Enhanced Short-Text Clustering Based on a Multirepresentation Fusion Model

Abstract

Short-text clustering has become increasingly important because of the exponential growth of textual data. However, traditional methods struggle with sparsity, limited context, and poor cluster cohesion. Recently, innovations have made word embedding algorithms, including Word2Vec, GloVe, and FastText, available for effectively learn semantic relationships. Although these methods have advantages, they tend to concentrate on semantics area. These limitations are addressed in this study by adopting an enhanced short-text clustering method based on multirepresentation. Specifically, this study integrates heterogeneous embeddings into a unified framework using weighted L2-normalized concatenation, followed by Principal Component Analysis (PCA)-based dimensionality reduction that retains 95% of the variance, thereby create a unified semantic representation. Unlike simple concatenation approaches of embeddings, pre-space normalization operations are added before the fusion operation to avoid the presence of dominant embedding bias. The proposed model was evaluated on four benchmark datasets, namely, StackOverflow, AgNews, BioASQ, and SearchSnippet, via the k-means and k-medoids clustering algorithms. The results demonstrate substantial improvements in the clustering quality metrics. Compared with the results of the baseline models Bag of Words (BoW)/Term Frequency-Inverse Document Frequency (TF-IDF), the average gains were statistically significant: 29% in silhouette score, 8% in adjusted Rand index, and 8% in normalized mutual information over the best (single) embedding baseline (GloVe with k-medoids) and 63%, 35%, and 40%, respectively, over traditional BoW/TF-IDF methods. Most improvements are found to be statistically significant at p < 0.05. Additionally, PCA was applied to reduce noise and dimensionality, potentially improving clustering outcomes.



Received: 5 February 2026 | Revised: 12 June 2026 | Accepted: 14 August 2026



Conflicts of Interest

The authors declare that they have no conflicts of interest to this work.



Data Availability Statement

Data are available from the corresponding author upon reasonable request.



Author Contribution Statement

Ali Sabah: Conceptualization, Methodology, Software, Formal analysis, Investigation, Resources, Writing – original draft, Writing – review & editing, Visualization. Dhurgham Ali Mohammed: Software, Validation, Investigation, Resources, Data curation, Writing – review & editing, Visualization. Zaid Alaa: Methodology, Validation, Formal analysis, Investigation, Resources, Data curation, Writing – review & editing, Visualization.