Automated Social Text Annotation With Joint Multilabel Attention Networks

Automated social text annotation is the task of suggesting a set tags for shared documents on media platforms. The automated process can reduce users' cognitive overhead in tagging and improve tag management better search, browsing, recommendation documents. It be formulated as multilabel classification problem. We propose novel deep learning-based method this problem design an attention-based neural network with semantic-based regularization, which mimic reading behavior to formulate document representation, leveraging semantic relations among labels. separately models title content each injects explicit, title-guided attention mechanism into sentence. To exploit correlation labels, we two loss regularizers, i.e., similarity subsumption, enforce output conform label semantics. model regularizers referred joint (JMAN). conducted comprehensive evaluation study compared JMAN state-of-the-art baseline models, using four large, real-world data sets. In terms F 1 , significantly outperformed bidirectional gated recurrent unit (Bi-GRU) relatively by around 12.8%-78.6% hierarchical (HAN) 3.9%-23.8%. demonstrates advantages convergence training speed. Further improvement performance was observed against latent Dirichlet allocation (LDA) support vector machine (SVM). When applying HAN Bi-GRU also boosted. found that dynamic update matrices (JMAN xmlns:xlink="http://www.w3.org/1999/xlink">d ) has potential further but at cost substantial memory warrants study.