Designing Afan Oromo Emotion Detection Model Using Deep Learning Method

Abstract

Social media is one of the most common virtual spaces where people share their opinions, feelings, and emotions in different languages. Many more peoples in Ethiopia, other foreign countries, speak the Afan Oromo language. It is a language of social media, various fictions, and provided literature. The study focused on detecting Afan Oromo emotions from social media text. Emotion is a feeling that represents a person's state of mind, such as happiness, anger, disgust, etc. Automatically identifying and classifying emotions from social media posts can greatly value in the commercial, political campaigns, decision making, and social benefit. Emotion detection has become a central study topic in the field of NLP. The main problem with social media posts is that they are extremely large, fragmented, disorganized, informal, and unstructured. Nevertheless, many organizations and individuals are very interested in knowing what other people think or feel about their services and products. Emotion detection is the process of automatically identifying and categorizing emotions to determine whether a writer's attitude towards a particular entity is happiness, sadness, anger, or fear. We used five deep learning models, namely CNN, LSTM, BiLSTM, CNN-LSTM, and CNN-BiLSTM, to develop emotion detection in Afan Oromo social media content, which is Facebook posts and comments. We conducted our experiments on the 2020 total dataset within six emotion classes (happiness, love, anger, disgust, fear, and sadness). The dataset was collected from Afan Oromo Facebook pages, such as BBC, OMN, OBN, and FBC by using Face pager software tools and after collection, manual annotation is made. Then the preprocessing techniques such as normalization, tokenization, and stop word removal were done. We used word embedding as feature encoding and Keras deep learning Python library was used to implement algorithms. We performed our experiments with the selected classifiers algorithms and for each classifier; we used 90% training dataset and 10% testing dataset. Our experimental result shows that the AOED models achieved an accuracy of 92%, 87%, 88%, 88%, and 90% for the models CNN, LSTM, BiLSTM, CNN-LSTM, and CNN-BiLSTM, respectively. Our proposed CNN model provided significant accuracy in classifying the text from Facebook posts and comments and outperformed. Overall, our study showed more promising results than previous related work for Afan Oromo emotion detection.