Assemble the shallow or integrate a deep? Toward a lightweight solution for glyph-aware Chinese text classification.

PLoS One

Center for the Studies of Information Resources, Wuhan University, Wuhan, Hubei, China.

Published: July 2023

As hieroglyphic languages, such as Chinese, differ from alphabetic languages, researchers have always been interested in using internal glyph features to enhance semantic representation. However, the models used in such studies are becoming increasingly computationally expensive, even for simple tasks like text classification. In this paper, we aim to balance model performance and computation cost in glyph-aware Chinese text classification tasks. To address this issue, we propose a lightweight ensemble learning method for glyph-aware Chinese text classification (LEGACT) that consists of typical shallow networks as base learners and machine learning classifiers as meta-learners. Through model design and a series of experiments, we demonstrate that an ensemble approach integrating shallow neural networks can achieve comparable results even when compared to large-scale transformer models. The contribution of this paper includes a lightweight yet powerful solution for glyph-aware Chinese text classification and empirical evidence of the significance of glyph features for hieroglyphic language representation. Moreover, this paper emphasizes the importance of assembling shallow neural networks with proper ensemble strategies to reduce computational workload in predictive tasks.

Download full-text PDF

Source
http://www.ncbi.nlm.nih.gov/pmc/articles/PMC10381045PMC
http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0289204PLOS

Publication Analysis

Top Keywords

text classification
20
glyph-aware chinese
16
chinese text
16
solution glyph-aware
8
glyph features
8
shallow neural
8
neural networks
8
chinese
5
text
5
classification
5

Similar Publications

Want AI Summaries of new PubMed Abstracts delivered to your In-box?

Enter search terms and have AI summaries delivered each week - change queries or unsubscribe any time!