[Paper Review] Self-interpretable Convolutional Neural Networks for Text Classification

Wei Zhao, Rahul Singh+3 more|arXiv (Cornell University)|May 18, 2021

Explainable Artificial Intelligence (XAI)10 references4 citations

TL;DR

This paper proposes a self-interpretable convolutional neural network (CNN) for text classification by leveraging local linear models derived from ReLU-based deep neural networks (DNNs). By mapping these local linear models back through max-pooling layers to original n-grams, the method produces sparse, interpretable feature representations while maintaining classification accuracy comparable to standard CNNs.

ABSTRACT

Deep learning models for natural language processing (NLP) are inherently complex and often viewed as black box in nature. This paper develops an approach for interpreting convolutional neural networks for text classification problems by exploiting the local-linear models inherent in ReLU-DNNs. The CNN model combines the word embedding through convolutional layers, filters them using max-pooling, and optimizes using a ReLU-DNN for classification. To get an overall self-interpretable model, the system of local linear models from the ReLU DNN are mapped back through the max-pool filter to the appropriate n-grams. Our results on experimental datasets demonstrate that our proposed technique produce parsimonious models that are self-interpretable and have comparable performance with respect to a more complex CNN model. We also study the impact of the complexity of the convolutional layers and the classification layers on the model performance.

Motivation & Objective

• To address the black-box nature of deep learning models in NLP by enabling intrinsic interpretability without sacrificing performance.
• To develop a method that maps local linear approximations from ReLU-DNNs back to original text n-grams for model interpretation.
• To evaluate the trade-off between model complexity in convolutional and classification layers and interpretability.
• To produce a parsimonious, self-interpretable model that retains competitive classification accuracy on text datasets.

Proposed method

• Utilizes the local-linear property inherent in ReLU-based deep neural networks to generate local explanations for predictions.
• Applies max-pooling in a standard CNN architecture to extract salient n-gram features from word embeddings.
• Maps the local linear models from the final dense layer back through the max-pooling operation to identify influential n-grams.
• Constructs a self-interpretable model by selecting only the most relevant n-grams based on the back-mapped local linear coefficients.
• Optimizes the model using standard CNN training with ReLU activation, preserving end-to-end differentiability.
• Employs a sparsity-inducing mechanism on the final classification layer to promote parsimonious feature selection.

Experimental results

Research questions

• RQ1 Can local linear models from ReLU-DNNs be effectively mapped back to original n-grams to produce interpretable text classification models?
• RQ2 How does the complexity of convolutional and classification layers affect model performance and interpretability?
• RQ3 To what extent can a self-interpretable CNN achieve performance comparable to a standard, non-interpretable CNN?
• RQ4 What is the impact of sparsity in feature selection on model accuracy and interpretability in text classification?

Key findings

• The proposed method achieves classification performance comparable to a standard CNN on benchmark text classification datasets.
• The model produces sparse, self-interpretable representations by identifying key n-grams responsible for predictions through back-mapped local linear models.
• Reducing the complexity of convolutional and classification layers leads to a trade-off between model performance and interpretability, with optimal configurations identified.
• The method successfully maps local linear approximations from the DNN to original n-grams, enabling direct interpretation of model decisions.
• The self-interpretable model maintains high accuracy while significantly improving transparency compared to standard black-box CNNs.
• Sparsity in the final layer enhances interpretability without substantial loss in predictive performance, validating the method's practicality.