COREFERENCE RESOLUTION FOR INDONESIAN TEXT WITH MENTION PAIR METHOD USING DEEP LEARNING

• Main Author: Auliarachman, Turfa
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/39717
• Institution: Institut Teknologi Bandung

There are only a few number studies on coreference resolution system for Indonesian text. In fact, there have been no studies using deep learning. Recent studies have shown that deep learning is effective in coreference resolution system. A coreference resolution system study that uses convolutional neural networks (Wu and Ma, 2017) proposes the use of a singleton classifier component before the system makes predictions for each markable pair. However, in this study (Wu and Ma, 2017) no experiments were conducted on the effectiveness of using singleton classifier component in improving coreference resolution system performance. In this study, a series of experiments were conducted to see whether the use of deep learning in the coreference resolution system for Indonesian text could outperform the baseline performance which received an F1 average of 44.97% and 60.22%. In addition, this experiment was also conducted to see the effect of using the singleton classifier on the overall performance of the coreference resolution system. The experiment was conducted using data used in the research in one of the baselines consisting of 759 training data and 108 test data. The coreference resolution system in this study outperformed the baseline by obtaining the highest F1 average of 67.37%. From a total of 18 coreference resolution systems made, the use of the trained singleton classifier component with the best weighted-F1 score (93%) improved the performance of 15 systems, with the highest F1 average of 63.27%. The remaining three systems are systems that already have good performance. The final score for all systems increases when the golden singleton classifier was used, with the highest F1 average of 80.00%.