1.1 What is sentiment analysis?

A human can read one review and tell you the customer is happy. The problem is scale: a brand may receive thousands of reviews, tweets, and messages every day. No team can read them all.

1.2 Why marketers care

Every message a customer writes is a tiny piece of feedback. Read together, they answer questions marketers ask constantly:

1.3 The three sentiment classes

In this dataset, every message is labelled as one of three classes. Here is a real example of each:

Knowledge Check — Section 1

2.1 The dataset structure

Our file has just three columns and 100 rows. Simple, but enough to learn the whole workflow.

2.2 How the classes are balanced

Counting the labels reveals the messages are not split evenly across the three classes:

2.3 Real text is messy

Real customer text is never tidy. Look closely at these actual rows and spot the problems:

• Broken grammar — "Excellent the product!" is not a normal sentence.
• Stuck-together words — "excellentchoice" should be two words.
• Stray words, mixed CAPS, punctuation, and emojis all mixed in.

Knowledge Check — Section 2

3.1 The sentiment analysis pipeline

Whatever method we use, the message travels through the same four steps:

3.2 Approach A — the dictionary method

Example on a real message — "Excellent product, but slow support and poor packaging."

Total score = +1 −1 −1 = −1  →  below zero  →  Negative

3.3 Approach B — the learning method

3.4 Comparing the two approaches

Knowledge Check — Section 3

4.1 Why we clean the text first

Computers match words exactly. To them, all of these look like different words:

4.2 The cleaning steps

We tidy the text one step at a time. Watch a real message get cleaned:

4.3 Don't throw away the emojis

It is tempting to delete emojis as "junk." In this dataset that would be a mistake — the emoji is a perfect clue to sentiment:

Knowledge Check — Section 4

5.1 Accuracy — the simplest score

To check a method, we compare its guesses against the human labels and count how many it got right.

$$\text{Accuracy} = \frac{\text{Number of correct predictions}}{\text{Total number of messages}}$$

Accuracy is easy to explain to a client — but on its own it can be misleading. Here's why.

5.2 The imbalance trap

Remember our data is 60% Positive. Imagine a lazy method that ignores the text and just guesses Positive every single time:

This is why analysts also check each class separately — especially the rare but important Negative messages, which a lazy method would miss entirely.

Knowledge Check — Section 5

6.1 Marketing applications

Once you can score sentiment automatically, a lot of marketing work becomes measurable:

• Brand health tracking — watch the positive vs negative share week by week.
• Campaign evaluation — did sentiment lift after a launch, ad, or promotion?
• Product feedback — mine negative reviews for the top recurring complaints.
• Customer service triage — route angry messages to a human first.
• Competitor comparison — measure how people feel about you vs rivals.

6.2 Where it goes wrong

Sentiment analysis is powerful but not magic. Watch out for:

6.3 Key takeaways

1. Sentiment analysis sorts text into Positive, Neutral, or Negative — measuring emotion at scale.
2. Always look at your data first: our dataset is imbalanced (60 / 24 / 16) and the text is messy.
3. The pipeline is: clean → find clues → decide → label.
4. Two approaches: the dictionary method (word scores) and the learning method (learns from examples).
5. Clean the text so words match — but keep useful clues like emojis.
6. Judge results with care: on imbalanced data, beat the "always guess the biggest class" baseline.