Evaluation of LSTM Explanations in Sentiment Classification Task

Abstract

Deep learning techniques produce impressive performance in many natural language processing tasks. However, it is still difficult to understand what the neural network learned during training and prediction. Recently, Explainable Artificial Intelligence (XAI) is becoming a popular technique to interpret deep neural networks. In this work, we extend the existing Layer-wise Relevance Propagation (LRP) framework and propose novel strategies on passing relevance through weighted linear and multiplicative connections in LSTM. Then we evaluate these explainable methods on a bidirectional LSTM classifier by performing four word-level experiments: sentiment decomposition, top representative words collection, word perturbation and case study. The results indicate that the epsilon-LRP-all method outperforms other methods in this task, due to its ability to generate reasonable word-level relevance, collect reliable sentiment words and detect negation patterns in text data. Our work provides an insight on explaining recurrent neural networks and adapting explainable methods to various applications.