Benchmarking Text Embedding Models for Multi-Dataset Semantic Textual Similarity: A Machine Learning-Based Evaluation Framework

The selection of optimal text embedding models remains a critical challenge in semantic textual similarity (STS) tasks, particularly when performance varies substantially across datasets.In this study, the comparative effectiveness of multiple state-of-the-art embedding models was systematically evaluated using a benchmarking framework based on established machine learning techniques.A range of embedding architectures was examined across diverse STS datasets, with similarity computations performed using Euclidean distance, cosine similarity, and Manhattan distance metrics.Performance evaluation was conducted through Pearson and Spearman correlation coefficients to ensure robust and interpretable assessments.The results revealed that GIST-Embedding-v0 consistently achieved the highest average correlation scores across all datasets, indicating strong generalizability.Nevertheless, MUG-B-1.6 demonstrated superior performance on datasets 2, 6, and 7, while UAE-Large-V1 outperformed other models on datasets 3 and 5, thereby underscoring the influence of dataset-specific characteristics on embedding model efficacy.These findings highlight the importance of adopting a dataset-aware approach in embedding model selection for STS tasks, rather than relying on a single universal model.Moreover, the observed performance divergence suggests that embedding architectures may encode semantic relationships differently depending on domain-specific linguistic features.By providing a detailed evaluation of model behavior across varied datasets, this study offers a methodological foundation for embedding selection in downstream NLP applications.The implications of this research extend to the development of more reliable, scalable, and context-sensitive STS systems, where model performance can be optimized based on empirical evidence rather than heuristics.These insights are expected to inform future investigations on embedding adaptation, hybrid model integration, and meta-learning strategies for semantic similarity tasks.