The comparative linguistic hypothesis formulated by Dr. Sungtae Son posits that ancient East Asian populations—specifically from Northeast Asian polities such as Gojoseon, Buyeo, Goguryeo, and Balhae—undertook trans-Pacific migrations between the 3rd and 10th centuries CE. Traversing the Aleutian Islands, these groups settled across the Americas, leaving structural and toponymic footprints in indigenous language families like Algonquian, Uto-Aztecan, Siouan, and Na-Dene.

While manual studies uncovered structural and phonetic overlaps—such as matching Nahuatl and Algonquian terms to ancient Korean roots—the approach relies on manual text analysis. Manual methods suffer from cognitive selection bias, struggle with diachronic consistency across sound shifts, and lack mathematical frameworks to rule out random coincidence. Modern computational linguistics resolves these limits through an automated four-stage AI mining pipeline.

Pipeline Architecture and Workflow

1. High-Volume Data Ingestion and Preprocessing: The ingestion engine queries the USGS Geographic Names Information System (GNIS)—the federal standard for geographic nomenclature under ANSI INCITS 446-2008—containing over two million records. The pipeline isolates over 50,000 native place and tribe names across feature classes like Stream, Spring, Populated Place, and Landform. Crucially, it extracts historic Variant_Name attributes to recover indigenous pronunciations obscured by European colonial spellings. Named Entity Recognition (NER) models filter out post-1492 Euro-American names.

    

2. Diachronic Phonological Mapping Engine: To resolve transcription biases and historical sound drift, toponyms are converted into International Phonetic Alphabet (IPA) vectors via Grapheme-to-Phoneme (G2P) conversion (Epitran) and PanPhon articulatory features. The system reverses historical sound shifts to match forms against Old Korean (3rd–10th century CE) and Middle Korean (10th–16th century CE):

    

   • Stop Contrast Expansion: Reconstructs European single stop transcriptions ($b/p, d/t, g/k$) back to Korean lenis, fortis, or aspirated stops.

      

   • Pre-Palatalization Reversal: Reverts modern affricates (/dʒ, tʃ/) before front vowels back to Old Korean alveolar stops (/d, t/), restoring forms like dip (house) behind modern jip.

      

   • Cluster Preservation & Coda Behavior: Enforces Middle Korean consonant cluster rules (e.g., golgi $\rightarrow$ modern gori) and models European vowel epenthesis or final stop omissions.

      

3. Algorithmic Alignment and Confidence Scoring: Rather than standard Levenshtein edit distance, the pipeline employs a modified ALINE algorithm based on articulatory feature distance. A Composite Confidence Score ($C$) is calculated as:

    
   $$C = w_1 \cdot S_{\text{phon}} + w_2 \cdot S_{\text{sem}} + w_3 \cdot S_{\text{morph}} – P_{\text{anach}}$$

   Where $S_{\text{phon}}$ measures feature-weighted phonetic similarity, $S_{\text{sem}}$ evaluates semantic-topographic class concordance (e.g., matching a GNIS Stream feature to archaic Korean ne/ni for river), $S_{\text{morph}}$ tracks agglutinative morphological markers, and $P_{\text{anach}}$ penalizes post-migration linguistic anachronisms. High confidence matches include terms like Tipi (dip + -i), Moccasin (mog-hwa-sin), and Aztlan (asadal).

    

4. Etymological Disambiguation and Statistical Validation: Candidates with $C \ge 0.75$ are audited against comparative indigenous dictionaries (e.g., Proto-Algonquian, Proto-Uto-Aztecan) to downgrade candidates fully explained by internal family developments. Universal onomatopoeia and post-1492 European loanwords are filtered out. Finally, Monte Carlo permutation testing shuffles toponym arrays against randomized East Asian lexicons across 100,000 iterations, validating candidate clusters only if alignment density exceeds chance at $p < 0.01$.