SEMANTIC ROLE LABELING FOR GENERATING TEMPLATE OF INDONESIAN NEWS SENTENCES

• Main Author: Edria Devina, Irene
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/40102
• Institution: Institut Teknologi Bandung

News templates have been widely used in automatic news generation. Indrayani and Khodra (2018) has generated Indonesian news automatically with manually defined templates. However, automated generation of Indonesian news sentences template had never been done before. Semantic Role Labeling (SRL) is a process of giving label to a word or phrase in a sentence according to its semantic role based on the sentence’s predicate. In this final project, a system for automatic template generation using SRL will be built. A dataset for Indonesian semantic role labeling does not exist publicly, so collecting Indonesian sentences and manually labeling them based on PropBank (Martha and Palmer, 2005) will be done. The SRL model is built based on the research by He et al. (2017). This final project will use BiLSTM with highway connection, dropout, recurrent dropout, and hard constraints to get the best configuration. The result from SRL will be saved as a template and the difference between automatically generated template and manually generated template will be analyzed. The best configuration, 2 layer BiLSTM with dropout, recurrent dropout, and hard constraints, managed to get F1 score of 0.92 for token level and 0.84 for sentence level. The automatically generated templates also possess similar quality with manually generated templates.