Multi-Stage NLP Classification; Sentiment, Intent, and RLHF Evaluation
Scope & Tasks: This project involved a dual-phase annotation process designed to improve Natural Language Processing (NLP) models. 1. Phase 1 (User Feedback): Performed multi-label classification on raw customer inputs. Tasks included identifying Sentiment (Positive, Negative, Neutral) and categorizing User Intent(Question, Complaint, Feedback, Request, Other). 2. Phase 2 (RLHF Evaluation): Conducted a Reinforcement Learning from Human Feedback (RLHF) audit on AI-generated responses. I evaluated model outputs based on Relevance, Accuracy, and Clarity, providing qualitative critiques to identify technical gaps and jargon-heavy explanations. Project Size: * Initial pilot batch of 20+ high-variance text samples representing diverse user personas (tech-savvy, frustrated, and general inquiry). Quality Measures Adhered to: * Consistency: Applied strict internal rubrics to ensure "Neutral" labels were reserved for factual data, while emotive language was correctly funneled into Positive/Negative categories. * Edge Case Resolution: Developed logic for ambiguous inputs (e.g., distinguishing between a "Complaint" and a "Request for Action"). * Clarity Benchmarking: Evaluated AI responses against a "Beginner Persona" standard to ensure technical concepts (like Blockchain) were accessible without sacrificing fundamental accuracy.