MACHINE LEARNING BASED SEMANTIC ROLE LABELING FOR AFAAN OROMOO TEXTS

Abstract

Semantic Role Labeler (SRL) is the one of the task in Natural Language Processing, which identifies semantic arguments of predicates and labels them with their semantic role in a given sentence. Different researches have been done on semantic role labeling especially for developed languages like English and Chinese. Those researchers have used different algorithms and different corpus as a dataset for executing their algorithms. Well-developed semantic role labeler increases the performance and effectiveness of different natural language processing applications such as machine translation, information extraction, question answering and others. Even though Afaan Oromoo is one of the languages, which is used by a large population, there is no research work done on this semantic role labeling for the language. For the given predicate, there are different participants in a sentence for making this predicate more meaningful. There is also the relationship between the given predicate of the sentence and other arguments. This relationship should be persistent whenever the arrangement of those arguments and predicates is changed. The sequence of predicates and other arguments is understandable by human beings, but it is difficult for machines in order to capture the similarities and differences in a meaning of verbs reflected in the argument. The goal of this study is to do semantic role labeling using a supervised machine learning system that draws inferences from input data without labeled answers. The proposed system has different tasks like data preprocessing, morphological analysis, semantic role annotation, feature extraction, training and classification. To make sure that the developed model is good enough for classification of an arguments of the predicates, basic evaluation metrics such as accuracy, precision, recall, and F-measure have been used in the study. The experiment of the study is done using Support Vector Machine, Decision Tree, and Naïve Bayes classification algorithms. Afaan Oromoo propositional bank (AOPropBank), which has a detail of each arguments of the sentences and their proper role, is developed from 1400 Afaan Oromoo sentences collected from newspapers, social medias, and blogs. The result obtained from our experiment shows that the performance of the developed model using NB, SVM, and DT classification algorithms achieved an accuracy of 75%, 76.63%, and 70.25% respectively. However, in order to increase the obtained results, further research work such as named entity recognition, word sense disambiguation are needed to be conducted