Fine-Grained Multi-label Sexism Classification Using a Semi-Supervised Multi-level Neural Approach

Abstract Sexism, a permeate form of oppression, causes profound suffering through various manifestations. Given the increasing number experiences sexism shared online, categorizing these recollections automatically can support battle against sexism, since it promote successful evaluations by gender studies researchers and government representatives engaged in policy making. In this paper, we examine fine-grained, multi-label classification accounts (reports) sexism. To best our knowledge, consider substantially more categories than any related prior work 23-class problem formulation. Moreover, present first semi-supervised for describing type(s) We devise self-training-based techniques tailor-made nature to utilize unlabeled samples augmenting labeled set. identify high textual diversity with respect existing set as desirable quality candidate instances develop methods incorporating into approach. also explore ways infusing class imbalance alleviation learning, independently conjunction method involving diversity. addition data augmentation methods, neural model which combines biLSTM attention domain-adapted BERT an end-to-end trainable manner. Further, formulate multi-level training approach models are sequentially trained using different levels granularity. loss function that exploits label confidence scores associated data. Several proposed outperform baselines on recently released dataset categorization across several standard metrics.