Sort Text Data Risks

Place each issue into the risk category it most directly represents. Leakage Drift Imbalance Ambiguity Privacy Replies from the same email thread appear in both train and test. Training examples are mostly from a pandemic-era policy that has since changed. Only 45 examples exist for legal_complaint, the highest-stakes label. Annotators disagree whether stop billing is billing_issue or cancellation_request. Raw chats include phone numbers and home addresses. The TF-IDF vectorizer was fit before the validation split. New slang appears in production complaints after a product launch. Spanish tickets represent 2 percent of training but 18 percent of live volume. Trap Using…

Sign up free — one personalized lesson every day, matched to your role and goals.

Already have an account? Sign in