Multi-label emotion classification of Urdu tweets.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Affective Computing

Background:

• Urdu is a major South Asian language with limited resources for computational emotion analysis.
• Existing emotion datasets are predominantly in English, necessitating Urdu-specific resources.
• The complex morphology and syntax of Urdu pose unique challenges for emotion detection.

Purpose of the Study:

• To create the first multi-label emotion dataset for Urdu tweets.
• To develop and evaluate baseline classifiers for multi-label emotion detection in Urdu.
• To explore various text representation techniques for Urdu emotion classification.

Main Methods:

• Dataset creation: 6,043 Urdu tweets annotated with six basic emotions.
• Classification approaches: Machine learning (RF, J48, SMO, AdaBoostM1, Bagging), Deep Learning (1D-CNN, LSTM, LSTM+CNN), and Transformer (BERT).
• Text representations: Stylometric features, pre-trained embeddings, n-grams (word and character-based).

Main Results:

• Evaluation of baseline classifiers using metrics like F1-score (micro and macro), accuracy, Hamming Loss, and Exact Match.
• Comparative analysis of different methodologies for Urdu emotion classification.
• Identification of effective text representations and models for the task.

Conclusions:

• The developed Urdu emotion dataset is a valuable resource for NLP research.
• The study provides insights into the performance of various models on Urdu multi-label emotion detection.
• Further research can build upon these baselines for improved Urdu affective computing.