Robust hierarchical model for joint span detection and aspect-based sentiment analysis in Vietnamese

Recent successes of large pre-trained language models highlighted deep neural networks' potential in solving complicated natural language processing tasks, but some remain unsolved, especially for non-English languages. Two of them are span detection and aspect-based sentiment analysis. The former is a general sequence tagging, and the latter is a hierarchical classification task. Both are challenging in various ways, most notably due to the lack of available resources and effective methods for training. To overcome these drawbacks, this paper introduces a fashionable and robust hierarchical model that learns both tasks simultaneously using a multi-task learning framework for Vietnamese datasets. Our model is trained using a tunable custom task-dependent loss and easily adapted to a wide range of similar tasks. Experimental results showed that our approach is superior to the existing baseline systems and achieved state-of-the-art results. We also make the source code publicly available as a starter kit for further investigation and research ^{1 1} https://github.com/datnnt1997/ViSA.