Evaluating the Performance of Transformers-Based Semantic Similarity Measures in Short-Text Clustering

Clustering short texts is an essential task in many natural language processing applications. Traditional methods of measuring text similarity, which rely on vector space models, struggle to capture the nuanced semantic information necessary for this task. However, in recent years, transformer models, specifically BERT, have exhibited impressive performance in capturing contextualized word representations, offering promise in addressing challenges related to short-text clustering. This paper presents a new implementation for evaluating sentence-transformers-based text similarity measures in the task of short-text. The experimental findings demonstrate the superiority of transformer-based measures, especially Siamese and USE similarity measures, in capturing the semantic relationships among a collection of text fragments, resulting in a high-quality clustering.