Exploring Deep Learning for Amharic Sentiment Analysis: In the Context of Political Discourse in Ethiopia
No Thumbnail Available
Date
2023-06
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
Deep learning, a potent machine learning technique, learns several layers of
representations or characteristics of the data to produce cutting-edge prediction outputs.
In addition to its popularity in a variety of fields of application, deep learning has been
success fully used in sentiment analysis on a large scale in recent years. Nowadays, social
media plays a significant role in swaying public opinion in favor of or against a
government or organization. Therefore, an effective strategy is necessary to analyze the
mood of any social media posting.
This research paper ex pl ores the application of deep learning that automatically extracts
sentiments from Amharic text, Sentiment analysis for Amharic in the political domain.
Social media platforms have become crucial for shaping public opinion and sentiment
toward governments and organizations. A lot of research is being conducted for the
development of sentiment analysis in different languages. However, sentiment analysis
for under-resourced languages like Amharic has received little attention.
The study proposes a supervised deep-learning for sentiment classification and
demonstrates its effectiveness in analyzing Amharic language data. Data was extracted
from FANA broadcasting corporation and Ethiopian Prime Minister ABlY Ahmed Ali's
official Face book page and manually annotated into positive and negative classes. And
this study examines the effectiveness of using a different training dataset and pre-trained
word embedding and tokenizers experiment on purposes and improving the text
categorization effectiveness of deep learning models.
The study compares the performance of five different deep learning models, including
LSTM, CNN, RNN, GRU, and bi-LSTM, with various architectures and parameters. The
experiments reveal that the GRU model achieved an accuracy of 82.49%, the LSTM
model achieved the highest accuracy of 79.4 1 % while the CNN model attained an
accuracy of 77% the bi-LSTM model obtained an accuracy of 74.2% and the R !N
model achieved an accuracy of 73.3%. These results suggest that using a large training
dataset, pre-trained embedding, pre-trained tokenize, and Adams optimizer can
significantly increase text categorization performance using deep learning models.
Description
Keywords
Sentiment analysis, Amharic language, deep learning, polities