Building Interpretable Interaction Trees for Deep NLP Models

This paper proposes a method to disentangle and quantify interactions among words that are encoded inside DNN for natural language processing. We construct tree encode salient extracted by the DNN. Six metrics proposed analyze properties of between constituents in sentence. The interaction is defined based on Shapley values words, which considered as an unbiased estimation word contributions network prediction. Our used BERT, ELMo, LSTM, CNN, Transformer networks. Experimental results have provided new perspective understand these DNNs, demonstrated effectiveness our method.