Sentiment_Data_Analysis_Using_Multinomial_Naive_Bayes

Overview

Techniques Used

• Bag-of-Words (BoW) feature representation
• Multinomial Naive Bayes
• Laplace smoothing (tested across different smoothing factors)

Experiments Included

1. Processed Text + Naive Bayes (smoothing factor comparison)
2. Raw Text + Naive Bayes (smoothing factor comparison)
3. Removing Rare Words (frequency threshold < 5 removed)
4. Impact of Sentiment on Review Lengths (distribution analysis)

---

Experiments & Key Results

Experiment 1 — Processed Text (BoW + Naive Bayes)

Preprocessing applied: tokenization + stopword removal.
Smoothing factors evaluated: 0.1, 0.3, 0.5, 1.0, 1.5, 1.8, 1.9, 2.0.

Best result: smoothing factor 1.8

• Precision 0.85592
• Recall 0.84009
• F1 0.84793

📌 Figure: Impact of smoothing factors on precision/recall/F1

---

Experiment 2 — Raw Text (BoW + Naive Bayes)

No preprocessing (raw reviews).
Smoothing factors evaluated: 0.1 → 3.0.

Best result: smoothing factor 0.5

• Precision 0.88033
• Recall 0.84914
• F1 0.86445

📌 Figure: Smoothing factors vs precision/recall/F1 using raw text

---

Experiment 3 — Removing Rare Words

Rare words appearing < 5 times are removed to reduce vocabulary size and noise.

• Rare words identified: 70,178
• Metrics after retraining:
  • Precision 0.8534
  • Recall 0.8425
  • F1 0.8479

Comparison (Cleaned Rare Words vs Processed Text):

Metric	Cleaned Rare Words	Processed Text
Precision	0.8534	0.8553
Recall	0.8425	0.8405
F1 Score	0.8479	0.8478

📌 Figure: Cleaned rare words vs processed text

---

Experiment 4 — Impact of Sentiment on Review Lengths

This experiment analyzes the relationship between review length and sentiment label.

Observation:

• Negative reviews are more concentrated around shorter lengths.
• Positive reviews show a slightly broader spread.
• There is no clear length threshold that cleanly separates positive vs negative, but length may still help as a feature in ambiguous cases.

📌 Figure: Histogram of review lengths by sentiment

---

Summary of Findings (Experiments 1–4)

• Raw text Naive Bayes performed better than processed text in terms of F1 score, suggesting raw text retains helpful context.
• Removing rare words provides similar overall performance while reducing complexity and potentially improving efficiency.
• Review length alone is not a reliable sentiment indicator, but trends suggest it could be a useful auxiliary feature.