semantic relation classification, cnn, oversampling, binary relevance

• Main Author: Insan Al-Amin, Muhammad
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/40466
• Institution: Institut Teknologi Bandung

Relation classification (RC) in the music domain is used to identify the relationship between two named entities (NE) automatically. The relations are used to update the music knowledge database which is then used in various applications such as related music information, music search engines, music recommendations, automatic generation of music news, and music encyclopedias. The purpose of this thesis is to produce RC models for Indonesian language music news. This thesis also produced an SRC Indonesian language corpus for the music domain. The hypothesis of this research is, making RC models using CNN will performs better than the baseline method, Support Vector Machine (SVM) and shallow learning when applied to Indonesian music news. A Convolutional Neural Network (CNN) architecture is used to train RC models using word vector representations as input. Oversampling methods such as SMOTE and word perturbation are used for handling unbalanced data. Multilabel problems are solved using the Binary Relevance method. This research only addresses relations between two named entities in a single sentence. Two forms of RC classification modeled in this research are single-label, and multilabel classification. NER modeling is not part of this research.