Moralization Toolkit

Analyze moral framing in texts.

Moralization detection with dictionaries and language models.

Moralizations - appeals to moral values such as fairness, care or freedom used to justify a position or demand - are a central persuasive strategy in political, media and science communication.
The Moralization Toolkit, developed at the CHAI Lab (University Heidelberg) detects them in German, English and French texts in two steps:
(1) DiMi, a lexicon-based approach, finds candidate passages via curated morality-indicating words;
(2) fine-tuned models and large language models then assess whether these passages are actually moralizing and, in the case of LLMs, provide explanations, protagonists, and moral values.
This two-step method was used to compile The Moralization Corpus (Becker et al., LREC 2026).

Analyze one text at a time.

Step 1 – Find Candidates: Dictionary-Based Search (DiMi)

The Dictionary of Morality Indicating Words (DiMi) searches your text for words from curated lexicons of morality-indicating words in German, English and French and returns each match with its surrounding context (Sentence with Match +/- 2 Sentences). For details on the method, see Becker et al. (2023).
Matches are moralization candidates only. Verify them with the language models in Step 2.

Input Language
loading lemmas…LemmaA lemma is a dictionary's base word form, such as run for runs or running. DiMi uses these word forms to find moralizing language in your text.Characters: 183/5000

Step 2 – Classify, Analyze, and Explain: Language Models

Choose between two types of multilingual models. The fine-tuned models XLM-RoBERTa and mmBERT run locally, are free to use and return a label with a confidence score. The LLMs Claude Haiku 4.5 and GPT-5 mini additionally explain their decision and identify the moral values and protagonists involved. Each LLM prediction costs 1 credit, and 20 free credits are available per day.
Automatic predictions can be wrong, so treat them as a starting point for analysis, not as ground truth.

Input Language
ModelModelLocal models: XLM-RoBERTa and mmBERT are open models fine-tuned specifically to detect moralization. They run locally, do not send your text to an external provider, and are free to use. Closed-source models: Claude Haiku 4.5 and GPT-5 mini are proprietary models accessed via external APIs. They can provide explanations, moral values, and protagonists, but your text is sent to the selected provider. Each prediction costs 1 credit, 20 credits per day are free.
Characters: 183/5000 · Cost: Free
API Token
No API token and out of credits?
Email us to request an API token, or wait 24 hours for your credits to refill automatically. The free tier includes 20 credits per day for Claude Haiku 4.5 and GPT-5 mini (1 credit per prediction). DiMi, XLM-RoBERTa and mmBERT are always free.
Contact us via email.