Interpreting Deep Text Quantification Models

Quantification learning is a relatively new deep learning task. Differing from a classic classification problem where the class of a single instance is predicted, a quantification model predicts the distribution of classes within a given set of instances. Quantification learning has applications in various domains. For example, in designing political campaign ads, it is important to know the proportion of different aspects voters care about. QuaNet is a recent deep learning quantification model that was shown to achieve good quantification performance. Like many deep learning models, there is no explanation about the contributions of different inputs QuaNet uses to predict a class distribution. In this study, we propose a method to provide such an explanation, which is important to increase users' trust in the model. Our method is the first work on interpreting deep learning quantification models.