Sentiment Analysis Tutorial: From Raw Text to Business Insights

01Beyond positive/negative/neutral

Three-class sentiment is a toy. The real value is aspect-based sentiment: which thing is the customer talking about, and how do they feel about it specifically?

02Step 1 — Baseline in 50 lines

from transformers import pipeline
nlp = pipeline("sentiment-analysis", model="cardiffnlp/twitter-roberta-base-sentiment-latest")

texts = ["Love the new feature!", "Crashed three times today.", "It's fine."]
for t in texts:
    r = nlp(t)[0]
    print(t, r["label"], round(r["score"],3))

03Step 2 — Aspect extraction

For aspect-based sentiment you need two outputs per review: the aspects mentioned (food, service, price) and the sentiment toward each.

prompt = """Extract aspects and sentiment from the review.
Return JSON: [{"aspect": str, "sentiment": "pos|neg|neu", "evidence": str}].
Review: "{text}"
"""
result = llm.chat(prompt.format(text=review)).json()

04Step 3 — Make it actionable

A dashboard of percentages does not drive action. Give product managers a feed of negative reviews per feature, with evidence quotes and a trend line.

05Common pitfalls

06How to evaluate honestly

07When to use a model versus an LLM

08