Diacritic Restoration and the Development of a Part-of-Speech Tagset for the Māori Language

Abstract

This thesis investigates two fundamental problems in natural language processing: diacritic restoration and part-of-speech tagging. Over the past three decades, statistical approaches to diacritic restoration and part-of-speech tagging have grown in interest as a consequence of the increasing availability of manually annotated training data in major languages such as English and French. However, these approaches are not practical for most minority languages, where appropriate training data is either non-existent or not publically available. Furthermore, before developing a part-of-speech tagging system, a suitable tagset is required for that language. In this thesis, we make the following contributions to bridge this gap:

Firstly, we propose a method for diacritic restoration based on naive Bayes classifiers that act at word-level. Classifications are based on a rich set of features, extracted automatically from training data in the form of diacritically marked text. This method requires no additional resources, which makes it language independent. The algorithm was evaluated on one language, namely Māori, and an accuracy exceeding 99% was observed.

Secondly, we present our work on creating one of the necessary resources for the development of a part-of-speech tagging system in Māori, that of a suitable tagset. The tagset described was developed in accordance with the EAGLES guidelines for morphosyntactic annotation of corpora, and was the result of in-depth analysis of the Māori grammar.