Abstract
In this thesis, we describe the development of several natural language processing tools and resources for the Andean language Cuzco Quechua as part of the SQUOIA project at the University of Zurich.
The main focus of this work lies on the implementation of a machine translation system for the language pair Spanish-Cuzco Quechua. Since the target language Quechua is not only a non-mainstream language in the field of computational linguistics, but also typologically quite different from the source language Spanish, several rather unusual problems became evident, and we had to find solutions in order to deal with them. Therefore, the first part of this thesis presents monolingual tools and resources that are not directly related to machine translation, but are nevertheless indispensable.
The main contributions of this thesis are as follows:
- We built a hybrid machine translation system that can translate Spanish text into Cuzco Quechua. The core system is a classical rule-based transfer engine, however, several statistical modules are included for tasks that cannot be resolved reliably with rules.
- We implemented a text normalization pipeline that automatically rewrites Quechua texts in different orthographies or dialects to the official Peruvian standard orthography. This includes a tool for the morphological analysis of Quechua words that achieves high coverage. Furthermore, we also created a slightly adapted version that can be used as spell checker back-end, in combination with a plug-in for the open-source productivity suite LibreOffice/OpenOffice.
- We built a Quechua dependency treebank of about 2000 annotated sentences, that provided not only training data for some of the translation modules, but also served as a source of verification, since it allows to observe the distribution of certain syntactic and morphological structures. Furthermore, we trained a statistical parser on the treebank and thus have now a complete pipeline to morphologically analyze, disambiguate and then parse Quechua texts.
Apart from the scientific interest in developing tools and applications for a language that is typologically distant from the main stream languages in computational linguistics, we hope that the various resources presented in this thesis will be useful not only for language learners and linguists, but also to Quechua speakers who want to use modern technology in their native language.