TRANSFER LEARNING FOR INDONESIAN NAMED ENTITY RECOGNITION

• Main Author: ADITYA KOSASIH (NIM : 13514012), JOSHUA
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/28317
• Institution: Institut Teknologi Bandung

<p align="justify"> <br /> <br /> Transfer learning is used in machine learning to apply the knowledge previously gained from source task to solve the target task. Current Indonesian named entity recognition systems do not have satisfying performance yet. On the other hand, Indonesian part-of-speech tagging systems have achieved quite satisfying result of 0.9262 F-1 score on Universal Dependencies v1.3 corpus. This final project investigated the impact of transfer learning in building Indonesian NER system using feature representation transfer from Indonesian POS tagging model. <br /> <br /> The model for this final project is a bidirectional GRU-CRF sequence labeller. The model can be grouped into two components, feature extraction component and labelling component. The feature extraction component combines the extracted feature from two different data levels, character level and word level. Transfer learning make use of this feature extraction component to do feature representation transfer. <br /> <br /> The optimum POS tagger model built in this final project-which has achieved 0.921 F-1 score on Universal Dependencies v2 corpus-is transferred to build the NER tagger model. The F-1 score was improved by 0.02 points when the model was trained on 100% of data, while with smaller number of training data-1% of the original-the result was improved by 0.15 points. <br /> <p align="justify">