In simple words: Define the label clearly before training a sentiment model.
Choose the question
Sentiment analysis asks whether text expresses a positive, negative, or neutral opinion. First decide what is being judged: the product, delivery, the whole review, or something else. Mixed opinions may need more than one label.
A lab workflow
Collect labelled reviews, split data, clean and tokenize text, convert it to numbers with counts, TF-IDF, or embeddings, train a classifier, evaluate it, and test a new review. Inspect mistaken predictions instead of trusting one accuracy number.
See the flow
reviews→labels→text numbers→model→evaluation
Simple example
“The phone is excellent, but delivery was late” can be positive for product quality and negative for delivery.
Remember: Define the label clearly before training a sentiment model.