Machine Learning
The group builds the pipelines that make data out of sources that were never data. The chain runs from page layout classification through table segmentation to the recognition of the handwritten entry itself, applied to historical censuses, parish records and nurse records at a scale hand transcription cannot reach.
The resulting resources are released publicly rather than kept in-house. HANA is a database of handwritten names, DARE recognises handwritten dates, and OccCANINE maps occupational descriptions to standard historical codes across languages. A separate strand applies machine learning inside estimation rather than before it, in causal machine learning and in generativemodels for time series.
Key research areas
| Handwritten text recognition for tabular sources Layout classification, table segmentation and recognition of the handwritten entries, built for the tabular records that historical economics depends on. |
| Public databases and models HANA for handwritten names, DARE for handwritten dates, and OccCANINE for occupational descriptions across languages. All three are publicly available. |
| Machine learning inside estimation Causal machine learning for effect heterogeneity, and generative adversarial networks for resampling a series where only one realisation exists. |
Who works on this
| Christian Møller Dahl Professor |
Applies machine learning to digitised historical records, and works on time series econometrics, quantile regression and causal inference. |
| Torben Skov Dyg Johansen Assistant Professor |
Works on machine learning for digitising historical records and on causal methods for allocating an intervention when capacity is limited. |
| Riccardo Di Francesco Postdoctoral Researcher |
Develops machine learning methods for estimating and interpreting heterogeneous causal effects, with applied work on worker health. |
Grants in this area
| Project | Funder | Period | Grant recipient |
| Moralization and Polarization Builds a long series of state-of-the-state speeches by American governors, from each state's entry into the Union to the present, and uses text analysis to date the point at which an issue acquires moral language. It then tests whether moralisation runs ahead of measured polarisation. |
Carlsberg Foundation | 2021- | Christian Møller Dahl 282,000 DKK |
| A More Equal School: How the Abolishment of Half-Time Schooling in Rural Primary Schools in Sweden Affected Life Chances and Inequality Uses the staggered abolition of half-time schooling in rural Swedish primary schools in the early twentieth century to estimate what additional classroom time was worth. Digitised birth records linked to censuses follow the affected cohorts into later life and into the next generation. |
Vetenskapsrådet | Christian Møller Dahl 362,559 DKK |