Explaining a Deep Neural Model with Hierarchical Attention for Aspect-Based Sentiment Classification Using Diagnostic Classifiers

LCR-Rot-hop++ is a state-of-art model for Aspect-Based Sentiment Classification. However, it is also a black-box model where the information encoded in each layer is not understood by the user. This study uses diagnostic classifiers, single layer neural networks, to evaluate the information encoded in each layer of the LCR-Rot-hop++ model. This is done by using various hypotheses designed to test for information deemed useful for sentiment analysis. We conclude that the model did not focus on identifying the aspect mentions associated with a word and the structure of the sentence. However, the model excelled in encoding information to identify which words are related to the target. Lastly, the model was able to encode to some extent information about the word sentiment and sentiments of the words related to the target.