Interpreting Convolutional Networks Trained on Textual Data

Abstract

There have been many advances in the artificial intelligence field due to the\nemergence of deep learning. In almost all sub-fields, artificial neural\nnetworks have reached or exceeded human-level performance. However, most of the\nmodels are not interpretable. As a result, it is hard to trust their decisions,\nespecially in life and death scenarios. In recent years, there has been a\nmovement toward creating explainable artificial intelligence, but most work to\ndate has concentrated on image processing models, as it is easier for humans to\nperceive visual patterns. There has been little work in other fields like\nnatural language processing. In this paper, we train a convolutional model on\ntextual data and analyze the global logic of the model by studying its filter\nvalues. In the end, we find the most important words in our corpus to our\nmodels logic and remove the rest (95%). New models trained on just the 5% most\nimportant words can achieve the same performance as the original model while\nreducing training time by more than half. Approaches such as this will help us\nto understand NLP models, explain their decisions according to their word\nchoices, and improve them by finding blind spots and biases.\n