تصنيف المشاعر في نصوص وسائل التواصل الاجتماعي باللهجة العربية باستخدام تقنيات التعلم العميق ومعالجة اللغة الطبيعية

Emotion Classification in Dialectal Arabic Social Media Texts Using Deep Learning and Natural Language Processing Techniques

Abstract

Due to the popularity and consensus on social media, emotion classification in textual data has received considerable attention and has been used in various applications such as sentiment analysis, public opinion mining, and mental health monitoring.

Arabic, and specifically dialectal Arabic, such as Iraqi Arabic, has its own set of unique challenges when it comes to language processing due to morphological complexity, the availability of labeled datasets, and the informal nature of many Arabic-language written forms on social media.

BERT’s strength is extracting context from the text by contextual representation of words, while LSTM capturing long-term dependencies of text sequences which helps in classification, this work proposes a hybrid deep learning approach based on BERT and LSTM to improve emotion classification in Arabic social media texts, specifically in Iraqi dialect.

Three pre-classified datasets were used for emotions, and the datasets used differed in dialects. The Iraqi dialect dataset was chosen, which is Iraqi Arabic Emotion Corpus (IAEC), the Saudi dialect dataset is Arabic Pan-Arab Emotion Dataset (ArPanEmo), and the Egyptian dialect dataset is Arabic Emotions Twitter Dataset (AETD). To remedy dialectal differences and data scarcity, various text preprocessing techniques such as normalization, emoji translation, and synonym-based data augmentation are implemented.

Based on the results we obtained when using BERT in classification, the accuracy rate was 81%, and when using the proposed hybrid model, the accuracy became 84%.

The results show the contribution of these summarized methods of how sequential and contextual learning could be combined to achieve the highest accuracy among the previous Arabic sentiment analysis models. the study presents credible contribution to Arabic natural language processing (NLP) as it proposes a scalable and efficient method for the detection of emotion in dialectal text.