ABSTRACTIVE AFAAN OROMO TEXT SUMMARIZATION FOR NEWS TEXTS POSTED ON FACEBOOK USING A DEEP LEARNING APPROACH.

Major Advisor: Mr.Kamal Mohammad (Ass.prof) In today’s digital world, one of the languages utilized to write digital news texts on Facebook is Afaan Oromo. Afaan Oromo news texts being posted on Facebook are typically lengthy and unstructured, which causes readers to lose interest. Issues with lengthy text raise the significance of ATS, which can summarize user posts. The majority of the earliest approaches involved in developing Afaan Oromo text summarization were extractive approaches, with only a small number involving an abstractive approach. According to a recent study, the research on abstractive Afaan Oromo text summarization focuses on finding solutions to the issue of producing summaries that are shorter in length than the original content by disregarding the necessity of producing relevant summaries with Afaan Oromo lengthy texts posted on Facebook. In this research, we have used 6327 Afaan Oromo news texts from various verified Facebook pages. We proposed an attention-based Se2Se model with LSTM and bi-directional attentive Se2Se RNN in our study to address issues of Afaan Oromo text summarization for news texts posted on Facebook. As our experimental result shows, training accuracy for the attention-based Se2Se model with LSTM is 80.31%, while validation accuracy is 77.13%. Bi-directional Attentive Se2Se RNN, which is the second model, achieves 80.98% training accuracy and 77.68% validation accuracy. The results indicate that bi-directional attention Se2Se RNN outperformed the attention-based Se2Se model with LSTM, scoring the highest values of 80.98% for training accuracy and 77.68% for validation accuracy. Using unseen data as input, we tested the performance of our model with our test data to evaluate the generated summary using rouge-1, rouge-2 (rouge-n), and rouge-l. The bi-directional attentive Se2Se RNN f-measure score is higher for rouge-1, rouge-2, and rouge-l at 0.13, 0.03, and 0.11, respectively, when compared to the attention-based Se2Se model with LSTM. This indicates that the summary produced by this model is superior to the attention-based Se2Se model with LSTM. Nevertheless, some initiatives may be taken in the field of ATS to enhance the model's performance by incorporating GPT-3.5, T5, and other learning strategies. Keywords: ATS, NLP, LSTM, bi-directional RNN, softmax