Amharic event text classification from social media using hybrid deep learning

Abstrak

This study aims to develop a hybrid deep-learning model for detecting and classifying Amharic text. Various natural language applications, such as information extraction, event extraction, conversation, text summarization, and require an automatic event classification. However, existing studies focused on classification, giving little attention to the preprocessing and feature extraction techniques. To address this problem, this work proposed a hybridized deep learning-based Amharic social media text event classification model. The model consists of word-to-vector (Word2vecv) word embedding techniques to capture the semantic and syntactic representation. Convolutional neural network (CNN) is used to extract short-length text features. Additionally, bidirectional long-short memory (Bi-LSTM) is used to extract features from long Amharic sentences and classify those events based on their classes. The dataset used for training and testing consists of 6,740 labeled Amharic text sentences, collected from social media. The result shows an accuracy of 94.8% in detecting and classifying Amharic text events.