Text Classification

• Build an end-to-end text classification pipeline
• Compare TF-IDF-linear vs embedding vs transformer approaches
• Evaluate multi-class text fairly (macro metrics, per-class confusion)

• It automates ticketing, moderation, triage and discovery — high-business-value, low-ceremony ML.
• A correct classification pipeline transfers to spam, intent, language and sentiment domains.
• Comparing approaches honestly (baseline-first) stops teams from overspending on models.
• Per-class errors are operational gold: misrouted support tickets cost dollars, tracked as confusion.

The clerk uses cues — dept, header, sender, keywords — to bucket each envelope. A classifier learns which cues matter and (crucially) which sorts it gets wrong, so the mailroom can fix the miss-routing where money leaks.

• Build the label-quality check first: duplicate texts, mislabeled classes, tiny minority classes = the real problem.
• Vectorize: TF-IDF baseline → optional embeddings → transformer embeddings (frozen or fine-tuned).
• Fit on stratified train, tune with stratified CV; optimize macro-F1 (each class equal) unless frequency-weighted matters.
• Evaluate the confusion matrix per class and by slice (short/long text, language, channel).
• Ship the simplest model that clears the bar; revisit as data grows.

• A support platform routes 80k tickets/day with a TF-IDF logistic model — explainable, <50ms, retrained nightly.
• A moderation system fine-tunes a small transformer for 'offensive' detection with macro-recall gates.
• An email triage product classifies intent (inquiry/complaint/order) and feeds downstream automation with its confidence.

• Always start at the linear TF-IDF baseline — it sets the bar transformers must genuinely beat.
• Class-stratified splits + macro-averaged scores unless the cost truly is frequency-weighted.
• Log prediction confidence; a threshold on confidence is often better than a fixed label for routing.
• Recheck label quality before blaming the model — 5% label noise caps any accuracy.

• Judging on accuracy when classes are imbalanced (a routing model with 90% 'general' class).
• Using an unbalanced test and trusting '91%' reported on the majority class.
• Fine-tuning the transformer on 200 labeled examples and calling it superior to TF-IDF.
• Deploying a model whose confidence says 'unsure' but rendering it as a hard label anyway.

• Print top-coefficient words per class for the linear model — a suspicious word means a labeled-noise leak.
• Slice confusion by text length; short texts may under-represent signal entirely.
• If a rare class is never predicted, its training share is too small — hard-sample or rebalance.
• Run everything on a 300-sample subset first; iterate on speed before burning the full corpus.