Accurate occupational classification is essential for educational and sociological research, yet manual coding of free-text survey responses remains expensive, inconsistent, and difficult to scale. This challenge is magnified in multilingual settings where inputs may combine several languages, contain severe spelling errors, or reflect limited respondent knowledge. We present an agentic retrieval-augmented generation (RAG) pipeline for the automatic assignment of International Standard Classification of Occupations (ISCO) codes to free-text parental occupation descriptions collected through Luxembourg’s Épreuves standardisées school monitoring program. The pipeline integrates two domain-specific knowledge stores – an enriched ISCO store and a Luxembourg-specific glossary – and delegates retrieval decisions to a large language model, allowing it to dynamically query the most relevant sources for each input. Applied to a dataset of 24,222 entries spanning up to nine languages, the pipeline produces results broadly consistent with Luxembourg’s occupational distribution as measured by the SILC survey. On a sample of 500 entries independently coded by a human expert, the pipeline achieved 79.80% accuracy at the 4-digit ISCO level with a mean absolute deviation (MAD) in ISEI scores of 1.34 points, compared to 54.20% accuracy and a MAD of 4.12 points for the human coder. We further introduce two novel quality indicators – classification certainty and respondent seriousness – and establish empirically motivated thresholds for their use in downstream analyses. To our knowledge, this is the first study to apply agentic RAG to occupational classification in a multilingual minority-language setting, and the first to address noise introduced by children as proxy respondents in large-scale survey data.
Frania et al. (Mon,) studied this question.