How to Automatically Spot Which Groups Politicians Talk About

Why This Matters

Political actors routinely invoke social groups—like workers, immigrants, or small businesses—to persuade, signal representation, and mobilize support. Yet reliably identifying which groups politicians mention in speeches, manifestos, or social media is difficult because group references depend on context and phrasing. Hauke Licht and Ronja Scezpanski address this measurement challenge by building a scalable, context-aware tool for detecting group mentions in political texts.

What the Paper Does

The authors develop a supervised, text-as-data method that flags the exact word spans in a text that refer to social groups. Rather than relying on static dictionaries or simple keyword searches, the approach uses human annotations to mark passages that contain group references and then trains a contextual language model to perform word-level classification.

How the Method Works

• Human annotators first label passages in political texts that refer to social groups, producing gold-standard training data.
• A pre-trained transformer language model is fine-tuned for contextualized, supervised classification at the token (word) level so it can identify the start and end of group mentions within sentences.
• Once trained, the model can be applied to unlabeled corpora to extract precise group-mention spans automatically.

Validation and Illustration

Licht and Scezpanski validate the procedure by applying it to political texts and then demonstrate two substantive applications focused on British party rhetoric. These applications show how the method can reveal which social groups parties invoke and how those invocations are distributed across different communications—insights that are difficult to obtain with coarser keyword approaches.

What This Enables

The technique makes it feasible to map references to social groups across large, heterogeneous text collections (speeches, manifestos, social media, press releases) and to do so with contextual sensitivity—capturing subtle or multiword group mentions and excluding non-group uses of the same words. That opens new possibilities for comparative analyses of party messaging, representation claims, and rhetorical targeting.

Broader Research Value

By turning group-mention detection into an automated, trainable task, this contribution supplies political scientists with a replicable measurement tool that complements existing text-as-data methods and supports finer-grained studies of political rhetoric and appeals to identity and interest groups.