I worked on a project on Sentiment Analysis for Customer Reviews
I worked on a project on Sentiment Analysis for Customer Reviews. I Labeled 10,000 product reviews as Positive, Negative, or Neutral. It was on Amazon product reviews (text data). Annotators received guidelines on sentiment labeling. For example “I love this product" which is Positive Annotators labeled a test batch, discussed edge cases (sarcasm), and aligned on rules. Reviews were split among annotators; each review got 3 labels to ensure agreement. Disagreements (1 Positive, 2 Neutral) were reviewed by a senior annotator. I faced some Challenges like Customers that make use of Sarcasm/subtlety like “Great, just what I needed" (could be Positive or Negative). Mixed sentiment like “The camera is great but battery life sucks" (Neutral). Domain-specific terms "This mic is fire" (Positive slang). High agreement post-discussion; model trained on this data achieved 85% accuracy in predicting sentiment.