TEXT CLASSIFICATION &

SENTIMENT ANALYSIS

A Complete Guide to
Understanding NLP Classification
Feature Engineering | ML Models | Sentiment Analysis | Evaluation Metrics
Dr. Ajay Prakash B V • AI & ML Department • Dr. Ambedkar Institute of Technology
Today's Roadmap — 9 Sections

0
1 Introduction to Text Classification

0
2 Real-Time Applications

0
3 Feature Engineering (BoW, TF-IDF, Embeddings)

0
4 Evaluation Metrics (Accuracy, Precision, Recall, F1)

0
5 Sentiment Analysis — Concepts & Definition

0
6 Types of Sentiment Analysis

0
7 Approaches (Lexicon-Based & ML)

0
8 Evaluation of Sentiment Models

0
9 Case Study — Aspect-Based Example

Text Classification & Sentiment Analysis 2 / 26

 Introduction to
SECTION 01
Text Classification

What it is, why it matters, and how it works

What Is Text Classification?

Text Classification is the task of automatically assigning a predefined category (label) to a piece of text based on its
content. Given raw text, the system determines which class it belongs to from a fixed set of categories.

Types of Text Classification:

Binary Multi-Class
2 classes: Spam / Not Spam 3+ classes: Sports / Politics / Tech
Ex: Email filtering Ex: News categorization

Multi-Label Hierarchical
Multiple labels per doc Labels in tree structure
Ex: Movie: [Action, Comedy] Ex: Electronics → Phones

Text classification is the most widely-used NLP task — spam filters, news feeds, chatbots, and search engines all depend on it.

Text Classification & Sentiment Analysis 4 / 26

 The Text Classification Pipeline — 5 Stages

→ → → →
Data Text Feature Model Evaluation
Collection Preprocessing Extraction Training & Deploy

Gather labeled text + Clean: lowercase, Convert text → Learn patterns (NB, Test accuracy, F1,
categories tokenize, stopwords numbers (TF-IDF) SVM, DT) deploy

Example — Spam Detection Pipeline:

Input: "Congratulations! You won a FREE iPhone! Click NOW!"

Preprocess → "congratulations won free iphone click" → TF-IDF → NB Model → SPAM

Text Classification & Sentiment Analysis 5 / 26

 Real-Time Applications
SECTION 02 of Text Classification

How companies use text classification in practice

 8 Major Applications — Industry Use Cases

Spam Detection Social Sentiment News Categorization Support Tickets

Gmail blocks 100M spam/day Brands monitor Twitter opinion Sports / Politics / Tech / Business Route to billing / logistics / tech

Document Org Medical Text Fake News Review Analysis

Extract sentiment from product
Classify contracts, invoices, records ICD codes, diagnosis, lab reports Flag misleading content at scale
reviews

McKinsey: NLP can automate 15-25% of knowledge worker tasks, saving millions of hours annually.

Text Classification & Sentiment Analysis 7 / 26

 Feature Engineering
SECTION 03
for Text Data

Converting text into numbers that ML models can process

Bag of Words (BoW) — Count Word Occurrences

BoW represents each document as a vector of word counts. Each unique word becomes a column. Word order is completely
ignored — like throwing words into a bag.

Doc 1: "I love NLP" | Doc 2: "I love ML" | Doc 3: "NLP and ML are great"

I love NLP ML and are great

Doc 1 1 1 1 0 0 0 0

Doc 2 1 1 0 1 0 0 0

Doc 3 0 0 1 1 1 1 1

Cons: Ignores word order, common words dominate, very

Pros: Simple, intuitive, easy to implement, good baseline.
sparse.

N-grams — Capturing Word Sequences:

Unigram (N=1) Bigram (N=2) Trigram (N=3)
["I love", "love natural", "natural ["I love natural", "love natural
["I", "love", "natural", "language"]
language"] language"]

Text Classification & Sentiment Analysis 9 / 26

TF-IDF — The Gold Standard for Text Features

TF-IDF upweights rare, informative words and downweights common ones. It answers: "How important is this word to THIS
specific document relative to ALL documents?"

TF(t,d) = Count(t in d) / Total words in d

IDF(t) = log(Total documents / Documents containing t)
TF-IDF(t,d) = TF(t,d) × IDF(t)

Numerical Example (1000 documents):

Word Appears in IDF Meaning

"machine" 10 / 1000 docs log(1000/10) = 2.0 Rare → HIGH IDF → Important!

"data" 200 / 1000 docs log(1000/200) = 0.7 Moderate frequency → Medium IDF

"the" 990 / 1000 docs log(1000/990) ≈ 0.004 Very common → NEAR-ZERO IDF

TF-IDF is the default choice for traditional ML text classifiers (NB, SVM). It consistently outperforms raw BoW counts because it filters out
uninformative common words automatically.

Text Classification & Sentiment Analysis 10 / 26

Advanced: Word Embeddings & BERT

Word2Vec / GloVe Sentence Embeddings BERT (Contextual)

Same word gets DIFFERENT vectors based

Each word = dense vector (100-300 dims). Entire sentence = single vector.
on context.
Captures semantic similarity. Methods: Doc2Vec, Sentence-BERT.
"bank" (financial) ≠ "bank" (river).
king − man + woman ≈ queen Captures overall meaning.
State-of-the-art for classification.

Feature Methods Comparison:

Method Dims Semantics? Context? Best For

Bag of Words Very High No No Simple baselines

TF-IDF Very High Partial No Traditional ML

Word2Vec/GloVe 100-300 Yes No Similarity tasks

BERT 768 Yes Yes State-of-the-art

Text Classification & Sentiment Analysis 11 / 26

 Evaluation Metrics for
SECTION 04 Text Classification

Measuring model performance beyond just accuracy

The Confusion Matrix — Foundation of All Metrics

Predicted: Positive Predicted: Negative TP: Correct positive

TN: Correct negative

Actual: Positive TP (True Positive) FN (False Negative)
FP: False alarm

Actual: Negative FP (False Positive) TN (True Negative) FN: Missed positive

Example: Spam detection on 100 emails (40 spam, 60 not spam)

Pred: Spam Pred: Not Spam From this matrix:

Accuracy = 92/100 = 92%
Actual: Spam TP = 35 FN = 5 Precision = 35/38 = 92.1%
Recall = 35/40 = 87.5%
Actual: Not Spam FP = 3 TN = 57 F1 = 2×(.921×.875)/(.921+.875) = 89.7%

Accuracy can be misleading! If 95% of emails are not spam, a model that always says 'not spam' gets 95% accuracy but catches ZERO spam.
Always check Precision, Recall, and F1.

Text Classification & Sentiment Analysis 13 / 26

Precision, Recall & F1-Score — Formulas & Meaning

Accuracy 92%
(TP + TN) / Total Overall correctness
Use when: Balanced classes

Precision 92.1%
TP / (TP + FP) Of predicted positive, how many correct?
Use when: Cost of false alarms is HIGH

Recall 87.5%
TP / (TP + FN) Of actual positive, how many found?
Use when: Cost of missing positives is HIGH

F1-Score 89.7%
2 × (P × R) / (P + R) Harmonic mean of P and R
Use when: Need balanced single metric

Text Classification & Sentiment Analysis 14 / 26

 Sentiment Analysis
Concepts & Definition
SECTION 05

Identifying emotions and opinions in text

What Is Sentiment Analysis?

Sentiment Analysis (Opinion Mining) is the computational task of identifying the emotional
tone, opinion, or attitude expressed in text. It answers: "How does the author feel about
this subject?"

Why It Matters:

Brand Monitoring Market Research

Track opinion across millions of social posts Gauge customer satisfaction at scale

Political Analysis Product Insights

Monitor public opinion in real-time Identify what customers love/hate

4 Levels of Analysis:

Document Level Sentence Level Aspect Level Word Level

Text Classification & Sentiment Analysis 16 / 26

 Types of
Sentiment Analysis
SECTION 06

Binary, Multi-Class, and Aspect-Based approaches

Binary & Multi-Class Sentiment Analysis

Binary Sentiment (2 classes) Multi-Class (3+ classes)

"Absolutely fantastic movie!" POSITIVE "Best phone ever! Love it!" 5★ Very Positive

"Terrible waste of time." NEGATIVE "Good, meets expectations." 4★ Positive

"Loved every moment!" POSITIVE "It's okay, nothing special." 3★ Neutral

"Worst product ever." NEGATIVE "Disappointed with quality." 2★ Negative

Aspect-Based Sentiment Analysis (ABSA) — Most Sophisticated

"The camera is excellent but battery life is terrible and it overheats during gaming."

Camera → POSITIVE Battery → NEGATIVE Heat Mgmt → NEGATIVE

Text Classification & Sentiment Analysis 18 / 26

 Approaches for
Sentiment Analysis
SECTION 07

Lexicon-Based vs Machine Learning methods

Lexicon-Based Approach — Using Sentiment Dictionaries

Uses a pre-built dictionary where each word has a polarity score. No training data needed — just look up words and combine
scores.

Word Score Sentiment Popular Lexicon Tools:

excellent +3 Strongly Positive
VADER Best for social media, emojis
good +1 Positive
okay 0 Neutral TextBlob Polarity + subjectivity scores
bad −1 Negative
SentiWordNet WordNet-based, detailed scores
terrible −3 Strongly Negative
AFINN 2,477 words rated −5 to +5

Worked Example:
"The camera quality is excellent but the battery is terrible"
With ABSA: Camera → Positive (+3), Battery →
excellent (+3) + terrible (−3) = 0 → Neutral overall
Negative (−3)

Text Classification & Sentiment Analysis 20 / 26

Machine Learning Approach — Learning from Labeled Data
Treats sentiment analysis as supervised classification: train a model on labeled examples
(text + known sentiment), then predict sentiment for new text. Adapts to specific domains
and vocabularies.

Collect Preprocess Extract Train Evaluate

Labeled Data
→ Text
→ Features
→ Classifier
→ & Predict

Lexicon vs Machine Learning — Head-to-Head:

Criterion Lexicon-Based Machine Learning

Training data? Not needed Required (labeled)

Domain adaptability Limited High (learns domain language)
Accuracy Moderate (70-80%) Higher (85-95%)
Handles negation Difficult Learns from data patterns
Setup effort Very low (plug & play) Higher (data + training)
Best for Quick analysis, no data Production, high accuracy
Text Classification & Sentiment Analysis 21 / 26
 Evaluation of
SECTION 08 Sentiment Models

Cross-validation, imbalance, and error analysis

Model Evaluation — Beyond Standard Metrics

Cross-Validation Dataset Imbalance

Amazon reviews: ~80% positive, ~20% negative. Model

Split data into k folds (typically k=5 or 10). Train on k−1 folds,
becomes biased toward majority class. Solutions: SMOTE
test on the remaining fold. Repeat k times and average
oversampling, undersampling, class weights, use F1 instead
results. Every data point is used for both training and testing.
of accuracy.

Common Error Types in Sentiment Analysis:

Sarcasm "Oh great, another update that breaks everything" Model sees 'great' → Positive

Negation "This product is not good" Model sees 'good' → Positive

Mixed
"Food was delicious but price outrageous" Conflicting signals cancel out
Sentiment

Implicit "Phone lasted 2 hours on full charge" No explicit negative words

Text Classification & Sentiment Analysis 23 / 26

 Case Study —
SECTION 09
Aspect-Based Analysis

Analyzing a real product review with all approaches

Case Study: "The phone design is excellent but the battery
drains quickly."
VADER (Lexicon-Based) Aspect-Based Analysis

"excellent" → +0.75 Phone Design → POSITIVE

"but" → contrast signal Evidence: "excellent"
"drains quickly" → −0.45
Battery Life → NEGATIVE
Compound ≈ +0.25 → Weakly Positive Evidence: "drains quickly"

Approach Comparison:

Approach Prediction Aspect-Level? Actionable?

VADER Weakly Positive No Limited — hides battery issue

Binary ML Positive No Limited — overall only
Aspect-Based Design: +, Battery: − Yes High — knows exactly what to fix
Teaching Point: ABSA is the most valuable for businesses — overall sentiment hides critical insights. The manufacturer knows the design is
praised but battery needs improvement.

Text Classification & Sentiment Analysis 25 / 26

KEY TAKEAWAYS

Text Classification Assigns predefined categories to text — foundation of NLP applications

Feature Engineering BoW → TF-IDF → Word2Vec → BERT: each level captures more meaning

Evaluation Use Accuracy + Precision + Recall + F1 + Confusion Matrix together

Sentiment Types Binary (Pos/Neg) → Multi-class (5★) → Aspect-based (per feature)

Two Approaches Lexicon (no training, quick) vs ML (higher accuracy, needs data)

ABSA is King Aspect-based analysis gives the most actionable business insights

Error Awareness Sarcasm, negation, mixed sentiment — know your model's weaknesses

Thank You! Questions?