NLP Model Performance: Raw vs Cleaned Data

# Raw vs Cleaned NLP: Does Data Cleaning Beat Model Choice? This notebook compares **two training pipelines** on the IMDB 50K movie reviews sentiment dataset: 1. **Raw text** — minimal preprocessing (just lowercase + basic tokenization) 2. **Cleaned text** — full NLP cleaning (remove HTML, URLs, punctuation, stopwords, lemmatize) The same **LogisticRegression** model is trained on both, then compared against a **RandomForest** on cleaned data to answer: *does simple cleaning improve results more than changing the model?*

# ── Setup & Imports ──────────────────────────────────────────────
%pip install -q kagglehub

import pandas as pd
import numpy as np
import re
import html

from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import (
    accuracy_score, precision_score, recall_score, f1_score,
    confusion_matrix, classification_report
)
from sklearn.model_selection import train_test_split
from sklearn.pipeline import Pipeline

…

Note: you may need to restart the kernel to use updated packages.
✅ Imports ready

# ── 1. Download Dataset ──────────────────────────────────────────
print("Downloading IMDB 50K Movie Reviews from Kaggle...")
path = kagglehub.dataset_download("lakshmi25npathi/imdb-dataset-of-50k-movie-reviews")
data_path = Path(path) / "IMDB Dataset.csv"

df = pd.read_csv(data_path)
print(f"✅ Loaded! Shape: {df.shape}")
print(f"Columns: {list(df.columns)}")
print(f"\nFirst 3 reviews:")
df.head(3)

Downloading IMDB 50K Movie Reviews from Kaggle...
Downloading to /root/.cache/kagglehub/datasets/lakshmi25npathi/imdb-dataset-of-50k-movie-reviews/1.archive...
Extracting files...
✅ Loaded! Shape: (50000, 2)
Columns: ['review', 'sentiment']

First 3 reviews:

## Step 1: Explore the Dataset Before building pipelines, let's understand the data quality — class balance, text length, and baseline noisy elements (HTML tags, URLs, punctuation).

# ── 2. EDA / Profile the Dataset ────────────────────────────────
df = pd.read_csv(data_path)  # fresh copy

print("─── Dataset Profile ───")
print(f"Rows: {df.shape[0]:,}   Columns: {df.shape[1]}")
print(f"\n--- Sentiment Distribution ---")
print(df['sentiment'].value_counts())
print(f"\n--- Missing values ---")
print(df.isnull().sum())

# Quick text stats
df['review_len'] = df['review'].str.len()
df['word_count'] = df['review'].str.split().str.len()
print(f"\n--- Text Length Stats (chars) ---")
print(df['review_len'].describe())
print(f"\n--- Word Count Stats ---")
print(df['word_count'].describe())

…

─── Dataset Profile ───
Rows: 50,000   Columns: 2

--- Sentiment Distribution ---
sentiment
positive    25000
negative    25000
Name: count, dtype: int64

--- Missing values ---
review       0
sentiment    0
dtype: int64

--- Text Length Stats (chars) ---
count    50000.000000
mean      1309.431020
std        989.728014
min         32.000000
25%        699.000000
50%        970.000000
75%       1590.250000
max      13704.000000
Name: review_len, dtype: float64

--- Word Count Stats ---
count…

# ── Visual EDA ──────────────────────────────────────────────────
fig, axes = plt.subplots(1, 3, figsize=(16, 4))

# 1. Class balance
df['sentiment'].value_counts().plot(kind='bar', ax=axes[0], color=['#4e79a7', '#e15759'])
axes[0].set_title('Class Balance', fontsize=13)
axes[0].set_ylabel('Count')
axes[0].tick_params(axis='x', rotation=0)

# 2. Review length distribution by sentiment
for sent in ['positive', 'negative']:
    subset = df[df['sentiment'] == sent]['review_len']
    axes[1].hist(subset, bins=60, alpha=0.6, label=sent, density=True)
axes[1].set_title('Review Length Distribution', fontsize=13)
axes[1].set_xlabel('Character Count')
axes[1].legend()

# 3. Estimate noise: % of reviews with HTML tags, URLs
…

HTML tags: 58.4% of reviews
URLs:      0.2% of reviews
Special chars: 100.0% of reviews

## Step 2: Text Preprocessing Pipelines Two parallel pipelines: - **Raw**: only lowercase + basic splitting (minimal) - **Cleaned**: strip HTML, decode entities, remove URLs/@mentions/punctuation, lowercase, remove stopwords, lemmatize

# ── 3. Text Preprocessing Pipelines ──────────────────────────────
%pip install -q nltk

import nltk
nltk.download('stopwords', quiet=True)
nltk.download('wordnet', quiet=True)
nltk.download('omw-1.4', quiet=True)
nltk.download('punkt_tab', quiet=True)
from nltk.corpus import stopwords
from nltk.stem import WordNetLemmatizer

STOPWORDS = set(stopwords.words('english'))
lemmatizer = WordNetLemmatizer()

def pipeline_raw(text: str) -> str:
    """Minimal: lowercase only."""
    return text.lower()

…