Khmer POS Tagging Using Conditional Random Fields

• Main Authors: Sokunsatya Sangvat, Charnyote Pluempitiwiriyawej
• Other Authors: Mahidol University
• Format: Conference or Workshop Item
• Published: 2019
• Subjects: Computer Science; Mathematics
• Online Access: https://repository.li.mahidol.ac.th/handle/123456789/45669

© 2018, Springer Nature Singapore Pte Ltd. The transformation-based approach with hybrid of rule-based and tri-gram have already been introduced for Khmer part-of-speech (POS) tagging. In this study, in order to further explore this topic, we present an alternative approach to Khmer POS tagging using Conditional Random Fields (CRFs). Since the features greatly affect the tagging accuracy, we investigate five groups of features and use them with the CRF model. First, we study different contextual information and use it as our baseline model. We then analyze the characteristics of Khmer and come up with three additional groups of language-related features including morphemes, word-shapes and name-entities. We also explore the use of lexicon as features to further improve the accuracy of our tagger. Our proposed approach has been evaluated on a corpus of 41,058 words and 27 POS tags. The comparative study has shown that our proposed approach produces a competitive accuracy compared to other Khmer POS tagging approaches.