>_ Jorge Uc / projects← All projects
~/research/nahuatl-nmt // MASTER'S THESIS
✦ Cum Laude distinctionNatural Language Processing2026

Giving historical language a digital voice.

Neural machine translation from Classical Nahuatl to English using NLLB-200, historical parallel data, and parameter-efficient fine-tuning.

Cover artwork for the Classical Nahuatl translation research
RESEARCH FILE / NMT-001NAHUATL → ENGLISH
ALIGNED CORPUS~40Ksentence pairs
BASE MODEL1.3BNLLB-200
APPROACHES03training strategies
HUMAN EVAL.3.45/5LoRA overall mean
01 / RESEARCH CONTEXT

Why this problem matters

Classical Nahuatl is a low-resource historical language. Translating it requires addressing issues that are less visible in high-resource languages.

01LANGUAGE

Complex morphology

Words can encode grammatical information through multiple affixes, complicating segmentation and translation.

MORPHOLOGY / TOKENIZATION
02DATA

Limited parallel text

Historical bilingual documents are scarce, making high-quality alignment and careful evaluation essential.

OCR / ALIGNMENT
03HISTORY

Historical spelling

Variation in orthography and the cultural context of colonial-era texts make automated translation challenging.

ORTHOGRAPHIC VARIATION
02 / CORPUS ENGINEERING

From manuscript to training data

The Florentine Codex provided the historical material for constructing an aligned dataset and a held-out test split.

> corpus.build_pipeline()DATA PREPARATION
01

Historical source

Florentine Codex material

02

OCR & cleaning

Extract and clean text

03

Sentence alignment

Create bilingual pairs

04

Tokenization

Extended NLLB vocabulary

05

Model training

Fine-tuning and evaluation

Approximately 40,000 aligned pairs • Approximately 10% reserved for testing
PUBLIC DATASET

FlorentineCodex / Hugging Face

Research corpus with a persistent DOI and citation metadata.

DOI: 10.57967/hf/11168 ↗
03 / MODEL DESIGN

One model. Three adaptation strategies.

NLLB-200 1.3B is an encoder–decoder Transformer. The research retained its tokenizer and extended the vocabulary for Classical Nahuatl.

> translation.architectureCONCEPTUAL FLOW
INPUTClassical NahuatlHistorical sentence
→
PREPROCESSTokenizer + embeddingsExtended vocabulary
→
MODELNLLB-200Encoder → decoder
→
OUTPUTEnglish textGenerated translation
01 PEFT

LoRA

Learnable low-rank updates to selected attention projections, limiting the number of trainable parameters.

Strongest chrF and TER
02 PARTIAL UPDATING

Layer Freezing

Keep selected layers fixed while fine-tuning other parameters of the pretrained network.

Comparison strategy
03 LAYER SAMPLING

LISA

Sample active layers during optimization to explore an alternative efficient training process.

Highest BLEU
04 / QUANTITATIVE EVIDENCE

What the automatic metrics show

Different metrics capture different aspects of translation quality. Higher BLEU and chrF are better; lower TER is better.

> evaluation.metricsTEST SET
StrategyBLEU ↑chrF ↑TER ↓
LoRA BALANCED3.027.076.0
Layer Freezing2.421.078.0
LISA3.824.080.0

Scores from this research experiment; these values do not imply production-ready translation quality.

KEY FINDINGThere is no single winner across all metrics.

LISA leads in BLEU, while LoRA achieves better chrF and TER, making LoRA the stronger balanced choice in these experiments.

BLEU / LISA
3.8
chrF / LoRA
27.0
TER / LoRA
76.0
Bars illustrate separate metrics; their lengths are not directly comparable.
05 / HUMAN-CENTERED EVALUATION

Beyond automated scores

A paired, blinded evaluation assessed generated translations on three distinct dimensions, each rated from 1 to 5.

> human_eval.summaryLIKERT / 1–5
LISA3.35 / 5Overall mean
Layer Freezing2.85 / 5Overall mean
Semantic adequacyAccuracy of meaning
FluencyReadability in English
Cultural fidelityHistorical and cultural terminology
EVALUATION PROTOCOL

Designed for a fair comparison

  1. Paired samplesThe same sentences were assessed across systems.
  2. Blinded orderingNeutral system labels with randomized order.
  3. Three dimensionsMeaning, fluency, and cultural fidelity.

Means are descriptive; they do not establish statistical significance. LISA scored slightly higher in semantic adequacy, while LoRA led in fluency and cultural fidelity.

> human_eval.comparisonHUMAN SCORES / 1–5

Translation quality, by dimension

Average human scores across the three evaluation criteria.

LoRALISALayer Freezing
3.45
3.55
2.80
Semantic adequacy
3.60
3.40
3.10
Fluency
3.30
3.10
2.65
Cultural fidelity
Descriptive means on a 1–5 Likert scale. Higher is better. The chart is rendered in HTML and CSS, not loaded from an image.
06 / QUALITATIVE OUTPUT

Translation examples

Four screenshots showing the project's translation examples. Select any image to expand it.

Translation example 1, desktop view
01 / Desktop example
Translation example 2, mobile view
02 / Mobile example
Translation example 3, desktop view
03 / Desktop example
Translation example 4, mobile view
04 / Mobile example
07 / RESEARCH REFLECTION

What I learned from this project

Beyond model scores, this research changed how I approach data preparation, experimentation, and the evaluation of language technologies.

RESEARCH TAKEAWAY / 2026

Building a machine translation system for a historical, low-resource language taught me that the quality of the data and the evaluation process matters as much as the choice of neural architecture. Each stage of the pipeline introduces assumptions that must be examined rather than taken for granted.

01 / DATA ENGINEERING

Good models begin with reliable data

Working with historical texts strengthened my skills in OCR cleanup, sentence alignment, corpus quality control, and reproducible dataset preparation. Small errors at the beginning can propagate through the entire translation pipeline.

OCRData cleaningAlignment
02 / LANGUAGE REPRESENTATION

Tokenization is a modeling decision

Extending the NLLB-200 vocabulary showed me how the representation of words and subwords affects a model's ability to process a morphologically rich language. Tokenization and embeddings cannot be treated as incidental preprocessing.

TokenizationEmbeddingsMorphology
03 / EXPERIMENTAL DESIGN

Efficiency requires trade-offs

Comparing LoRA, Layer Freezing, and LISA helped me reason about trainable parameters, computational constraints, and systematic comparisons. The most efficient strategy is not necessarily the best under every metric.

PEFTTransformersExperiments
04 / EVALUATION

A single metric never tells the whole story

BLEU, chrF, and TER produced different rankings. Designing a blinded human assessment reinforced the need to consider meaning, fluency, and cultural fidelity—especially when working with historically significant language.

BLEU / chrF / TERHuman evaluation
THE BIGGER LESSON

Research is about understanding limitations, not just reporting improvements.

This project made me more careful about what experimental evidence can—and cannot—support. The resulting models are research prototypes, and the open corpus creates a foundation for further investigation rather than a claim that the translation problem is solved.

07 / PUBLICATION & REPRODUCIBILITY

Research resources

Access the published research dataset or copy its citation for future reference.

THESIS / NOT LINKED

Full thesis document

The full PDF is not linked on this portfolio page. It can be added after publication or sharing is confirmed.

PDF link pending
DATASET CITATIONUc Torres, J. A. (2026). FlorentineCodex (Revision 9e3e0fb) [Dataset]. Hugging Face. https://doi.org/10.57967/hf/11168
Expanded translation example