Bryan Anthony Hong’s research while affiliated with University of the Philippines Manila and other places

Filipino is a changing language that poses several challenges. Our goal is to develop a bidirectional English-Filipino Machine Translation (MT) system using a hybrid approach to learn rules from examples. The first phase was an English to Filipino MT system that required several language resources. The problem lies on its dependency over the annotated grammar which is currently unavailable for Filipino, which makes reverse translation impossible. Phase 2 addresses this limitation by using information taken from English and Filipino POS Taggers. The seed rules are generated by aligning the POS tags from English and Filipino examples, including their constraints. To perform compositionality, the system deduces the constituent labels by using the longest adjacent POS tags found in both the English and Filipino rule. The system groups together similar rules and generalizes it to encompass a wider range of unseen examples.

Machine translation (MT) has been used to address inherent problems from human translators. However, the quality of machine translations are usually unacceptable. Researches have focused on improving quality by incorporating machine learning for translation. An example of which is TWiRL which translates English to Filipino sentences. However, TWiRL's approach presented a strict requirement of a grammar that limits the system's flexibility and feasibility. This architecture will not be applicable for reverse Filipino to English translation. To address the problems of TWiRL, REAL Translation provides an alternative approach in translation by removing the grammar requirement. Instead, POS taggers and Semantic Analyzers takes its place to deduce rules from bilingual corpus.