A novel explainable structure for text classification

With the development of deep learning, text classification has achieved very good results, but the poor interpretability of the model still limits its application in practical scenarios to a certain extent. Many explainable text classifiers extract words from a sentence and then observe their effect on increasing or decreasing classification accuracy. However, in many cases, the relationship between words in a sentence is interdependent and closely related. On account of the above, selecting words individually often has little effect on the classification results. To address the above situation, we propose a new model which treats interpretability as an intrinsic property, using constituent trees to generate continuous interpretable words instead of isolated words and it achieves good results on several datasets.