摘要Language disorder,a common manifestation of Alzheimer's disease(AD),has attracted widespread attention in recent years.This paper uses a novel natural language processing(NLP)method,compared with latest deep learning technology,to detect AD and explore the lexical performance.Our proposed approach is based on two stages.First,the dialogue contents are summarized into two categories with the same category.Second,term frequency-inverse document frequency(TF-IDF)algorithm is used to extract the keywords of transcripts,and the similarity of keywords between the groups was calculated separately by cosine distance.Several deep learning methods are used to compare the performance.In the meanwhile,keywords with the best performance are used to analyze AD patients'lexical performance.In the Predictive Challenge of Alzheimer's Disease held by iFlytek in 2019,the proposed AD diagnosis model achieves a better performance in binary classification by adjusting the number of keywords.The F1 score of the model has a considerable improvement over the baseline of 75.4％,and the training process of which is simple and efficient.We analyze the keywords of the model and find that AD patients use less noun and verb than normal controls.A computer-assisted AD diagnosis model on small Chinese dataset is proposed in this paper,which provides a potential way for assisting diagnosis of AD and analyzing lexical performance in clinical setting.