The Buried Etymology

Two computational linguistics papers demonstrate that language preserves history invisibly — cultural evolution encoded in patterns that speakers don't notice but algorithms can detect.

Jafari et al. (arXiv: 2604.01467) build a dynamic atlas of Persian poetic symbolism by mapping symbolic networks across 129,451 poems organized by Hijri century. Poetic symbols — wine, gardens, courtly imagery, sacred references — function as dynamic systems whose connections strengthen, weaken, and rewire over centuries. Wine vessels and garden imagery intensified in later periods. Courtly vocabulary weighted toward earlier centuries. Cross-connections between symbolic families weakened over time, suggesting cultural compartmentalization as traditions aged.

Rao (arXiv: 2604.01425) shows that random forest classifiers, given only word embedding features, can distinguish whether Hindi synonyms originate from Sanskrit or Perso-Arabic — even when the words are semantically indistinguishable. Usage patterns preserve etymological traces that are invisible to speakers but present in distributional statistics. The words mean the same thing. They appear in different contexts. That contextual difference encodes centuries of cultural contact between Sanskritic and Persianate literary traditions.

The structural claim: language is a recording medium that preserves cultural history in distributional patterns, not in the words themselves. Persian poems encode the evolution of symbolic culture across centuries — not in the content of individual symbols but in the network of connections between them. Hindi preserves the history of Sanskrit and Perso-Arabic contact — not in word meanings but in the statistical contexts where synonyms appear.

Neither pattern is accessible through traditional reading. A Persian literature scholar recognizes that garden imagery is important — but the quantitative dynamics of how garden symbolism connected to other symbolic families, and how those connections rewired century by century, requires computation at scale. A Hindi speaker might sense that two synonyms have slightly different registers — but the precise statistical signature that separates Sanskrit-origin from Perso-Arabic-origin words is invisible to human intuition.

Jafari et al.'s network analysis reveals something specific: symbolic families didn't just change frequency — they changed connectivity. Early Persian poetry had dense cross-connections between symbolic domains (wine linked to mysticism linked to courtly imagery). Later poetry separated these domains, each operating more independently. This is cultural compartmentalization made visible — a process that happened over centuries without anyone deciding it, recorded in the aggregate statistics of how poets chose their metaphors.

Rao's finding is even more striking in its implications. Two words that mean the same thing, that can be substituted for each other in any sentence without changing the meaning, that no speaker would distinguish semantically — these words carry different distributional signatures because of events that happened centuries ago. The Mughal court's Persianate culture and the Sanskrit literary tradition each left fingerprints in how their vocabulary gets used, and those fingerprints persist into modern Hindi despite the words themselves being treated as interchangeable.

Both papers suggest that language is a deeper archive than it appears. The surface level — what words mean, what poems say — is accessible to human readers. The distributional level — how words co-occur, how symbols connect, how contexts cluster — is an archive of cultural history that only becomes readable at computational scale.

The question this raises for any system that processes language: what historical and cultural information is your model absorbing from distributional patterns, and how is it affecting downstream behavior? If Hindi word embeddings carry etymological traces from centuries-old cultural contact, what other invisible historical patterns are encoded in the corpora that train modern language models?