Annotator-Specific Rationales for Fine-Grained NLI Explanations
A novel framework simultaneously addresses the prediction of labels specific to annotators and the generation of explanations in natural language inference. This method employs a User Passport system to tailor predictions based on the identity and demographic information of the annotator. Two architectures for explanation are presented: a post-hoc prompt-based explainer and a prefixed bridge explainer, which translates classifier representations conditioned on annotators into a generative model. Findings indicate that the inclusion of explanation modeling enhances performance on a dataset featuring disaggregated NLI annotations along with rationales provided by annotators.
Key facts
- Framework jointly models annotator-specific label prediction and explanations
- Uses User Passport mechanism for conditioning on annotator identity and demographics
- Introduces post-hoc prompt-based and prefixed bridge explainer architectures
- Prefixed bridge explainer transfers classifier representations into generative model
- Dataset includes disaggregated NLI annotations and annotator-provided rationales
- Incorporating explanation modeling substantially improves results
- Published on arXiv with ID 2604.21667
- Focuses on fine-grained signals of individual perspectives
Entities
Institutions
- arXiv