Evaluating LSTM Networks, HMM and WFST in Malay Part-of-Speech Tagging

Abstract

Long short term memory (LSTM) networks have been gaining popularity in modeling sequential data such as phoneme recognition, speech translation, language modeling, speech synthesis, chatbot-like dialog systems and others. This paper investigates the attention-based encoder-decoder LSTM networks in Malay part-of-speech (POS) tagging when it is compared to weighted finite state transducer (WFST) and hidden Markov model (HMM). The attractiveness of LSTM networks is its strength in modeling long distance dependencies. Malay POS tagging is examined from two different conditions: with and without morphological information. The experiment results show that LSTM networks that are trained without any explicit morphological knowledge perform nearly equally with WFST but better than HMM approach that is trained with morphological information.