Psychological counseling chatbots have many advantages such as low cost and disregard for time and place. However, some technical requirements limit the use and effectiveness of counseling chatbots, such as how to make autoregressive models such as GPT able to understand complex counseling texts, how to be more attuned to the subject matter of the visitor's question to avoid meaningless answers, and how to generate gentle conversational language in the. To address the above problems this paper proposes a MMIAT-GPT (Maximum Mutual Information ALBERT-TextCNN-GPT) model, which first uses ALBERT word vector model with strong text comprehension ability to extract complex semantic features in psychological counseling text. Then it is classified by TextCNN model to output the type of psychological problems for visitors or psychologists for auxiliary treatment, and spliced with the questions, adding transition statements to be inputted into the GPT model together to improve the ability of the GPT model to understand the psychological counseling questioning. Finally, the designed network structure is experimented on the extended PsyQA dataset, which outperforms the GPT2 network model in all results.