, rather than making direct predictions on the entire large label set. Experimental results demonstrate the effectiveness of our approach in improving the performance of XMC.

Paper Type: long

Research Area: NLP Applications

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English