Stylistic Pattern Guided Tip Extraction from Music Reviews
Jingya Zang, Anguo Dong · 2022
Reviews are important sources for users to obtain helpful information and make quicker decisions. However, reviews of songs are characterized by a vast number and noise. It is hard for users to grasp meaningful information from a large number of reviews and make decisions. To solve this problem, prior studies propose to provide users with tips. Tips are sentences that are characterized as short, concise, empathetic, and self-contained. Previous studies find that stylistic patterns play an important role in tip identification. However, no prior studies learn the stylistic pattern well. In this paper, we create a heterogeneous graph and propose a dynamic graph convolution net-based model to extract tips by considering stylistic patterns. Experimental results show that our model achieves 74.73% and 73.05% for accuracy and macro F1, outperforming the baseline models by 3.46% and 2.38%.