Objective: Identify human values expressed by each specific actor in a news article and avoid over-prediction or value distortion.
Provided data:
Factual news article
List of main actors
Output:
For each actor, predict a set of values with direction:
Aligned with human values
Contradictory to human values
Error definition:
Omission error: missed gold labels
Hallucination error : predicted values not in gold labels
Reversal error : predicted opposite direction (aligned ↔ contradictory)
Objective: Perform sentence-level paraphrasing while keeping original semantics and human values of entire article unchanged
Provided data:
News article
List of main Actors
Specified paraphrasing type (each with ~100 samples):
Negation Switching | Entailment Extension | Opposite Polarity | Ellipsis / Compression
Sentence-level applicable transformations
Output:
Rewritten article
Output Requirements:
Faithfulness to original news & Natural and fluent news style
Must preserve human value expressions
Objective: Generate a concise news summary while preserving core human values.
Provided data:
Full news article (document-level)
List of main Actors
Compression rate: (min_words ~ max_words)
Output:
Generated summary
Output Requirements:
Faithfulness to original news
Readable and fluent summary style
Core value must be retained; omission will be penalized heavily.
Omit supporting/marginal values is accepted, but their inclusion provides small bonus rewards.
No hallucination or value distortion