Qwen TextCNN and BERT models for enhanced multilabel news classification in mobile apps

Area of Science:

• Natural Language Processing
• Machine Learning for Mobile Applications

Background:

• Mobile news classification systems are complex and large-scale.
• Evaluating traditional models against Large Language Models (LLMs) is crucial for optimizing mobile news categorization.

Purpose of the Study:

• To conduct a comparative study of traditional classification models and LLMs for multi-label news categorization in Chinese mobile applications.
• To assess the performance of BERT, Qwen (instruction-tuned and LoRA fine-tuned), TextCNN, LSTM, and MLP classifiers.

Main Methods:

• Comparative analysis of TextCNN, BERT, and various LLMs (Qwen).
• Evaluation of models using instruction tuning and Low-Rank Adaptation (LoRA) fine-tuning techniques.
• Assessment of classifier performance on balanced and imbalanced datasets, focusing on multi-label and binary classification tasks.

Main Results:

• BERT models demonstrate superior performance in multi-label classification on balanced datasets.
• TextCNN is more effective for binary classification tasks.
• LSTM and MLP classifiers achieve high accuracy with text instruction prompts, outperforming random embeddings.

Conclusions:

• Model selection for mobile news classification requires balancing technical capabilities with deployment constraints.
• Despite class imbalance challenges (low macro F1 scores), the relative performance analysis validates the findings.
• The study provides insights into optimizing automotive news classification within mobile applications.