Debate Metrics is an experimental tool for LLM-assisted statistical analysis of speeches in the German Bundestag.
It scores observable, topic-independent communication patterns in parliamentary debate text, including:
- discourse quality, such as argument clarity, reasoning structure, and engagement with counterarguments
- rhetorical behaviour, such as escalation/de-escalation signals, clarity, polemics, and cooperation signals
The public website is debatemetrics.pax77.org.
Debate Metrics processes parliamentary speech text and produces structured scores for selected communication patterns. The goal is to make rhetorical and discourse-related features of debate text more comparable across parties and LLM providers.
The analysis focuses on how something is argued, not whether the statement is true, persuasive, morally justified, or politically effective.
Debate Metrics does not:
- fact-check political claims
- determine whether a speaker is right or wrong
- measure political competence or democratic legitimacy
- endorse or rank parties, speakers, or positions
- replace close reading, historical context, or political analysis
Scores should be interpreted as empirical indicators derived from a defined annotation and modelling process, not as objective judgments about political value.
Parliamentary debates contain rich information about political communication, but large-scale comparison is difficult when analysis depends only on manual reading.
Debate Metrics explores whether LLM-assisted annotation can help produce reproducible, topic-independent measurements of discourse and rhetorical behaviour while preserving transparency about limitations and uncertainty.
Debate Metrics may be useful for:
- comparing changes in rhetorical style over time
- studying escalation and de-escalation patterns in parliamentary debate
- supporting exploratory political communication research
- generating structured data for further statistical analysis
The project is experimental. The scoring system is designed to capture observable textual features rather than hidden intent, factual accuracy, or audience impact.
Because LLM-assisted analysis can be sensitive to prompt design, model choice, text segmentation, and coding definitions, results should be treated as provisional and interpreted with care.
Where possible, the project aims to make prompts, scoring categories, processing steps, and validation logic inspectable.
The project analyzes speeches from the German Bundestag.
Depending on the repository structure, this section should explain:
- where the source data comes from
- whether raw data is included in the repository
- how speeches are segmented
- what metadata is used, such as party, date, agenda item, or legislative period
- any licensing or attribution requirements for the source material
- Processed text, metadata, annotations, and scores generated by this project are derived from the official Bundestag source material but are not official Bundestag publications. Any preprocessing, segmentation, normalization, annotation, or scoring is performed by Debate Metrics and may introduce errors.
See CONTRIBUTING.md for local setup, workflow, and pull request guidance.
Debate Metrics is licensed under the GNU General Public License v3.0 only. See LICENSE.