EmotionClassification

Pet project in evaluating different models (both deep learning and traditional) for the NLP task of classifying emotions

Dataset used:

The dataset used for this project is an emotion classification dataset from HuggingFace, containing twitter messages classified into 6 emotions:

• anger
• fear
• joy
• love
• sadness
• surprise

Preprocessing used:

• Word Counts
• TFIDF
• Resampling to correct for imbalanced data (not used for every)

Models:

Model Used	Training Accuracy	Validation Accuracy
Naive Bayes (resampling + TFIDF)	0.93	0.80
Logistic Regression (with spacy embeddings)	0.335	0.35
Simple Decision Trees (word_counts)	0.998	0.844
Gradient Boosted Trees (word_counts)	0.997	0.834
LSTMs	---	0.8335

LSTM Details

Neural Network Architecture

$$ Embedding (64) \rightarrow LSTM (32) \rightarrow Linear (6) \rightarrow LogSoftMax $$

Training History

Loss history	Accuracy history