jbfish00/BYU-LING-Thesis

Repository for my MA Linguistics thesis work in low-resource Machine Translation (primary workspace for code, drafts, and related materials). # Domain-Specific Fine-Tuning for English-to-Georgian Machine Translation of LDS Religious Texts **BYU Linguistics MA Thesis** | Jacob B. Fisher ## Overview This thesis investigates how domain-specific fine-tuning improves machine translation (MT) quality for English-to-Georgian, with a focus on Latter-day Saint (LDS) religious texts. Three models are compared: 1. **Baseline** — NLLB-200-1.3B (Meta's No Language Left Behind, pre-trained on 200+ languages) 2. **General-FT** — NLLB-200-1.3B fine-tuned on ~66,000 general LDS parallel sentences (General Conference talks, church articles) 3. **Scriptural-FT** — NLLB-200-1.3B fine-tuned on ~43,229 scriptural parallel sentences (Bible, Book of Mormon, D&C, Pearl of Great Price) Fine-tuning uses **LoRA** (Low-Rank Adaptation) for parameter-efficient training. Models are evaluated with automatic metrics (BLEU, chrF, xCOMET) and a large-scale **human evaluation** (2,217 error annotations from 3 evaluators across 450 segments). ### Key Findings - Fine-tuned models reduce total errors by **34-40%** on scriptural text - **Register errors** (formal/archaic language mismatch) drop by **92-99%** — from 125 in the baseline to 9 (General-FT) and 1 (Scriptural-FT) - Both fine-tuned models significantly outperform the baseline (p **Note on data availability:** The General Language data (General Conference talks, church articles) cannot be publicly distributed due to copyright restrictions. The Scriptural corpus is included. Model checkpoint weights (`.safetensors`/`.pt` files, 2+ GB each) are excluded via `.gitignore`. --- ## Data Pipeline ``` Web sources (bible.com, churchofjesuschrist.org) │ ▼ scrape_georgian_bible.py scrape_georgian_book_of_mormon.py scrape_georgian_dc_pgp.py scrape_general_conference_2025.py Raw HTML → per-scripture CSV files │ ▼ combine_standard_works_parallel_data.py align_scripture_sentences.py combine_general_church_data.py merge_conference_data_2012_2025.py split_conference_talks_by_paragraph.py …