Areas of work — Juan S. Gómez-Cruces

Natural language processing

Text classification and measurement

Large language models make it easier to analyse text, but their results still need to be checked. I define what a measure is intended to capture, test it against human-coded examples, and document the process so the analysis can be repeated.

Classification of open-ended survey responses, interviews, documents and online text
Zero-shot and fine-tuned pipelines, benchmarked against hand-coded samples
Multilingual corpora, including Spanish and Portuguese sources
Containerised tools and documentation for in-house use
Review of an existing automated measure before publication

Opinion & behaviour

Surveys, experiments and causal inference

I study attitudes and behaviour and estimate whether programmes, campaigns or policies have had an effect. Where possible, I specify the design in advance and report the uncertainty around the results.

Survey and experiment design, questionnaire review and fielding support
Causal inference on programme, campaign and policy effects
Measurement of attitudes, trust and approval over time, including forecasting
Secondary analysis of existing survey series

Platforms & society

Digital environments and institutions

I examine how online platforms shape information, participation and governance, and help organisations assess what the available evidence can and cannot tell them.

Country and regional assessments of digital governance and internet freedom
Monitoring designs for information environments and coordinated behaviour
Collection of platform and API data within legal and ethical constraints
Briefings and expert input for non-specialist audiences

Forms of collaboration.

Commissioned research

A defined study

A written methodology, delivered with the data, the analysis code and a report.

Advisory

Retained or per day

Design review, participation in expert groups, and assessment of a study before fieldwork begins.

Software & training

Built to hand over

Containerised applications for in-house use, and workshops on NLP and large language models.

Partnerships

Consortium participation

Co-applicant or work-package lead in Horizon Europe, DFG, foundation and donor-funded projects.

Methodological commitments.

Construct before model

A measure is only as informative as the construct it operationalises. The concept is specified first, and the method is chosen to fit it.

Validation against human coding

Automated measures are benchmarked against hand-coded data and reported with agreement statistics, including the cases and languages in which performance is weaker.

Reproducibility

Version control, containers and documented workflows, so that a result can be reproduced when a reviewer, an auditor or a colleague asks how a figure was obtained.

Areas of work.