SentimentAnalysis_CoreMLDemo_EmotionClassifier_iOS_SwiftUI

This is an iOS demo application for emotion classification using CoreML and SwiftUI. The project follows MVVM architecture with async/await and actor-based model management.

The CoreML model was converted from a Scikit-learn pipeline using coremltools Python package, including:

• DictVectorizer for feature extraction
• LinearSVC for classification

---

Dataset

The training dataset is from Kaggle: Emotions Dataset for NLP. But I have added the preprocessed and formatted CSV data in the converter folder for direct use in training.

Steps:

1. Original text files converted to CSV with two columns:

   • text → user sentence or phrase
   • emotion → label (e.g., joy, sadness, anger, fear, love, surprise)

2. Python example to convert .txt → .csv:

import csv
from pathlib import Path

# Input and output file paths
input_file = Path("val.txt")
output_file = Path("val.csv")

# Ensure input exists
if not input_file.exists():
    raise FileNotFoundError(f"❌ Input file not found: {input_file}")

# Open the input text file and create the output CSV file
with input_file.open("r", encoding="utf-8") as infile, output_file.open("w", newline="", encoding="utf-8") as outfile:
    writer = csv.writer(outfile)
    writer.writerow(["text", "emotion"])  # CSV header
    
    count = 0
    for line in infile:
        line = line.strip()
        if not line:
            continue
        if ";" in line:
            text, emotion = line.split(";", 1)
            writer.writerow([text.strip(), emotion.strip()])
            count += 1
        else:
            print(f"⚠️ Skipping malformed line: {line}")
    
print(f"✅ Conversion complete! {count} entries saved to '{output_file.name}'")

---

Python Setup

1. Install Python 3.9+
2. Install dependencies:

pip install pandas numpy nltk scikit-learn coremltools

3. Download NLTK data:

Step 1: Open Python 3 in your terminal

python3

Step 2: Run the following commands in the Python prompt (>>>)

>>> import nltk
>>> nltk.download('punkt')
>>> nltk.download('stopwords')

Step 3: Exit Python

>>> exit()

---

Training the Model

Run:

python3 train_emotion_model.py

Steps performed:

• Load CSV data
• Preprocess text
• Extract bag-of-words features
• Train LinearSVC using GridSearchCV
• Save CoreML model as EmotionClassifier.mlmodel

---

How the Model Learns (Step by Step Example)

This section explains how the emotion classification model learns from training data using a step-by-step example with multiple iterations.

---

Example Training Dataset

Text	Emotion
"I am happy today"	joy
"I feel sad"	sadness
"I am really angry"	anger
"I love my family"	love
"I am scared of the dark"	fear
"Wow, I didn’t expect that!"	surprise
"I am ecstatic and joyful!"	joy
"I am very upset and angry"	anger
"I adore my friends"	love
"The night makes me anxious"	fear
"I am shocked by the news"	surprise
"Feeling sad and lonely"	sadness

---

Step 1: Convert Text → Features (Bag-of-Words)

Take "I am happy today" → tokenize, remove stopwords:

['happy', 'today']

Feature dict (all words from dataset considered):

{'happy':1, 'today':1, 'sad':0, 'angry':0, 'love':0, 'scared':0, 'wow':0, 'expect':0, 'ecstatic':0, 'joyful':0, 'upset':0, 'family':0, 'friends':0, 'anxious':0, 'shocked':0, 'lonely':0}

---

Step 2: Initialize Weights

Each emotion has a weight vector for all words (initialized to 0):

weights_joy = [happy=0, today=0, sad=0, angry=0, love=0, scared=0, wow=0, expect=0, ecstatic=0, joyful=0, upset=0, family=0, friends=0, anxious=0, shocked=0, lonely=0]
weights_sadness = same...
weights_anger = same...
weights_love = same...
weights_fear = same...
weights_surprise = same...

---

Step 3: Training Iterations

Iteration 1: "I am happy today" → joy

Update weights:

• Increase weights for the true emotion for the words present.
• Decrease weights for all other emotions for the words present.

weights_joy      = [happy=2, today=1, ...]
weights_sadness  = [happy=-1, today=-0.5, ...]
weights_anger    = [happy=-1, today=-0.5, ...]
weights_love     = [happy=-0.5, today=-0.2, ...]
weights_fear     = [happy=-0.5, today=-0.2, ...]
weights_surprise = [happy=-0.5, today=-0.2, ...]

Iteration 2: "I feel sad" → sadness

weights_sadness  = [happy=-1, today=-0.5, sad=2, ...]
weights_joy      = [happy=2, today=1, sad=-1, ...]
weights_anger    = [happy=-1, today=-0.5, sad=-1, ...]
weights_love     = [happy=-0.5, today=-0.2, sad=-0.5, ...]
weights_fear     = [happy=-0.5, today=-0.2, sad=-0.5, ...]
weights_surprise = [happy=-0.5, today=-0.2, sad=-0.5, ...]

Iteration 3: "I am really angry" → anger

weights_anger   = [angry=2, really=1, ...]
weights_joy     = decrease weights for angry, really
weights_sadness = decrease weights for angry, really
weights_love    = decrease weights for angry, really
weights_fear    = decrease weights for angry, really
weights_surprise= decrease weights for angry, really

Iteration 4: "I love my family" → love

weights_love    = [love=2, family=1, ...]

Iteration 5: "I am scared of the dark" → fear

weights_fear    = [scared=2, dark=1, ...]

Iteration 6: "Wow, I didn’t expect that!" → surprise

weights_surprise = [wow=2, expect=1, ...]

Iteration 7: "I am ecstatic and joyful!" → joy

weights_joy      = [happy=2, today=1, ecstatic=1, joyful=1, ...]
weights_other   = decrease weights for ecstatic, joyful

Iteration 8: "I am very upset and angry" → anger

weights_anger    = [angry=3, upset=2, ...]
weights_other    = decrease weights for angry, upset

Iteration 9: "I adore my friends" → love

weights_love     = [love=3, family=1, friends=1, ...]

Iteration 10: "The night makes me anxious" → fear

weights_fear     = [scared=2, dark=1, anxious=1, ...]

Iteration 11: "I am shocked by the news" → surprise

weights_surprise = [wow=2, expect=1, shocked=1, ...]

Iteration 12: "Feeling sad and lonely" → sadness

weights_sadness  = [sad=3, lonely=1, ...]

Final Weights Table (Word → Emotion)

Word	Joy	Sadness	Anger	Love	Fear	Surprise
happy	2	-1	-1	-0.5	-0.5	-0.5
today	1	-0.5	-0.5	-0.2	-0.2	-0.2
sad	-1	3	-1	-0.5	-0.5	-0.5
angry	-1	-1	3	-0.5	-0.5	-0.5
love	-0.5	-0.5	-0.5	3	-0.5	-0.5
scared	-0.5	-0.5	-0.5	-0.5	2	-0.5
wow	-0.5	-0.5	-0.5	-0.5	-0.5	2
expect	-0.5	-0.5	-0.5	-0.5	-0.5	1
ecstatic	1	-0.5	-0.5	-0.5	-0.5	-0.5
joyful	1	-0.5	-0.5	-0.5	-0.5	-0.5
upset	-0.5	-0.5	2	-0.5	-0.5	-0.5
family	-0.5	-0.5	-0.5	1	-0.5	-0.5
friends	-0.5	-0.5	-0.5	1	-0.5	-0.5
anxious	-0.5	-0.5	-0.5	-0.5	1	-0.5
shocked	-0.5	-0.5	-0.5	-0.5	-0.5	1
lonely	-0.5	1	-0.5	-0.5	-0.5	-0.5

Note: Positive values indicate the strength of association between a word and an emotion. Negative values reduce association with other emotions. This table is used for predicting emotions of new sentences.

---

Step 4: Prediction Formula

When a new sentence comes in, the model computes:

score_emotion = sum(word_count * weight_for_that_emotion)

Example: Emotion Prediction for a Sentence

Sentence: "I feel so happy and excited today!"

Steps:

1. Tokenize ['i', 'feel', 'so', 'happy', 'and', 'excited', 'today']

2. Remove stopwords ['happy', 'excited', 'today']

3. Create feature dictionary {'happy': 1, 'excited': 1, 'today': 1}

4. Compute score_emotion for each class

Emotion	Score Calculation	Score
Joy	2(happy) + 1(excited) + 1(today)	4
Sadness	-1(happy) + -0.5(excited) + -0.5(today)	-2
Anger	-1(happy) + -0.5(excited) + -0.5(today)	-2
Love	-0.5(happy) + -0.5(excited) + -0.5(today)	-1.5
Fear	-0.5(happy) + -0.5(excited) + -0.5(today)	-1.5
Surprise	-0.5(happy) + -0.5(excited) + -0.5(today)	-1.5

Predicted Emotion: Joy (highest score = 4)

---

Evaluation Metrics Explained

This section provides detailed explanations of all the evaluation metrics used in the training report of the emotion classification model.

---

1. Precision

• Definition: Measures how many of the predicted instances for a class are actually correct.
• Formula:

$$Precision = True Positives / (True Positives + False Positives)$$

• Example:

  • Model predicts 10 texts as joy
  • 8 of them are actually joy
  • Precision = 8 / 10 = 0.8

• Interpretation: High precision means fewer false positives.

---

2. Recall (Sensitivity or True Positive Rate)

• Definition: Measures how many of the actual instances of a class were correctly predicted.
• Formula:

$$Recall = True Positives / (True Positives + False Negatives)$$

• Example:

  • There are 12 actual joy texts
  • Model correctly predicted 8 of them
  • Recall = 8 / 12 ≈ 0.667

• Interpretation: High recall means fewer false negatives.

---

3. F1-Score

• Definition: Harmonic mean of precision and recall, balancing both metrics.
• Formula:

$$F1 = 2 * (Precision * Recall) / (Precision + Recall)$$

• Example:

  • Precision = 0.8, Recall = 0.667
  • F1 = 2 * (0.8 * 0.667) / (0.8 + 0.667) ≈ 0.727

• Interpretation: High F1-score → model is both accurate and complete for that class.

---

4. Support

• Definition: Number of actual instances of each class in the dataset.

• Example:

  • If there are 432 anger texts in validation set → support = 432

• Interpretation: Helps understand class distribution and performance per class.

---

5. Accuracy

• Definition: Overall ratio of correct predictions to total predictions.
• Formula:

$$Accuracy = Total Correct Predictions / Total Predictions$$

• Interpretation: Overall measure of model correctness.

---

6. Macro Average (Macro Avg)

• Definition: Average of precision, recall, and F1-score across all classes equally, ignoring class imbalance.
• Interpretation: Treats all classes as equally important.

---

7. Weighted Average (Weighted Avg)

• Definition: Average of precision, recall, and F1-score weighted by number of instances (support) per class.
• Interpretation: Reflects performance considering class distribution.

---

8. Confusion Matrix

• Definition: Table showing the counts of actual vs predicted classes.
• Structure: Rows = actual classes, Columns = predicted classes
• Example:

Actual \ Predicted	joy	sadness	anger	love	fear	surprise
joy	8	3	1	0	0	0

• Interpretation: Helps identify which classes the model confuses. Each cell indicates how many instances of an actual class were predicted as a particular class.

---

Actual Test Report Discussion

This section explains the training and validation results obtained from the emotion classification model.

---

Training and Validation Accuracy

• Training Accuracy: 0.9817

  • Indicates that the model correctly predicted 98.17% of training examples.
  • High accuracy shows the model has learned the patterns in the training data well.

• Validation Accuracy: 0.8978

  • Indicates 89.78% correct predictions on unseen validation data.
  • Slightly lower than training accuracy, showing a good balance and not much overfitting.

---

Classification Report

Emotion	Precision	Recall	F1-Score	Support
anger	0.889	0.873	0.881	432
fear	0.869	0.889	0.879	387
joy	0.901	0.927	0.914	1072
love	0.804	0.785	0.795	261
sadness	0.937	0.929	0.933	933
surprise	0.887	0.748	0.811	115

Observations

• Best performing classes: joy and sadness with high precision, recall, and F1-score. The model easily recognizes these due to larger support.
• Lower performing class: surprise with lower recall (0.748), likely due to fewer examples and overlapping features with other classes.
• Balanced classes: anger and fear show balanced performance (precision ~0.87-0.89, recall ~0.87-0.89).
• Love: Moderate performance, possibly due to limited vocabulary and overlap with joy/love sentiments.

---

Confusion Matrix

[[377   7  21   3  24   0]
 [ 11 344   9   2  14   7]
 [ 18   3 994  39  15   3]
 [  2   1  50 205   3   0]
 [ 16  19  24   6 867   1]
 [  0  22   5   0   2  86]]

Confusion Matrix with Emotion Labels

| Actual \ Predicted | joy | sadness | anger | love | fear | surprise |
| ------------------ | --- | ------- | ----- | ---- | ---- | -------- |
| joy                | 377 | 7       | 21    | 3    | 24   | 0        |
| sadness            | 11  | 344     | 9     | 2    | 14   | 7        |
| anger              | 18  | 3       | 994   | 39   | 15   | 3        |
| love               | 2   | 1       | 50    | 205  | 3    | 0        |
| fear               | 16  | 19      | 24    | 6    | 867  | 1        |
| surprise           | 0   | 22      | 5     | 0    | 2    | 86       |

Interpretation

• Rows = Actual emotions
• Columns = Predicted emotions
• Diagonal values = Correct predictions
• Off-diagonal values = Misclassifications

Example Insights

• 21 joy texts misclassified as anger
• 22 surprise texts misclassified as sadness
• Diagonal values indicate how well each class is predicted

---

Key Takeaways

1. The model is highly accurate for common emotions like joy and sadness.
2. Performance drops for rare or subtle emotions like surprise and love.
3. Confusion matrix provides insight for future improvements such as adding more training examples for underrepresented classes.
4. Overall, the model generalizes well to unseen data with 89.78% validation accuracy.

---

iOS App Usage

1. Open EmotionClassifier.xcodeproj in Xcode 14+
2. Run on simulator/device
3. Enter text → tap Analyze → shows predicted emotion & confidence

---

Tech Stack

• Swift 5 / SwiftUI
• CoreML
• MVVM (async/await, actor)
• Python (Scikit-learn → CoreML)

---

License

MIT License

---

Author

• Muhammad Haroon
• Email: mianmharoon72@gmail.com
• LinkedIn: https://www.linkedin.com/in/mian-haroon

---

References

• CoreML Tools Documentation
• Emotions Dataset on Kaggle
• Scikit-learn Documentation
• SwiftUI Documentation