A machine learning application that analyzes natural-language text and classifies it into one of six human emotion categories:
Sadness · Joy · Love · Anger · Fear · Surprise
The project implements an end-to-end NLP pipeline, from text preprocessing and TF-IDF feature engineering to model comparison, hyperparameter tuning, unseen test evaluation, and real-time prediction through a Streamlit web application.
| Metric | Final Test Score |
|---|---|
| Accuracy | 89.25% |
| Macro F1 | 85.11% |
| Weighted F1 | 89.40% |
Final model: Linear SVM C: 0.5 Class weighting: Balanced TF-IDF features: 32,992 Total samples: 19,999
- Built an end-to-end NLP classification pipeline
- Engineered 32,992 TF-IDF features using unigrams and bigrams
- Compared Logistic Regression, Linear SVM, and Multinomial Naive Bayes
- Performed systematic hyperparameter tuning
- Addressed class imbalance using
class_weight="balanced" - Evaluated the final model on a completely unseen test set
- Built a reusable prediction pipeline
- Developed a polished Streamlit web application
- Performed per-class error and confusion-matrix analysis
The system follows this pipeline:
User Text
↓
Text Preprocessing
↓
TF-IDF Vectorization
↓
Class-Balanced Linear SVM
↓
Predicted Emotion
↓
Human-Friendly Result
For example:
Input:
"I am feeling very happy today"
Output:
Joy
| Label | Emotion | Description |
|---|---|---|
| 0 | Sadness | A sense of heaviness, loss, or low mood |
| 1 | Joy | A feeling of happiness, warmth, or excitement |
| 2 | Love | A feeling of affection, warmth, or connection |
| 3 | Anger | A strong feeling of frustration or displeasure |
| 4 | Fear | A feeling of worry, uncertainty, or apprehension |
| 5 | Surprise | A reaction to something unexpected or sudden |
The project uses 19,999 labeled text samples divided into training, validation, and test sets.
| Split | Samples |
|---|---|
| Training | 15,999 |
| Validation | 2,000 |
| Test | 2,000 |
| Total | 19,999 |
The dataset is imbalanced, with some emotions represented significantly more frequently than others.
| Emotion | Samples |
|---|---|
| Joy | 5,361 |
| Sadness | 4,666 |
| Anger | 2,159 |
| Fear | 1,937 |
| Love | 1,304 |
| Surprise | 572 |
This imbalance was considered during model development and addressed in the final Linear SVM using balanced class weights.
The text was transformed into numerical features using:
TfidfVectorizer(
max_df=0.95,
min_df=2,
ngram_range=(1, 2),
sublinear_tf=True
)The resulting representation contains:
32,992 features
Both unigrams and bigrams were used to capture individual words as well as short phrases.
For example, a sentence can generate features such as:
feel
humiliated
didnt
didnt feel
feel humiliated
The resulting TF-IDF matrices are stored as sparse CSR matrices, making them efficient for high-dimensional text classification.
Three machine learning algorithms were evaluated on the validation set.
| Model | Accuracy | Macro F1 | Weighted F1 |
|---|---|---|---|
| Logistic Regression | 87.55% | 84.88% | 87.65% |
| Linear SVM | 90.30% | 87.35% | 90.32% |
| Multinomial Naive Bayes | 62.95% | 35.27% | 54.01% |
Linear SVM produced the strongest overall validation performance and was selected for further tuning.
The Linear SVM was evaluated using different values of C and class-weight strategies.
The selected configuration was:
C = 0.5
class_weight = balanced
Accuracy: 90.35%
Macro F1: 87.73%
Weighted F1: 90.41%
The use of balanced class weights improved the model's ability to account for less-represented emotion categories.
After model selection and hyperparameter tuning, the final model was evaluated on the held-out test set.
| Metric | Score |
|---|---|
| Accuracy | 89.25% |
| Macro F1 | 85.11% |
| Weighted F1 | 89.40% |
| Emotion | Precision | Recall | F1 Score |
|---|---|---|---|
| Sadness | 0.94 | 0.92 | 0.93 |
| Joy | 0.93 | 0.90 | 0.92 |
| Love | 0.72 | 0.85 | 0.78 |
| Anger | 0.88 | 0.90 | 0.89 |
| Fear | 0.88 | 0.85 | 0.87 |
| Surprise | 0.67 | 0.79 | 0.72 |
The model performs particularly well on sadness, joy, and anger.
The lower performance on surprise reflects both its smaller representation in the dataset and the natural linguistic overlap between emotions such as surprise and fear.
The final test-set recall shows where the model performs most reliably:
| Emotion | Recall |
|---|---|
| Sadness | 91.74% |
| Joy | 90.22% |
| Anger | 90.18% |
| Love | 84.91% |
| Fear | 84.82% |
| Surprise | 78.79% |
Common areas of semantic overlap include:
Love ↔ Joy
Fear ↔ Surprise
Sadness ↔ Anger
This demonstrates an important characteristic of emotion classification: emotional categories are not always linguistically independent.
The trained model is integrated into a Streamlit web application.
The interface allows users to:
- Enter natural-language text
- Submit the text for classification
- View the detected emotion
- Read a human-friendly description
- Explore the six supported emotions
- View model performance
- Learn about responsible use
The interface was designed with a clean, dark, human-centered aesthetic so that the prediction remains the primary focus.
text-emotion-detection/
│
├── app/
│ └── app.py
│
├── data/
│ └── ...
│
├── models/
│ ├── final_svm_model.pkl
│ ├── tfidf_vectorizer.pkl
│ └── final_evaluation_results.csv
│
├── src/
│ ├── predict.py
│ └── ...
│
├── notebooks/
│ └── ...
│
├── requirements.txt
├── .gitignore
└── README.md
- Python 3.10+
- Git
git clone https://github.com/YOUR_USERNAME/text-emotion-detection.git
cd text-emotion-detectionReplace YOUR_USERNAME with your GitHub username.
python -m venv .venv
.venv\Scripts\activatepython3 -m venv .venv
source .venv/bin/activatepip install -r requirements.txtstreamlit run app/app.pyThe application will open through the local Streamlit server.
The trained model can also be used directly through the prediction pipeline.
from src.predict import predict_emotion
text = "I am feeling wonderful today"
emotion = predict_emotion(text)
print(emotion)Example output:
joy
The prediction pipeline uses the saved TF-IDF vectorizer and trained Linear SVM model, allowing predictions without retraining.
The trained components are included in the repository:
models/
├── final_svm_model.pkl
├── tfidf_vectorizer.pkl
└── final_evaluation_results.csv
The final class-balanced Linear SVM used for emotion prediction.
The fitted TF-IDF vectorizer containing the learned vocabulary and feature configuration.
The final test-set evaluation results.
- Python
- scikit-learn
- Linear SVM
- Logistic Regression
- Multinomial Naive Bayes
- TF-IDF
- Unigrams
- Bigrams
- Sparse matrix representation
- pandas
- NumPy
- SciPy
- Streamlit
The final TF-IDF representation contains 32,992 dimensions and is highly sparse.
Linear SVM is well suited to this type of high-dimensional text representation because it can efficiently learn linear decision boundaries without requiring the computational cost of nonlinear kernels.
The experimental comparison also showed that Linear SVM provided the strongest validation performance among the evaluated baseline models.
This project is a text classification system, not a psychological assessment system.
The model cannot reliably determine a person's actual emotional or psychological state.
Important limitations include:
- Text may express multiple emotions simultaneously
- Emotional meaning can depend strongly on context
- Sarcasm and irony are difficult to interpret
- Very short text may contain insufficient information
- Similar emotions can share vocabulary
- Model performance depends on the characteristics of the training dataset
- The model may make incorrect predictions on unfamiliar language patterns
Therefore, predictions should be interpreted as machine-learning classifications of text, not definitive statements about a person's feelings.
This project is intended for:
- Educational purposes
- NLP experimentation
- Machine learning demonstrations
- Research and prototyping
It should not be used for:
- Medical diagnosis
- Psychological diagnosis
- Mental-health assessment
- High-stakes decisions about individuals
Human emotion is complex and contextual. A classifier can identify patterns in language, but it cannot directly observe a person's internal emotional state.
The current TF-IDF + Linear SVM approach provides a strong and efficient baseline.
Potential improvements include:
Evaluate transformer-based approaches such as:
- BERT
- RoBERTa
- DistilBERT
- Increase minority-class samples
- Apply targeted data augmentation
- Explore additional emotion datasets
- Improve handling of ambiguous examples
Investigate:
- SHAP
- LIME
- Feature-level explanations
The current model uses Linear SVM decision scores rather than calibrated probabilities.
Future versions could investigate probability calibration to provide more interpretable prediction confidence.
This project demonstrates that a carefully engineered traditional NLP pipeline can achieve strong performance on multi-class emotion classification without requiring a large neural network.
The final system combines:
TF-IDF
+
Class-Balanced Linear SVM
+
Streamlit
to achieve:
89.25% Test Accuracy
85.11% Macro F1
89.40% Weighted F1
More importantly, the project covers the complete machine-learning workflow:
data preparation → feature engineering → model comparison → hyperparameter tuning → evaluation → error analysis → deployment
Sinan PS
Machine Learning · Natural Language Processing · Python
Note: Emotion classification is inherently subjective. A model prediction describes a pattern identified in the supplied text; it does not establish the actual emotional or psychological state of the person who wrote it.