I recently started working on my first NLP and Sentiment Analysis project using the Amazon Reviews dataset on Kaggle.

I expected to find a normal CSV file that I could load using pandas.read_csv().

Instead, I found files like this:

train.ft.txt.bz2
test.ft.txt.bz2

At first, I was confused.

• What is a .bz2 file?
• How do I read it?
• Why isn't there a CSV file?

After some searching and experimentation, I discovered that the dataset is stored in a compressed BZip2 format. Fortunately, Python provides a built-in bz2 library that can read these files directly.

Step 1: Reading the compressed file

The first step was simply reading the compressed file.

import bz2 
import pandas as pd 
file_path = "/kaggle/input/datasets/bittlingmayer/amazonreviews/train.ft.txt.bz2"
with bz2.open(file_path, "rt", encoding="utf-8") as f:
 lines = f.readlines() 
print(lines[0])

The output looked something like this:
__label__2 Stunning even for the non-gamer: This soundtrack was beautiful...

At this point, I noticed that every line starts with a label followed by the review text.

Step 2: Separating Labels and Reviews

Next, I separated the sentiment label from the review text.

labels = []
sentences = []

for line in lines:
    label = 1 if line.split(" ")[0] == "__label__2" else 0
    sentence = line.split(" ", 1)[1]

    labels.append(label)
    sentences.append(sentence)

The dataset uses:

• __label__1 → Negative Review
• __label__2 → Positive Review

Since machine learning models work better with numerical values, I converted them into:

• 0 → Negative
• 1 → Positive