Abstract:
Semantic role labeling is a sentence-level semantic analysis of texts. It is a task of assigning role labels for arguments of predicates in sentences to answer mainly who did what to whom. It is widely used in developing many Natural Language Processing applications such as Question Answering, Information Extraction, and Machine Translation.
Several studies in this field have been done for different languages. Different languages have different behavior and structure with each other. The existing study on Amharic semantic role labeling focused on simple sentences that contain one verbal predicate. The sentences had also clearly mentioned subject and direct object. In this research, we have developed a semantic role labeler for Amharic sentences containing multiple predicates or that are compound sentences. The dataset was collected from Addis Admass Newspaper, Fana broadcasting corporate, and Walta media and communication corporate websites. In this study, the variant of Recurrent Neural Network algorithm called Long Short-Term Memory is used along with Conditional Random Field. The proposed model achieved a promising result.
viii