Graph-Based Clustering of Short Texts Using Word Embedding Similarity
Supakpong Jinarat, Ratchakoon Pruengkarn · International Journal of Advanced Computer Science and Applications · 2025
The exponential growth of short textual content on the Internet, such as social media posts and search snippets, necessitates effective text mining techniques. Short text clustering, a critical tool for organizing this data, contends with two primary challenges: data sparsity, which undermines the quality of traditional clustering methods, and the poor interpretability of machine-generated cluster labels. This study introduces the Semantic Word Graph (SWG) algorithm, a novel graph-based approach designed to address both of these issues simultaneously. Our methodology begins by constructing a global word graph where nodes represent unique terms from the corpus, and edges are weighted by the semantic similarity of word pairs, calculated using a pre-trained Word2Vec model. Cohesive communities of words are then identified using the Louvain method, and documents are assigned to clusters based on these communities. Meaningful cluster labels are generated by ranking representative nouns within each community. To validate our approach, the SWG algorithm was evaluated on three benchmark datasets (AG News, Tweet, and SearchSnippets) and compared against established methods, including Lingo, Suffix Tree Clustering (STC), and K-means. Quantitative results, measured by the F-score, show that SWG achieved up to 0.89 F-score on AG News, 0.85 on Tweets, and 0.82 on SearchSnippets, consistently outperforming baseline algorithms in clustering quality. Further-more, a qualitative analysis confirms that SWG produces more coherent and topically comprehensive cluster labels, improving interpretability. This study concludes that the SWG algorithm is a robust and effective framework for enhancing both the accuracy and interpretability of short text clustering. Future research could explore integrating contextual embeddings such as BERT to capture deeper semantic relationships, optimizing the similarity threshold dynamically for different datasets, and scaling the algorithm to handle larger, real-time streaming text data. These directions would further improve the applicability of SWG in diverse domains such as social media analytics, news aggregation, and real-time topic detection.