1,800 movie reviews annotated with human-extracted rationales for sentiment classification. Each entry pairs a review document with specific text spans that justify the assigned positive or negative label.
Use Cases
- Train a rationale extraction model using the rationale and review text features.
- Evaluate the alignment of model attention mechanisms with human-provided rationale spans.
- Benchmark explainable AI (XAI) methods on the task of evidence identification in natural language.
Strengths
- 1,800 movie reviews with human-selected rationales.
- Includes specific text spans that serve as evidence for document-level labels.
- Provides document-level text paired with localized evidence markers.