Multi-Label Classification of Indonesian Financial Risk News using Transformer-based Multi-Task Learning
Miftahul Mahfuzh, Ayu Purwarianti · 2022
Multi-label classification is one of the task groups in text classification, which can predict more than one class in the text. A commonly used approach for multi-label task is binary relevance technique which usually needs large amounts of resources and suffers from loss of knowledge on characteristics of other labels in data during training. In this study, we propose to implement multi-task learning during Transformers finetune on binary relevance data, which aims to reduce resource consumption by utilizing parameter sharing and increase performance by obtaining other label characteristics during training. Our experiments had shown that our method outperformed vanilla multi-label finetune. Using 1599 instances of training data and 400 instances of testing data with 16 classes, our method achieved.8817 for classification and achieved.9083 in F1 for classification of 3 sampling classes.