Year: 2026
Type: dissertation
Abstract: This dissertation explores the uses of Large Language Models (LLMs) for enhancing Low-Resource Language (LRL) data. LLMs are models trained with substantial datasets to perform language tasks such as translation, text generation, information retrieval, and summarization, among others. The primary applications of LLMs for LRLs explored here are data denoising and morphological generation. Six LRLs are used throughout this research: K'iche', Seri, Kaqchikel, Huave, Uspanteko, and Nahuatl. These languages are Mesoamerican languages spoken in Mexico and Guatemala. They are used as examples of possibilities for LRL research using LLMs; however, the findings are meant to be applicable to other LRLs as well. Noisy data refers to data that contains errors or inconsistencies. For language data, data denoising is a process where a dataset is standardized if there is an existing standard, else it is normalized- made consistent through manual or computational methods. This dissertation tests computational methods of denoising through fine-tuning and prompting pre-trained LLMs. For denoising with LLMs, two approaches are used. The first approach uses a pretrained text-to-text LLM fine-tuned with limited data for each language. The second approach uses the instruction-tuned LLM GPT-4o. This LLMs approach explores various techniques of prompt engineering (the writing of prompts to elicit certain output) and demonstrates SOTA denoising with relatively small unlabeled datasets using a method termed "unlabeled corpus analysis." Morphological generation is the process of generating the inflected forms of a given word stem. This experiment attempts morphological generation for K’iche' with the instruction-tuned model, GPT-4o, through prompt engineering. K'iche' is an agglutinative language. Since written resources are scarce, not all inflected word forms can be gathered from texts. Morphological generation could theoretically be used to combat data scarcity in LRLs. This experiment in morphological generation builds on the prompting techniques used in the instruction-based denoising task and demonstrates the limitations of GPT-4o in performing some NLP tasks for LRLs. Data denoising and morphological generation are both important tasks for enhancing LRL data. Data denoising improves the quality of language data, while morphological generation improves the quantity. Both have the potential to make a dataset more conducive to linguistic research and practical applications.
Author: Wing, Cheyenne
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7162479607