Machine Learning Disambiguation of Quechua Verb Morphology

Abstract

We have implemented a rule-based prototype of a Spanish-to-Cuzco Quechua MT system enhanced through the addition of statistical components. The greatest difficulty during the translation process is to generate the correct Quechua verb form in subordinated clauses. The prototype has several rules that decide which verb form should be used in a given context. However, matching the context in order to apply the correct rule depends crucially on the parsing quality of the Spanish input. As the form of the subordinated verb depends heavily on the conjunction in the subordinated Spanish clause and the semantics of the main verb, we extracted this information from two treebanks and trained different classifiers on this data. We tested the best classifier on a set of 4 texts, increasing the correct subordinated verb forms from 80% to 89%.