Machine learning identifies 438 ancient substrate words in Sulawesi languages that predate Austronesian origins

Researchers analyzed 1,357 basic vocabulary forms from six Sulawesi Austronesian languages to detect non-inherited words

XGBoost classifier achieved 0.763 AUC by analyzing 26 phonological features to distinguish inherited from non-mainstream vocabulary

Identified substrate forms show distinctive phonological patterns: longer length, more consonant clusters, higher glottal stops, and fewer Austronesian prefixes

26.5% of analyzed vocabulary (438 forms) identified as potential candidates for pre-Austronesian substrate origin through computational methods