How do you implement text classification for production NLP systems?

Text classification assigns labels to text documents. Implementation: (1) Traditional ML pipeline: TF-IDF vectorization + logistic regression or SVM, fast inference, good for small datasets, and highly interpretable. (2) Deep learning: fine-tune a pre-trained transformer (DistilBERT for speed, RoBERTa for accuracy) on labeled data, achieving state-of-the-art accuracy. (3) Zero/few-shot: use an LLM with prompt engineering for classification without training data, useful for rapid prototyping or infrequent categories. Training data preparation: collect representative examples for each class, handle class imbalance (oversampling minority classes, class-weighted loss, or synthetic data generation), clean text (remove HTML, normalize encoding, handle emojis), and split into train/validation/test with stratification. Fine-tuning: use Hugging Face Trainer with AutoModelForSequenceClassification, set learning rate 2e-5 to 5e-5, batch size 16-32, train for 3-5 epochs with early stopping on validation loss, and use gradient accumulation for larger effective batch sizes on limited GPU memory. Multi-label classification: use sigmoid activation instead of softmax, BCEWithLogitsLoss, and threshold tuning per label. Hierarchical classification: implement parent-child label relationships, train separate models per level or a single model with hierarchical loss. Production: serve via REST API with batch inference support, implement input preprocessing pipeline (tokenization, length truncation), cache predictions for repeated inputs, and implement A/B testing for model updates. Monitoring: track prediction distribution drift (alert if class frequencies change significantly), measure latency percentiles, and implement feedback loop for continuous improvement. Evaluation: precision, recall, F1 per class and macro-averaged, confusion matrix analysis, and error analysis on misclassified examples.