Exploring Hierarchical Multi-Label Text Classification Models using Attention-Based Approaches for Vietnamese language

The Hierarchical Attention-based Recurrent Neural Network (HARNN) is a system designed to categorize documents efficiently, taking into account both the content of the texts and their hierarchical category structure. This system is comprised of three primary components: the Document Representation Layer (DRL), which is used for semantic encoding, the Hierarchical Attention-based Recurrent Layer (HARL), that models dependencies between different hierarchical levels, and the Hybrid Predicting Layer (HPL), which is responsible for accurate category predictions. In this research, we put HARNN to the test, using a dataset of Vietnamese articles from VnExpress. We then contrast the performance of four different word embeddings (Word2Vec, FastText, PhoBERT, and BERT multilingual). Additionally, we introduce a domain-based approach for the HARNN model to compare the accuracy with the original manner. Experimental findings indicate that HARNN performs effectively in the context of Vietnamese language and that our domain-based approach can be advantageous in specific domains HMTC task.