A POS Tagging Model for Vietnamese Social Media Text Using BiLSTM-CRF with Rich Features

This paper deals with the task of part-of-speech (POS) tagging for Vietnamese social media text, which poses several challenges compared with tagging for conventional text. We introduce a POS tagging model that takes advantages of deep learning and manually engineered features to overcome the challenges of the task. The main part of the model consists of several bidirectional long short-term memory (BiLSTM) layers that are used to learn intermediate representations of sentences from features extracted at both the character and the word levels. Conditional random field (CRF) is then used on top of those BiL-STM layers, at the inference layer, to predict the most suitable POS tags. We leverage various types of manually engineered features in addition to automatically learned features to capture the characteristics of Vietnamese social media data and therefore improve the performance of the model. Experimental results on a public POS tagging corpus for Vietnamese social media text show that our model outperforms previous work [4] by a large margin, reaching 91.9% accuracy with 27% error rate reduction. The results also reveal the effectiveness of using both automatically learned and manually designed features in a deep learning framework when only a limited amount of training data is available.