Hybrid BERT-LSTM Model for Emotion Detection in Iraqi Arabic Text

Emotion detection in natural language processing is crucial for understanding human communication through sentiment analysis. This paper is motivated by these concerns and discusses the difficulties pertinent to emotion detection in Arabic text, concentrating on emotions expressed in the Iraqi dialect as those are not fully addressed yet. This work proposes a hybrid model involving pre-trained bidirectional encoder representations from transformers and long short-term memory networks to detect five specific emotions (anger, sadness, happiness, fear, and surprise). 1,365 Facebook posts have been collected from the Face Emotion dataset that represents the original test set and split into train/test for the classifier. The main preprocessing stage was text cleaning, emoji translation, and data augmentation via synonym replacement, which succeeded in improving the overall results. The results showed that the suggested model dominated the benchmarks with a global F1 score of 83%. More significantly, the results further demonstrate that emotion recognition in context is necessary and indicate that transfer learning between different language dialects can be realized. Among other findings, this study contributes to the broader field of emotion detection, particularly in under-resourced languages. Also, this study provides a base for further research regarding ARABIK (Arabic NLP applications).