📄 PolInterviews — A Dataset of German Politician Public Broadcast Interviews

Summary

In Germany, televised political interviews strongly shape public opinion. Yet researchers have no clean, open transcripts of them, which limits computational analysis of political talk. The authors address this gap with PolInterviews. It is a corpus of German public-broadcast interviews with senior politicians, all collected from YouTube for 2020–2024. The selected politicians are party leaders, secretaries general, and federal and state ministers or premiers. After an initial list was drafted, backward searches found further interviews with the same politicians.

The audio was transcribed with Whisper. Speakers were separated using ECAPA-TDNN segment embeddings and agglomerative clustering. Two research assistants then checked and corrected every transcript by hand. The result is a tidy table with one sentence per row, carrying a speaker ID, a position in the video, and, where kept, a timestamp. Politicians are linked to their Wikidata IDs. The corpus covers two Bundestag terms and supports studies of agenda-setting, self-presentation, argumentation and interviewer–interviewee dynamics. The authors stress that it is useful but not exhaustive.

Key points

• The corpus contains 99 interviews with 33 high-ranking German politicians and 28,146 sentences in total. It is openly released on Harvard Dataverse.
• It draws on five formats: Maischberger (MAISC), Frag Selbst (FRAGS, built from YouTube comment questions), ARD Sommerinterviews (ARDSO), Caren Miosga (CARMI) and assorted shorter ZDF clips (ZDFWE).
• The processing pipeline has three stages:
  • Whisper produces segment-level transcription with timestamps.
  • ECAPA-TDNN embeddings plus agglomerative clustering perform diarization.
  • Two assistants carry out full manual review.
• The table has seven variables: format_id, year, video_id, speaker, text, timestamp and order_id. Politicians are identified by Wikidata QID (e.g. Q61053 for Olaf Scholz), and all interviewers share the code Q999. Some timestamps were lost during manual correction.
• Most videos last 20–30 minutes. Politicians average about 2,360 words per interview, while interviewers average about 1,257. Christian Lindner appears most often, followed by Söder, Merz and Habeck.

Interesting ideas

• I1. Speakers are keyed by Wikidata QIDs rather than free-text names, so the corpus joins directly with other resources such as LegislatoR or ParlSpeech without fuzzy name matching.
• I2. Individuals get unique IDs, while the interchangeable role of interviewer gets one shared code (Q999), separating "who exactly" from "what role".
• I3. An automatic pipeline produces a draft (ASR plus clustering diarization), and humans review every output, trading cost for a fully corrected gold set.
• I4. Each sentence row stores an explicit order_id inside its video, so conversational sequence survives even where timestamps were dropped.
• I5. One-on-one question–answer interviews form naturally paired question and response turns, with roles cleanly labelled.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (world): Mova's entity layer could anchor recognised real-world persons to QIDs, letting the world of a text inherit Wikidata facts and stay comparable across documents.
• I2 (frames): dialogue frames could track a turn's role (asker vs. answerer) separately from the speaker's identity, so speech-act analysis still works when the person is unknown or irrelevant.
• I3 (learn): silver parses agreed by witness parsers could be routed through the same "machine draft, human pass" step to build small fully verified evaluation slices.
• I4 (world): event and turn ordering could rest on an explicit sequence index rather than fragile surface cues, keeping the order of events reconstructable after edits or repairs.
• I5 (answer): such paired turns suggest a test design in which Mova's reader must judge whether a reply actually addresses the question, a harder check than extractive QA.