Language in psychotherapy provides important information about the process and outcome of counseling. Qualitative coding is a traditional and commonly used method for process and outcome research. However, it costs much time and money, and it is hard to deal with a large scale of psychotherapy data. Considering the limitations of qualitative coding, Language Inquiry and Word Count(LIWC) has become an alternative and convenient method to explore the relationship between language use and psychological status. Evidence showed that LIWC words revealed people’s mental health and thinking styles. The present study analyzed the association between clients’ language use during psychotherapy and their symptoms in a Chinese context, to identify the possibility of predicting clients’ improvement by analyzing natural language. Methods: The recordings of psychotherapy and the outcome questionnaires of 28 clients were collected. Most of participants were female because data was collected in the mental health center of a normal university where male to female ratio was 1:8. Each client received 5.5 sessions of counselling averagely. 13 therapists took part in the study and all of them were trained. 10 recordings were not included in the analysis due to loss of data. The remaining 144 recordings were transcribed into texts by research assistants word by word. After checking mistakes, data cleaning was conducted including cutting words, cutting off talk turns. And then the LIWC words frequencies were calculated by scripts written in python 3.6. According to previous study, raw word frequencies were transformed into Z score. The principal component analysis was conducted to find out the main factors of LIWC features associated with clients’ symptoms. Moreover, the main factors of LIWC features were used to predict the clients’ improvement across sessions in a multilevel model. Results: Results showed that 31 categories of LIWC features were associated with clients’ symptom level significantly. Results of principal component analysis revealed that 8 factors interpreted 75.11% of the variance, including biological feelings words, social relationship words, time words, cognition words, function words, emotion words, superfluous words and filler words. Among the factors, frequencies of biological feelings words, function words, emotion words, cognition words and filler words could predict the symptom level of clients in a general linear regression model. To determine the associations between the change of LIWC words across sessions and the change of clients’ symptom level. In a multilevel model, it was promising to find that biological feelings and emotion words can be used to predict the outcome of psychotherapy controlling the time variable. Conclusions: As a preliminary research applied text analysis in psychotherapy, the present study revealed that the words in psychotherapy have potential to predict clients symptoms. Although there are some limitations such as limited participants, rough LIWC features and simple algorithm, the present study gives us some implications as following. With the development of artificial intelligence, more sophisticated text analysis tools even audio analysis could be applied in monitoring the process and outcome of psychotherapy automatically which would help clients and therapists a lot. Future work can focus on a more accurate monitoring model of psychotherapy by combing more useful features and algorithms.