mlburnham/deberta-v3-large-polistance-affect-v1.1

Model Description

This model adapts Moritz Laurer's zero shot model for political texts. It is currently trained for zero-shot classification of stances towards political groups and people, although it should also preform well for topic and issue stance classification. Further capabilities will be added and benchmarked as more training data is developed.

Training Data

The model was trained using the PoliStance Affect and PoliStance Affect_QT datasets.

• Polistance Affect: ~27,000 political texts about U.S. politicians and political groups that have been triple coded for stance.
• Polistance Affect QT: A set of quote tweets about U.S. politicians that pose a particularly challenging classification task.

The test set for both datasets contains documents about six politicians that were not included in the training set in order to evaluate zero-shot classification performance.

Evaluation

Results below are performance on the PoliStance Affect test set.