On Empirical Evaluation of Deep Architectures for Indonesian POS Tagging Problem

Abstract

Models with deep architectures have been state-of-the-arts technologies in many natural language problems such as text classification, name entity recognition, language models, and Part-of-Speech (POS) tagging. Usually, the models are trained with large number of data to produce satisfactory results. In Indonesian POS tagging problems, we must deal with small number of data. In this paper, we evaluate models with deep architectures for Indonesian POS tagging problems to find the best structures for Indonesian POS tagging. Models with various number of hidden layers are investigated. We also investigate the effect of adding a regularization method such as dropout on the performance. The experimental results show that the model with 2 hidden layers shows to have better accuracy than the models with deeper structures.