Abstract:

The linguistic patterns, reflected in the words people use, reveal information about psychological state and psychopathological traits. Individuals with depression are distinguishable from healthy people in their patterns of language use. Identifying the linguistic patterns, as an innovative and convenient approach, would help predict and better diagnosis depression. 
In studies with traditional psychological methods, researchers examined the depressed language using constructed-response data. Specifically, participants are required to write or give speech on a topic (for example, participants were asked to write about their "deepest thoughts and feelings about their personal relationship" or talk about their “future career”). Participants’ responses (texts or transcribed texts) were then quantified and analyzed with the Linguistic Inquiry and Word Count (LIWC) software. Researchers found that the psychopathological traits of depressed patients were related to their linguistic patterns: the state of self-focused attention, social relationship quality and cognitive bias were respectively related to the frequency of first person singular pronouns, first person plural pronouns and emotional words in spoken and written language. It is worth to note that some variables (such as gender, communication environment and the person who are communicated with) were found moderated the relationship between language use and depression. Other factors affecting the results include the small and homogeneous sample, the limited time period researchers could track, and the weak correlations between certain words and depression. In sum, with the traditional psychological methods, researchers could hardly collect massive, longitudinal data from heterogenous participants to achieve stable results.
Recently, studies using social media data have partly addressed these flaws. Social media provides rich and informative data. It is efficient and economical to detect depressed patients from social media users with machine learning method. What researchers do is to build a predictive model by setting the users’ social media data (for example, demographics and the linguistic content of posts) as inputs, users’ score on the standardized depression scales or their depression diagnosis as outputs. The linguistic patterns of depressed patients were generated during the feature engineering or the modeling. LIWC software is also used as the major linguistic analysis tool. These studies replicated previous findings: depressed patients used more first person singular pronouns and negative emotional words, and less first person plural pronouns and positive emotional words. Furthermore, with the social media data, researchers found more language indicators of depressed patients in daily life: they use more death words, anxiety words, swear words, anger words, religion words, health words, as well as causality words, and less second and third person pronouns. These findings are explainable by several theoretical models of depression (e.g. behavioral model, cognitive model and interpersonal model). In addition, social media based research found that the language indicators of depressed users changed in a positive way after they participated in on-line depression communities. The change was strengthened as the number of communication increased. 
Taken together, studies using traditional psychological methods and research with machine learning complemented each other, and increased our understanding of the linguistic patterns of depressed patients. Future studies could explore other language indicators specific to depression, examine the factors that affecting word use patterns, test whether the language indicators found among Westerns fit Chinese, and strengthen related theoretical research.

Key words: depression, language use, social media, prediction, diagnosis