The Impact of Genre on Nasalization and Shortened Forms in Haitian Creole: Exploratory Work on Natural Language Processing for Low-Resource Languages

Abstract

Large Language Models (LLMs) are becoming increasingly prevalent in today's world, especially as it relates to Natural Language Processing (NLP) tasks, such as translation. Still, Artificial Intelligence (AI) models such as LLMs face a "second language problem", suffering from reduced performance on low-resource languages due to being pretrained primarily on English data. In order to close that gap, this project trains POS tagger and lemmatizer models from Stanza on Haitian Creole, one such low-resource language. It uses a baseline corpus of educational text and news excerpts, as well as an augmented corpus with additional literary material, to examine the impact of genre on written Haitian Creole -- specifically focusing on nasalization and shortened forms. The main findings suggest a large effect of genre on both linguistic phenomena, as well as the benefits of additional training data in increasing the accuracy of both part of speech tagging and lemmatizing tools regardless of genre. On a linguistic level, several questions are left open as to what about an author or culture of written language makes specific genres more nasalized or shortened. However, this project concretely shows the efficacy of a wider variety of training data for the development of Haitian Creole NLP capabilities, and establishes the merit of low-resource languages within the computational linguistic space.