Emotion Classification of Afaan Oromo Texts Using Deep Learning

Abstract

Emotion during communication is always expressed as joy, sadness, anger, surprise, hate, fear and so on. Despite the fact that textual communication is a more popular mode of communication, the fast adoption of social media has elevated it to a new level. Now a day people can easily convey their emotions through social media. Emotions can be expressed in a variety of ways, including voice, texts, and nonverbal communication such as facial expressions and gestures. Afaan Oromo is a native Afro-Asiatic language that is widely spoken throughout Ethiopia as well as nearby horn of Africa nations like Kenya, Djibouti, and Somalia. On social media, many users post, comment, and tweet about many subjects using Afaan Oromo language. Emotion classification has been studied for resource rich languages like English. However, the dataset and models available cannot be applied for Afaan Oromo due to its different syntax and scripting style. Previously emotion classifications for Afaan Oromo were made by one study. However the study has limitations in terms of dataset, text representation, and emotion category considerations. In this study, the researcher developed Afaan Oromo text emotion classification model using deep learning. The researcher collected Afaan Oromo dataset using web scraping techniques from Facebook and YouTube channels. We then annotate the collected Afaan Oromo dataset by four Afaan Oromo teachers. The dataset were labeled for seven emotion classes of anger, happy, disgust, sadness, fear, surprise and neutral. Then, annotated text is preprocessed for further manipulations. For text representation, the researcher used word2vec word embedding trained on 400k Afaan Oromo sentences. In the study a classifier based on different deep learning algorithm like GRU, LSTM, Bi-LSTM and a combination of CNN and Bi-LSTM is trained to classify emotions. The result of the experiment shows that the combination of CNN and Bi-LSTM algorithm has shown the best performance as compared to other algorithms. The combination of CNN and Bi-LSTM scored an accuracy of 88% and F-score of 83% on the test set