Anna Vissens

The Guardian’s Data Science and Data Projects teams, in collaboration with University College London, developed an in-house machine learning model to measure linguistic sentiment towards immigration in debates in the UK House of Commons. The model is designed to distinguish sentiment directed specifically at immigration from more general emotionally charged political language.
The project began with a set of core editorial questions: how can we assess whether parliamentary discourse on immigration has shifted over time? Can we meaningfully compare sentiment in recent debates with speeches from the 1950s and 1960s when immigration was a highly contested political issue? And can such changes be measured in a statistically robust and transparent way, suitable for journalistic scrutiny?
Over two years of collaboration between journalists, data scientists, and academic experts, we developed and validated a bespoke sentiment model capable of answering these questions. Alongside this, we experimented with using Large Language Models (LLMs) to support the labelling of training data, critically evaluating whether they could improve the efficiency and consistency of annotation without undermining analytical reliability.
Attendees will take away:
-A practical framework for measuring targeted sentiment in political texts, rather than relying on generic tone analysis.
-Hard-learnt lessons from translating open-ended editorial questions into robust machine learning approaches.
-Insights into the strengths and limitations of using LLMs for data annotation in sensitive, high-stakes domains.
-Guidance on applying these methods to long-running political or civic datasets to support accountability reporting.