Prototypical Convolutional Neural Network for a Phrase-Based Explanation of Sentiment Classification

The attention mechanisms are often used to support an interpretation of neural network based classification of texts by highlighting words to which the network paid attention while making a prediction. Following recent studies, the attention technique does not always provide a faithful explanation of the model. Thus, in this paper we study another idea of prototype-based neural networks. Although for texts they obtain promising results, they may provide explanations in the form of comparisons of whole (potentially long) documents or also run into problems with providing reliable explanations. To overcome it, in this work a new prototype-based convolutional neural architecture for text classification is introduced, which provides predictions' explanations in the form of similarities to phrases from the training set. The experimental evaluation demonstrates that the proposed network obtains similar classification performance to the black-box convolutional networks while providing faithful explanations. Moreover, it is shown that a new method for dynamic tuning of the number of prototypes introduced in this paper offers performance gains against static tuning.

Automated summary

This paper introduces a new prototype-based convolutional neural architecture for text classification, which offers faithful predictions' explanations compared to traditional attention mechanisms. It also demonstrates that dynamic tuning of the number of prototypes can lead to performance gains.