Recently, a technique called Layer-wise Relevance Propagation (LRP) was shownto deliver insightful explanations in the form of input space relevances forunderstanding feed-forward neural network classification decisions. In thepresent work, we extend the usage of LRP to recurrent neural networks. Wepropose a specific propagation rule applicable to multiplicative connections asthey arise in recurrent network architectures such as LSTMs and GRUs. We applyour technique to a word-based bi-directional LSTM model on a five-classsentiment prediction task, and evaluate the resulting LRP relevances bothqualitatively and quantitatively, obtaining better results than agradient-based related method which was used in previous work.
展开▼
