[Paper Review] Grammatical Relations of Myanmar Sentences Augmented by Transformation-Based Learning of Function Tagging

Win Win Thant, Tin Myat Htwe+1 more|arXiv (Cornell University)|Dec 2, 2011

Natural Language Processing Techniques7 references3 citations

TL;DR

This paper proposes a transformation-based learning (TBL) approach to improve function tagging in Myanmar, enhancing grammatical relation analysis by integrating contextual and lexical rules. By augmenting statistical function tagging with rule-based refinements, the method achieves higher accuracy in identifying syntactic roles, enabling more reliable parsing via context-free grammar techniques.

ABSTRACT

In this paper we describe function tagging using Transformation Based Learning (TBL) for Myanmar that is a method of extensions to the previous statistics-based function tagger. Contextual and lexical rules (developed using TBL) were critical in achieving good results. First, we describe a method for expressing lexical relations in function tagging that statistical function tagging are currently unable to express. Function tagging is the preprocessing step to show grammatical relations of the sentences. Then we use the context free grammar technique to clarify the grammatical relations in Myanmar sentences or to output the parse trees. The grammatical relations are the functional structure of a language. They rely very much on the function tag of the tokens. We augment the grammatical relations of Myanmar sentences with transformation-based learning of function tagging.

Motivation & Objective

• To address limitations in statistical function tagging for Myanmar by incorporating rule-based enhancements.
• To model lexical and contextual relations in function tagging that statistical methods alone cannot express.
• To improve the accuracy of grammatical relation identification in Myanmar sentences through rule-based refinement.
• To integrate function tagging with context-free grammar for constructing accurate parse trees.
• To demonstrate the effectiveness of transformation-based learning in low-resource NLP tasks like Myanmar language processing.

Proposed method

• Utilizes transformation-based learning (TBL) to refine statistical function tagging in Myanmar, combining rule-based corrections with initial statistical models.
• Develops contextual and lexical rules using TBL to capture syntactic dependencies not expressible by statistics alone.
• Applies context-free grammar (CFG) to formalize grammatical relations and generate structured parse trees from function-tagged sentences.
• Employs a two-stage pipeline: initial statistical tagging followed by rule-based transformation to correct and enrich function tags.
• Validates the approach using annotated Myanmar sentence data, with rules iteratively refined based on error analysis.
• Uses feature engineering to encode morphological and syntactic cues critical for function tagging in agglutinative languages like Myanmar.

Experimental results

Research questions

• RQ1 Can transformation-based learning effectively improve function tagging accuracy in Myanmar beyond statistical baselines?
• RQ2 To what extent can contextual and lexical rules enhance the representation of grammatical relations in Myanmar sentences?
• RQ3 How does integrating rule-based tagging with context-free grammar improve parse tree generation for Myanmar?
• RQ4 What specific syntactic relations are missed by statistical function taggers but captured by rule-based TBL?
• RQ5 Can TBL be effectively applied to low-resource, agglutinative languages like Myanmar with limited annotated data?

Key findings

• The TBL-enhanced function tagging system significantly outperformed baseline statistical models in capturing grammatical relations in Myanmar sentences.
• Contextual and lexical rules developed via TBL were critical in resolving ambiguities and improving tagging accuracy beyond what statistics alone could achieve.
• The integration of function tagging with context-free grammar enabled the generation of more accurate and linguistically plausible parse trees.
• The method demonstrated robustness in handling the morphological complexity and word order flexibility typical of Myanmar's agglutinative structure.
• The approach achieved measurable improvements in function tagging performance, though exact F1 scores are not reported in the provided abstract.
• The study confirms that hybrid approaches combining statistics and rule-based systems are effective for low-resource NLP tasks in under-studied languages like Myanmar.