AFAAN OROMO MULTI LABEL NEWS TEXT CLASSIFICATION USING DEEP LEARNING APPROACH

Abstract

The development of the internet has made Afaan Oromo's writings widely available both offline and online. Automatic text classification is an invertible response to the continuously expanding amount of information resources available. The practice of classifying a text or document into predetermined categories is called text categorization, or text classification (TC). It is recommended that this study employ the deep learning technique with word embedding on Afaan Oromo's multi label news text classifications. Since extracting feature values from news articles is difficult, this research suggests a deep learning strategy for news text categorization. The word order information, which is essential for classifying news texts, was not taken into consideration by the earlier approaches, which classified the text data by using a bag of words to represent the words in the text. Despite their low time complexity, the earlier models do not adequately account for the context and any semantic links between text words. Furthermore, as the number of characteristics and classes increased, the models' accuracy decreased. This thesis uses CNN, LSTM, BRNN, GRU and algorithms to perform deep learning approaches for Afaan Oromo multilabel news text categorization. To create a pre-trained word embedding model, utilize the Afaan Oromo news domain. Then, train our data using the CNN, LSTM, BRNN and GRU model create the classification process to recommend the best for the problem at hand. In this study, the models for the Afaan Oromo languages were constructed using newly gathered and annotated news datasets totaling nine thousand forty one (9041). After completing the following procedures, the Afaan Oromo News text documents are classified using those deep learning algorithms. Preprocessing, word embedding, deep network construction, output determination, model training, and classification. The semantics of the page are finished with word2vec, which uses neural network architecture to map similar words into a single vector. Consequently, the vector representations of words serve as the input for the deep network development component. Training, testing, and validation datasets are used to evaluate the model using accuracy and loss; finally, the performance of our models is compared. LSTM performs exceptionally well, scoring 98.71% accuracy and 98.71% precision, while CNN, BRNN and GRU score 94%, 94%, 96.40%, 97.48%, and 94.42%,92.4% respectively.