Sentence level Sentiment Analysis of Afaan Oromoo text-based from social media using Deep Learning

Abstract

The main objective of this research is to develop sentence-level sentiment analysis for Afaan Oromoo text using deep learning models. Finding out how using Emojis with Afaan Oromo text affects the labeling dataset and model's performance is the other task of the research. Sentiment analysis is a field of NLP that classifies the feeling of people into positive and negative sentiments. The data (comments and feedback) that was presented on social media are a list of unorganized and unclassified raw data that made it challenging to accurately and quickly ascertain customer sentiment in real-time. To overcome these problems, this study assessed different techniques from data collection to model building. Data that contain Emojis and text was collected from Facebook public pages of OBN, FBC Afaan Oromo, BBC Afaan Oromo, and VOA Afaan Oromo, as well as the Vision Entertainment and Aanaa Entertainment YouTube channels. And this study, was prepared and preprocessed using various techniques of NLP and proposed CNN, LSTM, and BiLSTM models with word2Vec feature extractions. According to the experiment's findings, using Emojis alongside Afaan Oromo text decreased the performance of the LSTM, BiLSTM, and CNN models from 73.34% to 72.03, 72.88 to 71.88%, and 74.58% to 72.66% respectively with skip-gram feature extractions. This study also evaluates the proposed models on a binary dataset and achieves an accuracy of 89.60% on LSTM, 88.32% on BiLSTM, 87.52% on CNN by using skip-gram, and 87.68% on LSTM, 87.36% on BiLSTM, and 88.64% on CNN with CBOW. Finally, by comparison of all the proposed models, the CNN model outperforms the other two models with an accuracy of 74.58% on the multi-class datasets with skip-gram feature extractions. Therefore, this study chose the CNN model for Afaan Oromo text-based Sentence level sentiment analysis classifications.