Build and iteratively improve sentiment classifier

You are asked to build a sentiment classifier (for example, positive/neutral/negative) for user-generated text. You have already shipped a first version, and the interviewer wants to go deep on the project. Explain:

• How you set up the problem (label definitions, classes, prediction unit, handling of multiple languages and emoji).
• Why you selected your modeling method (baseline versus deep model) and which alternatives you weighed.
• Your data pipeline and labeling approach (human annotation, weak supervision, distant labels, class imbalance).
• How you measured the model (metrics, train/validation split, leakage risks) and what error analysis you performed.
• How you improved the system over time based on what you found (data cleaning, feature/model changes, thresholding, calibration).
• What you learned while iterating and what you would do next. Overview: This question tests applied machine learning and natural language processing ability, covering problem setup (labels, classes, prediction unit, multilingual and emoji handling), modeling trade-offs, data pipeline and labeling choices, evaluation and error analysis, and iterative refinement.