APSI (Automated Political Stance Identification) is a theory-driven text analysis tool that estimates political positions expressed in text using Natural Language Inference (NLI).

APSI analyzes text across three dimensions: Economic Left–Right ideology, Support for Liberal Democratic values, and Populist vs. Pluralist rhetoric.

APSI does not train a supervised classifier on labeled stance datasets. Instead, it uses a hypothesis-based scoring framework with a pretrained political NLI model, enabling zero-shot and interpretable results.

APSI uses the Political DEBATE model (Burnham et al.), which is specialized for political discourse analysis.

It is built on a multilingual transformer base (mDeBERTa-v3 NLI) and then fine-tuned on political NLI data.

NLI is a method where a model estimates whether a text supports (entails) a hypothesis statement, rather than relying on keywords alone.

Zero-shot means APSI can estimate political stances without training on a labeled stance dataset for your specific task.

APSI runs a topic relevance pre-check. If the text does not strongly match political topic hypotheses (threshold 0.6), it returns no score.

The tool returns score = NA and an explanation that it cannot infer a political stance from the text.

Hypotheses are derived from political theory and expert survey codebooks, then written as natural-language statements suitable for NLI scoring.

APSI averages entailment probabilities for each side of a dimension, takes the difference, and scales it to a 0–10 score (clipped to stay within bounds).

Lower values indicate more pluralist rhetoric; higher values indicate more populist rhetoric. The tool maps score ranges to labels like “Pluralist”, “Moderate”, “Populist”, etc.

Lower scores indicate more economically left-leaning language; higher scores indicate more economically right-leaning language.

Lower scores indicate weaker support for liberal democratic values; higher scores indicate stronger support for liberal democratic values.

Confidence reflects internal consistency of the model’s hypothesis probabilities (e.g., lower variance means higher consistency).

Contradiction is flagged when both sides of a dimension receive comparatively strong support, suggesting mixed or conflicting signals in the text.

APSI outputs a score, interpretation label, confidence level, contradiction status, and indicators of which hypotheses influenced the score.

You can analyze speeches, articles, social media posts, policy statements, interviews, and transcripts—any text where political language might appear.

No. APSI analyzes only the text provided and does not know the author, audience, or circumstances surrounding the text.

APSI measures rhetorical framing in language, not verified beliefs, intent, or factual truth.

APSI was evaluated by comparing its scores to political science expert ratings on a validation set of political texts (20 texts per dimension).

A total of 147 expert responses were collected across the three dimensions.

The base multilingual NLI model uses millions of hypothesis–premise training pairs (2.73M) spanning 26 languages.

PolNLI is a political natural language inference dataset (200k hypothesis–premise pairs) used to fine-tune the Political DEBATE model.

The political fine-tuning supports multiple political NLP tasks such as stance detection, topic classification, and related text classification tasks.

APSI is theory-driven, transparent, interpretable, reproducible, and scalable—designed to enable auditing of how scores are produced.

Yes. Outputs are approximate and can be inaccurate or incomplete due to complex political language, model limitations, or bias.

No. APSI should not be used for legal, employment, or automated decision-making affecting people. Human interpretation is required.

APSI is intended for research, journalism, think tanks, civil society organizations, and critical analysis of political discourse.

APSI depends on hypothesis design, can reflect training-data bias, measures rhetoric rather than intent, and lacks real-world context.