We construct an elderly corpus with a control group (EC) comprising narratives of elderly people with Mild Cognitive Impairment (MCI), healthy elderly people, and younger people in order to develop a method to classify the elderly into healthy and MCI by analyzing the corpus. To do so, we carry out three tasks (picture description task: PDT, episode picture description task: EDT, and animation description task: ADT) to participants (n = 80) and their voices to the tasks are recorded and manually transcribed. 60 out of the participants are the elderly and classified into MCI and healthy control based on Mini Mental State Examination (MMSE). Then, language features such as Type Token Ratio and Idea Density are extracted by analyzing the elderly people’s data and machine learning models are built with the extracted features. In the experiments, our classification model using combined language features obtained from all tasks’ data achieved the highest performance (AUC = 0.85). The results indicate that it would be important to carry out multiple tasks to detect the elderly with MCI.