ZBJ Demand Order Text Classification

Project Background:

• Designed an automated text classification system for ZBJ.com demand orders.
• Objective: Accurately identify user intent from free-text orders to match relevant service providers.

Data Collection & Processing:

• Collected 1M+ demand orders covering 15 primary categories and 100+ tertiary categories.
• Cleaned data: Removed meaningless symbols, repeated sentences, and irrelevant text.
• Text segmentation: Custom dictionary segmentation using Ansj.
• Vectorization: Word2Vec embeddings trained on domain corpus; TF-IDF and BOW for baseline models.
• Labeling: Combined manual annotation with automated heuristics.

Modeling & Optimization:

• Algorithms: Naive Bayes, CNN, RNN.
• CNN Model:
  • Input: Fixed-length sequences with zero-padding for short orders.
  • Convolutional layers with multiple kernel sizes (2-8), max-pooling, batch normalization.
  • Fully connected layers with dropout and softmax for classification.
• RNN Model:
  • Input sequences padded/truncated.
  • LSTM layers with dropout, gradient clipping.
• Hyperparameter tuning: Learning rate, batch size, hidden layer neurons, activation functions, iteration rounds.
• Ensemble strategy: Voting across Naive Bayes, CNN, RNN to improve robustness.

Workflow:

1. Preprocess text and remove noise.
2. Apply NER and keyword extraction for feature enrichment.
3. Train models and perform cross-validation.
4. Ensemble predictions for final category output.

Project Outcome:

• Naive Bayes accuracy: 90%, CNN accuracy: 85%, RNN accuracy: 87%.
• Improved service provider matching, optimized search engine results, and increased user satisfaction.