Previously we have showed that some kinds of recurrent neural networks can distinguish the stories of Mild Cognitive Impairment (MCI) persons from those of healthy adults. In this research, we examine how these judges are made by visualizing the inside of neural networks. First, we apply the grad-CAM scores to the time-series word sequence from the view of their meanings and get the heatmaps. We find some characteristics regarding usage of words, for example, the usage of feeling words, subjective words, pause words. Secondarily we apply the same method to the sequence of a decomposed part of speech and extract some interesting patterns. Finally, we develop the attention encoder to visualize the strength of relationship between words. According to these experimental results, we conclude that MCI may have an impact on their linguistic characteristics
