Multilabel Emotion Recognition in Urdu Using Attention-Based Hybrid Model and Fractional Order Backpropagation

Emotion detection in text has been widely researched for high-resource languages, yet low-resource languages such as Urdu remain underexplored. The present work addresses this gap by developing a multilabel emotion detection framework for Urdu using a fractional-order backpropagation learning approach. The work employs two complementary sentence embedding techniques—one supervised (paraphrase-multilingual) and one unsupervised (transformer-based sequential denoising autoencoder)—in combination with long short-term memory (LSTM) and logistic regression classifiers. Experiments are conducted on the publicly available Urdu Emothreat dataset released as a part of 2022 Forum for Information Retrieval (FIRE), containing 7800 labeled sentences across seven emotion categories. The proposed fractional backpropagation method achieves a consistent 4% average improvement in F1-score over conventional integer-order backpropagation, highlighting its potential for more stable and generalized learning in low-resource settings. The best-performing model, based on supervised LSTM embeddings, attains a weighted precision of 66%, recall of 85%, and F1-score of 74%. These findings establish a strong baseline for Urdu emotion recognition and demonstrate the practical significance of fractional-order optimization for multilingual and low-resource NLP tasks.