Measuring the Culture of Useful Knowledge and the Roots of Innovation

I develop a method that uses the language surrounding places, people, and other entities in dated text to measure their attributes over time, and use it to ask whether a local orientation toward useful knowledge predicts innovation and growth. Patterns of word use form a numerical map, or word embedding, in which fixed word lists define attribute scales for scoring each entity. Neither scales nor scores use outcomes, and publication dates restrict scores to prior text. Applied to 38.6 billion context-word occurrences in nearly 2.7 million English-language works from 1500-1899, the method recovers independently measured elevation and agricultural dependence and combines learning, science, practical improvement, and print in a useful-knowledge index. Within countries, a one-standard-deviation increase in the index is associated with 0.105 standard deviations higher log(1+patents) during 1980-2014 and a 3.4 percentage-point increase in the probability of any patent. The association remains 0.062 standard deviations in 7,011 cells without a recorded birthplace of a notable intellectual. An index reconstructed from pre-1750 contexts predicts 1980-2014 log patenting within countries at 0.053 standard deviations after controlling for textual coverage, with positive estimates in Europe, Asia, and Latin America. Training the embedding solely on pre-1750 text leaves the full-sample association positive. Within countries, the 1500-1899 index also predicts average nighttime light over 1992-2024 and regional GDP per capita in each region's latest available year, 2016-2020. In European cities, the pre-1750 index predicts faster post-1750 growth near coal, a complementarity concentrated in Britain. The findings support Mokyr's view of useful knowledge as a local innovative capacity and an Allen-Mokyr synthesis for Britain's coal-and-steam take-off.