A Hybrid Model with New Word Weighting for Fast Filtering Spam Short Texts.

Area of Science:

• Natural Language Processing
• Machine Learning
• Information Security

Background:

• Short text messages (SMS, microblogs, apps) are prone to spam due to their low cost and wide reach.
• Classifying short texts is challenging due to their brevity, sparsity, rapid evolution, and informal language.
• Existing methods like Hidden Markov Models (HMM) struggle with new, informally written words.

Purpose of the Study:

• To propose a hybrid model for fast and accurate short text filtering, specifically addressing the challenge of informal writing.
• To improve spam detection in short messages by effectively weighting new words.
• To enhance the performance of short text classification beyond traditional HMM capabilities.

Main Methods:

• A hybrid model integrating an Artificial Neural Network (ANN) for new word weighting and an HMM for spam filtering was developed.
• The ANN calculates new word weights based on neighboring word weights and predicted spam/ham probabilities.
• Performance was evaluated on benchmark datasets including SMS, movie reviews, and customer reviews.

Main Results:

• The hybrid model demonstrated significantly higher operational speed compared to deep learning models.
• Experimental results showed the hybrid model outperformed other prominent machine learning algorithms in short text classification.
• The proposed model achieved a favorable balance between filtering throughput and classification accuracy.

Conclusions:

• The hybrid ANN-HMM model effectively addresses the informal writing challenge in short text classification.
• This approach offers a robust solution for high-accuracy, high-speed spam filtering of short messages.
• The model provides a superior alternative to existing methods for combating spam in the era of diverse digital communication platforms.