Hyperparameter settings for DSN-STC.

In this paper, we present a novel deep Siamese network with a multi-scale hybrid feature extraction architecture, named DSN-STC (Deep Siamese Network for Short Text Clustering), that significantly improves the clustering of short text. A key innovation of our approach is a specialized transformation mechanism that maps pre-trained word embeddings into cluster-aware text representations. In this new latent space, the proposed model minimizes the overall overlapping between clusters while improving the cohesion within each cluster. This results in considerable improvements in clustering performance. Since short texts inherently contain both sequential context and localized patterns within their limited context, in this paper a hybrid approach is used by combining both recurrent layers and multi-scale convolutional neural networks to maximize the extractable feature sets from their limited context. This architecture allows us to capture the sequential features and local dependencies by recurrent layer and convolutional layers respectively which leads to generating a more accurate and rich representation for each short text. To evaluate our architecture and because our main focus is on clustering Persian short text, several experiments are conducted in which the results show that the DSN-STC outperforms other approaches in clustering accuracy (ACC) and normalized mutual information (NMI) metrics. Also to further test the proposed architecture’s generalizability and adaptability in other languages, DSN-STC is evaluated on 2 English benchmark datasets where it consistently outperformed previous approaches in both metrics. These results highlight the model’s ability to learn robust and cluster-aware feature representations that are highly useful for effective short text clustering.