AMHARIC STANCE CLASSIFICATION USING DEEP LEARNING

GIRMA, BEWKETU MOLLA

dc.contributor.author	GIRMA, BEWKETU MOLLA
dc.date.accessioned	2022-03-18T06:24:57Z
dc.date.available	2022-03-18T06:24:57Z
dc.date.issued	2021-10
dc.identifier.uri	http://ir.bdu.edu.et/handle/123456789/13206
dc.description.abstract	These days the increasing numbers of social media are enabling individuals to express their respective ideas or views about different topics. People now can express their stance towards any topics by commenting or tweeting on social media like Facebook, YouTube and Twitter. Stance classification is a task to automatically determine whether the owner of a text is in support, against, or neutral towards a topic or target. Stance classification has been researched for high-resource languages like English. However, existing datasets and models for high-resource languages cannot be applied for the Amharic due to variations in context, morphology and character representation. As far as our knowledge is concerned, there is no research done on Amharic stance classification. Stance classification needs a defined target or topic, to assess the overall attitude toward the target or topic. In this study we use the approach of multi-task learning to build Amharic stance classification model. Our model jointly learns sentiment classification, target identification and stance classification tasks at the same time. We collect Amharic corpus from social media by employing web-scraping. To prepare our dataset, we filter the collected corpus using keywords to extract comments or tweets that more describe four selected targets or topics (Abiy Ahmed, Green Legacy, Tigray War, and New Currency) of our study. For data annotation we build an android stance annotator application that has cloud-based data storage. Using the application, we annotate the filtered dataset with five annotators. For text representation, we build a Word2Vec embedding model using the collected corpus. To build our model we use a combination of the CNN and Bi-LSTM algorithm. We employ the CNN for text feature extraction and stacked Bi-LSTM followed by fully connected layers as a classifier. Our model uses the sentiment and target information as auxiliary tasks. We see that using sentiment information as auxiliary task can improve performance of Amharic stance classification. From our experiment we observe that Amharic sentiment and stance do not always align for the same comment or tweet. We experiment with different deep learning algorithms for their performance on Amharic stance classification. Our multi-task model shows an improved performance as compared to single task deep learning models that needs multiple independent models for each task. Our model achieves F1-score of 86% for task 1 (target identification), 65% for task 2 (stance classification) and 50% for task 3 (sentiment classification). Keywords: Amharic sentiment and stance classification, Amharic stance target identification, multi-task learning	en_US
dc.language.iso	en_US	en_US
dc.subject	computer science	en_US
dc.title	AMHARIC STANCE CLASSIFICATION USING DEEP LEARNING	en_US
dc.type	Thesis	en_US