Toxic-Content-Classification-System: A Multi-Stage NLP and Multimodal AI Content Moderation System

A multi-stage NLP and multimodal AI system integrating LSTM, BLIP, LoRA, and Llama Guard for content understanding, moderation, and generation, combining classical deep learning, Transformer architecture, and modern safety-oriented large language models.

In today's internet environment, content moderation has become one of the core challenges for platform operations. With the explosive growth of user-generated content, traditional manual moderation methods can no longer meet the needs of real-time and large-scale operations. Toxic-Content-Classification-System is a multi-stage natural language processing (NLP) and multimodal AI system specifically designed to address content understanding, moderation, and generation challenges in the real world. This system integrates classical deep learning, Transformer-based architectures, and modern safety-oriented large language models into a unified production-level pipeline.

The system consists of four main components, each targeting specific content moderation needs:

The text classification module uses a Long Short-Term Memory (LSTM) network architecture, a classic deep learning model for processing sequence data. The system implements the following technical workflow:

• Text Preprocessing: Includes text cleaning, tokenization, and padding to ensure consistency and quality of input data
• Word Embedding Layer: Converts text into dense vector representations to capture semantic relationships
• LSTM Sequence Modeling: Uses the memory mechanism of LSTM to handle long text dependencies
• Bidirectional LSTM and Dropout: Optional bidirectional architecture enhances context understanding, and Dropout regularization prevents overfitting

The entire inference process starts with input text; after preprocessing and encoding, the model outputs classification prediction results. The system uses metrics such as accuracy, precision, recall, F1 score, and confusion matrix for comprehensive evaluation.

To meet the moderation needs of image content, the system integrates the BLIP (Bootstrapping Language-Image Pre-training) model to enable image understanding and caption generation:

• Accepts image input and generates natural language captions
• Converts image content into auditable text form
• Results are stored in MongoDB Atlas database for easy subsequent analysis and traceability

This multimodal capability allows the system to handle complex scenarios involving images, expanding the scope of application of traditional text-only moderation systems.

To adapt to domain-specific moderation needs, the system implements Parameter-Efficient Fine-Tuning (PEFT) technology:

• LoRA (Low-Rank Adaptation): By injecting low-rank matrices into attention layers, it significantly reduces training costs and memory usage
• DistilBERT Tokenizer: Uses a lightweight Transformer model for text encoding
• Complete Training and Validation Pipeline: Supports model fine-tuning on custom datasets

The advantage of this method is that it only requires training a small number of parameters to achieve model adaptation while maintaining the general capabilities of the base model. The fine-tuned model can perform more accurate identification of specific types of harmful content.

The system's most notable feature is the integration of Llama Guard for zero-shot content moderation:

• No Fine-Tuning Required: Uses the generalization ability of large language models to perform classification directly
• Prompt Engineering: Guides model output through carefully designed prompt templates
• Multi-Dimensional Detection: Identifies toxic content, harmful language, and policy violations

This zero-shot method greatly reduces the cost of deploying new moderation categories, enabling the system to quickly adapt to emerging types of violating content.

The project uses Python as the main development language and relies on the following core libraries:

• Scikit-learn: Traditional machine learning algorithms and evaluation metrics
• Pandas & NumPy: Data processing and numerical computation
• PyTorch: Deep learning model training and inference
• NLTK: Basic natural language processing tools
• Streamlit: Interactive web application deployment
• MongoDB Atlas: Persistent data storage
• Weights & Biases (W&B): Experiment tracking and model version management