Unsupervised Dependency Parsing for Myanmar Language

Abstract

Parsing natural language is an important intermediate step for natural language processing field of any language and any natural language applications such as machine translation, information extraction, text analytics, and speech recognition systems. Parsing is examining the structure of sentence in terms of relationships between phrases or words of sentence and can be carried out by syntax, or constituency, or phrasal parsing and dependency parsing. Dependency structure is simpler and better than syntax or phrase structure to represent language semantic and syntactic information. Dependency parsing provides directed links of the connection of linguistic unit (words) in sentence. Dependency structures and parsing have been more applied in natural language applications such as machine translation, and provide better performance results. Motivated research areas of unsupervised dependency parsing from raw sentence without requiring any annotated resources have achieved a big improvement in fifteen years ago and some resources and annotated treebanks of some languages have been shared to improve multilingual parsing purposes. As a result, unsupervised dependency parsing becomes a probable way to obtain dependency information of low or under resource languages and more applied. Myanmar language has free word order nature, many styles for sentence writing, and no resource for dependency information. Therefore, it is still cost- and time-consuming, and difficult to add manually dependency structures of Myanmar words. According to these issues, this dissertation is the first proposed work for dependency parsing based on transition-based dependency parsing method that uses transition predictions of neural network classifier for Myanmar language. An adaptable Myanmar POS tag scheme which is related to Universal part-of-speech (UPOS) tags and dependencies has been also firstly defined and proposed to apply unsupervised dependency parsing. Myanmar dependency treebank has been annotated to build Myanmar parsing model to parse Myanmar sentences. Evaluation experiments of the new Myanmar parsing model have been executed. The proposed dependency parsing method has parsed well new Myanmar test sentences. Accuracies scores of experiments and evaluations of parsing performance are measured by undirected attachment score (UAS) and label attachment score (LAS). Most UAS and LAS result scores of parsing experiments and evaluations are over 89% and 84% in general. Accuracies scores and result trees are acceptable.