Elucidation of the Linguistic Characteristics of Mild Cognitive Impairment Person by Visualizing the Inside of Neural Network

Abstract

It is known that some kinds of neural networks can distinguish the stories of Mild Cognitive Impairment (MCI) persons from those of healthy adults. However, these results do not show the reasons why they can classify them. In this paper, we examine how these judges are made by visualizing the inside of neural networks and try to clarify the Japanese linguistic characteristics of MCI person. First, we apply the gradient-based weight calculation to the convolutional neural network for the time-series word sequence and get the heatmaps to examine the characteristics from the view of its meaning. We find, for example, the difference between MCI person and healthy adult in their scope of interest, the objectivities or subjectivities, the pace or pause in their stories, and the usage of words to express their emotion. Secondarily, we apply the same method to the sequence of decomposed part of speech sequences to see some differences, for example, in the length of sentence or story, or in the frequently appearing sequence patterns, from the view of syntax. Finally, we develop the attention encoder to visualize the strength of the relationship between words and find that the relations are narrow and restricted in case of MCI persons. According to these experimental results, we conclude that MCI may infect the linguistic characteristics even though it does not impair the qualities of their daily lives.