A hierarchical neural network model for Japanese toward detecting mild cognitive impairment

We found that some signs of mild cognitive impairment (MCI) might be presented in a structure of a sentence and a relation between sentences talked by a man, and develop a neural network model which has an analogy with the hierarchical structure of speakers, topics, sentences and words in Japanese. We build our model based on 2-layered bi-directional LSTM, corresponding to words-sentences and sentences-topics hierarchy. As a layer corresponding to speakers, we use a linear classifier with self-attention. The test result shows a largely improved AUC, compared with another test by using the normal 2-layered bi-directional LSTM with TBPTT. The result also indicates that there are some characteristic patterns in a talk by an elderly person with MCI. We classify the character vectors of topics generated from our model through learning into clusters whose number is 1/10 of the number of persons in our data. Since these clusters have almost less than 10% or more than 90% rate of positive share, we conclude that we can develop a screening method based on a talk in Japanese by an elderly person in the near future.