In recent years, deep learning models have become very powerful – even outperforminghumans on a variety of tasks. This enables more real-world applications, including alsosensitive fields such as medical diagnoses or jurisdiction. Besides achieving sufficientlygood performance, the requirement to justify and explain the models’ decisions is becomingincreasingly important.This work aims to enable a broader application of a specific model class that is inherentlyinterpretable, namely explain-then-predict models, by reducing the annotation cost of theexplanations. We focus on the ExPred model as a representative of explain-then-predictmodels.We investigate its dependency on rationale annotations, a special kind of explanation,through training using gradually fewer rationale-labeled instances. Furthermore, we ex-plore different approaches that aim to reduce the number of human-labeled instancesrequired during training, such as active learning and weak supervision.Our results show that even with only a fraction of instances annotated with rationalesfrom the original dataset, ExPred still achieves good performance (within 95% of theperformance when using 100% annotation). Depending on the dataset, only a few thousandannotated rationales are required. Using weak supervision, this can be further reduced, atleast in specific settings. On the Movie Reviews dataset, we achieve good performance withonly 5% of the original rational labels. The tested off-the-shelf active learning methods donot provide any benefit over randomly selecting instances to label. However, the extensivebehavior analysis enables the future design of active learning methods that are tailoredto explain-then-predict models. We start by proposing an active learning method thatoutperforms the random baseline on the Movie Reviews dataset.