DSN-STC: Leveraging Siamese networks for optimized short text clustering.

Area of Science:

• Natural Language Processing
• Machine Learning
• Artificial Intelligence

Background:

• Short texts present unique challenges for clustering due to limited context.
• Existing methods struggle to effectively capture both sequential and local patterns in short texts.

Purpose of the Study:

• To develop a novel deep Siamese network (DSN-STC) for improved short text clustering.
• To create cluster-aware text representations that enhance intra-cluster cohesion and inter-cluster separation.

Main Methods:

• A multi-scale hybrid feature extraction architecture combining recurrent and convolutional neural networks.
• A specialized transformation mechanism mapping pre-trained word embeddings into a cluster-aware latent space.
• Minimizing cluster overlapping while maximizing intra-cluster cohesion.

Main Results:

• DSN-STC significantly outperforms existing approaches in clustering Persian short text, showing improvements in accuracy (ACC) and normalized mutual information (NMI).
• The model demonstrates generalizability and adaptability by achieving superior performance on English benchmark datasets.
• The proposed architecture effectively learns robust, cluster-aware feature representations.

Conclusions:

• DSN-STC offers a powerful solution for short text clustering, particularly for languages with limited digital resources.
• The hybrid feature extraction and cluster-aware representation learning are key to the model's success.
• The findings underscore the potential of deep learning architectures for nuanced text analysis and clustering.