SEMANTIC ROLE LABELING FOR AFAAN OROMO TEXT USING DEEPLEARNING

Major Advisor: Mr.Kamal Mohammad (Ass.prof) Natural Language Processing (NLP) is a subfield of artificial intelligence that investigates how people and computers interact. This research focuses on the development of Afaan Oromo semantic role labeling, which is widely used in information extraction, etc. It aims to classify and predict tokens of a given corpus into predefined semantic role labeling classes like predicate, manner, agent, patient, theme, beneficiary, locative, instrument, temporary, and others. Previously conducted research in this area made use of machine learning. But recent SRL studies using deep learning have shown improved performance over the feature-based ones for many languages. Therefore, we hypothesize that the performance of SRL systems for Afaan Oromo text can also be improved by using deep neural networks. We collected 4527 Afaan Oromo simple sentences for our study, labeled them with their semantic roles, and prepared the Afaan Oromo proposition bank (AOProBank). We annotated our corpus according to the IOB format, which indicates the beginning, middle, and other input values of the constituent (argument) of the sentence. In this work, artificial recurrent neural networks (RNNs) with memory, including long short-term memory (LSTM) and BiLSTM, are used to model Afaan Oromo semantic role labeling. In order to memorize longer input data sequences, the LSTM-based models have additional gates. By forwarding and backward traversing the input data, bidirectional LSTMs (BiLSTM) offer additional training. The findings demonstrate that BiLSTM-based modeling, which is based on additional training of data, provides better performance than regular LSTM-based modeling. The model achieved the following performances: precision, recall, and f1-scores of 86%, 86%, and 85.9% for the LSTM model whereas the BiLSTM model scored 88%, 91%, and 89%, respectively. It was discovered that the BiLSTM model offers better performance than the LSTM model. There are no expert-based semantic roles previously annotated in the sentences utilized in this investigation. Future research can test this study using semantic roles that have been expertly annotated. Keywords: Afaan Oromo semantic role labeling. BiLSTM, LSTM, NLP, RNN, softmax,regularization