SEMANTIC RELATION EXTRACTION FOR AFAAN OROMO TEXT USING DEEP LEARNING APPROACH

Main Advisor: Mr. Wakgari Dibaba (Ass.Prof ) In the discipline of Natural Language Processing (NLP), Relation extraction is a crucial semantic processing problem. The aim is to determine the semantic relationship between e1 and e2, given a phrase S and two annotated entities, entity1 and entity2, using a set of preset relation types. Numerous applications, including text mining, question answering, information extraction, and more, can benefit from semantic relation extraction. To obtain high-level features, certain state of the art foreign language systems still rely on lexical resources of NLP tools like dependency parsers and named entity recognitions. The fact that crucial information can be anywhere throughout the sentence presents another difficulty. To challenge these problems, we suggest the Afaan Oromo semantic relation extraction scheme, which uses a deep learning methodology. As one of the modern deep learning methods, the (Bi-LSTM) bidirectional long short-term memory network with attention mechanism is used. It extracts significant semantic information from sentences and allows for multi-level automatic feature representation learning from data. The suggested model consists of multiple mechanisms. The first is word embedding, which is the mapping of words into low-dimensional vectors. By acquiring new features across domains, this feature learning technique is utilized to extract relations from text written in Afaan Oromo. It's possible that not all contextual information will show up in one way of the interaction. The third attention mechanism builds a weight vector, which is then multiplied to generate a sentence-level feature vector by combining word-level data from every time step. For this thesis, we create the Afaan Oromo Relation Extraction Dataset from the Afaan Oromo literature, and we test our model by using to change hyper parameters and classifying data set into 80/20 to get the best performance. The assessment techniques of precision, recall, and F-score are commonly utilized to assess the efficiency of the proposed system, with 93.23% accuracy and an F1-score of 78.00%. It generates the proposed bidirectional long short-term memory model. It obtains good results utilizing only word embedding as input features, without the need for lexical resources or natural language processing (NLP) systems. Keywords: Deep learning, Word Embedding, Bi directional long short term memory. Semantic Relation Extraction