A graph-based model for semantic textual similarity measurement

Abstract

Measuring semantic similarity between sentence pairs is a fundamental problem in Natural Language Processing with applications in various domains, including machine translation, speech recognition, automatic question answering, and text summarization. Despite its significance, accurately assessing semantic similarity remains a challenging task, particularly for underrepresented languages such as Vietnamese. Existing methods have yet to fully leverage the unique linguistic characteristics of Vietnamese for semantic similarity measurement. To address this limitation, we propose GBNet-STS (Graph-Based Network for Semantic Textual Similarity), a novel framework for measuring the semantic similarity of Vietnamese sentence pairs. GBNet-STS integrates lexical-grammatical similarity scores and distributional semantic similarity scores within a multi-layered graph-based model. By capturing different semantic perspectives through multiple interconnected layers, our approach provides a more comprehensive and robust similarity estimation. Experimental results demonstrate that GBNet-STS outperforms traditional methods, achieving state-of-the-art performance in Vietnamese semantic similarity tasks.