Comparative Analysis of Deep Learning Models for Aspect Level Amharic News Sentiment Analysis

Abstract

Nowadays, in the era of social media, customer input and feedback have significant impact on firm’s services and products. Companies face a significant challenge in extracting meaningful information from this unstructured, unorganized, massive, and fragmented data. Some research works has been done on Amharic sentiment analysis (AMSA), however none of them have looked at the aspect level by utilizing a deep learning approach. This work focuses on sentiment analysis of Amharic text utilizing aspect level with a hybrid deep learning approach. The dataset was acquired from Amhara Media Corporation's official Facebook page in Microsoft Excel format. Comment exporter software was used to create a dataset of 10,000 in excel format. Different machine learning techniques such as Convolutional neural network (CNN), Long short-term Memory (LSTM), CNN-LSTM and CNN-GRU were used to train and test the dataset. The result of the study shows that LSTM model performed better than other models with training accuracy of 99.10% having a very little difference from CNN-GRU model with 99.08% training accuracy.