Shaaban, Y., Korashy, H., Medhat, W. (2022). Arabic emotion cause extraction using deep learning. The Egyptian Journal of Language Engineering, 9(2), 23-39. doi: 10.21608/ejle.2022.150056.1032
Yasmin Shaaban; Hoda Korashy; Walaa Medhat. "Arabic emotion cause extraction using deep learning". The Egyptian Journal of Language Engineering, 9, 2, 2022, 23-39. doi: 10.21608/ejle.2022.150056.1032
Shaaban, Y., Korashy, H., Medhat, W. (2022). 'Arabic emotion cause extraction using deep learning', The Egyptian Journal of Language Engineering, 9(2), pp. 23-39. doi: 10.21608/ejle.2022.150056.1032
Shaaban, Y., Korashy, H., Medhat, W. Arabic emotion cause extraction using deep learning. The Egyptian Journal of Language Engineering, 2022; 9(2): 23-39. doi: 10.21608/ejle.2022.150056.1032
Arabic emotion cause extraction using deep learning
1Computer Engineering, Faculty of Engineering, Ain Shams University ,Cairo , Egypt
2Faculty of Engineering, Ain Shams University
3Department of Information technology and computer science/ FCAI, Nile University/ Benha University, GIZA/BENHA, EGYPT
Abstract
Emotion cause extraction is a challenging task nowadays. Causes behind emotions are extracted from textual data. Emotion cause extraction has many applications such as extracting causes from reviews that are extracted from social networks and recommender websites where users give their feedback. The resources in this field are limited. There are some corpora built for western languages like English and far east languages like Chinese. Arabic language resources in this field are very limited. This paper introduces emotion cause detection in Arabic Language. A dialectal Arabic annotated corpus is built for the purpose of emotion cause extraction. The data collected from many resources. Sequence labelling techniques are applied with IOB2 scheme using BiLSTM-CRF algorithm and BERT-CRF algorithm. BERT-CRF outperforms BiLSTM-CRF in both span-level and token-level measure evaluation. BERT-CRF achieves a 0.29 F1 score in case of span-level measure evaluation and a 0.84 F1 score in case of token-level measure evaluation.
View on SCiNiTO
Top