Exploring the Feasibility of Assessing Stress, Depression, and Anxiety Using Text-Based Language Analysis

Data Collection

This study examined whether the textual characteristics of emotional-state responses differed according to stress, depression, and anxiety levels among adults (aged 20–64 years) and older adults (aged 65 years and above). Data were collected over a 98-day period, from December 19, 2023, to March 26, 2024, in the Seoul and Gyeonggi regions through both face-to-face and online surveys assessing mental health status and perceptions of agro-healing.

Although the original survey targeted individuals across the entire life span, including children, adolescents, adults, and older adults, the present study analyzed only responses from adults and older adults whose open-ended narrative responses met the criteria for sufficient length and completeness. The survey was designed to assess participants’ mental health status and perceptions of agro-healing. In this study, responses to the open-ended questions included at the end of the questionnaire (i.e., narratives describing participants’ current emotional state and recently memorable emotions or events) were analyzed alongside self-report measures of mental health status, including the PHQ-9, GAD-7, and BEPSI-K.

Participants were recruited through a professional survey agency with a nationwide online panel in South Korea, and the agency’s adult and older adult panels were used to administer the online survey. In parallel with the online data collection, face-to-face surveys were conducted at senior welfare centers in the Seoul and Gyeonggi regions to facilitate additional participation from older adults. These in-person survey sessions were arranged through prior consultation with the participating institutions, and respondents were informed that they could take breaks at any time if they experienced fatigue during the survey process. Upon completion of the survey, online participants received platform-based reward points, whereas participants in the face-to-face survey received a small potted plant as compensation.

A total of 268 valid responses from adults and older adults were collected and included in the final analysis. This study was approved by the Institutional Review Board (IRB) of Konkuk University (Approval No. 7001355-202310- HR-715), and all research procedures were conducted in strict accordance with established ethical guidelines.

Assessment Measures

Participants’ mental health status was assessed using self-report questionnaires measuring stress, depression, and anxiety. Stress levels were evaluated using the Korean version of the Brief Encounter Psychosocial Instrument (BEPSI-K) (Bae et al., 1992; Frank and Zyzanski, 1988). The instrument consists of five items rated on a 5-point Likert scale ranging from 0 to 4. Higher mean item scores indicate greater levels of stress, with scores below 1.8 classified as low stress, scores between 1.8 and 2.8 as moderate stress, and scores of 2.8 or higher as high stress. The reported reliability of the original instrument was Cronbach’s α = .84, whereas the reliability coefficient calculated for the present sample was Cronbach’s α = .869.

Depression levels were measured using the Patient Health Questionnaire-9 (PHQ-9) (Han et al., 2008; Spitzer et al., 1999). The PHQ-9 comprises nine items rated on a 4-point Likert scale (0–3), with total scores categorized as follows: 0–4 (minimal or no depressive symptoms), 5–9 (mild depression), 10–19 (moderate depression), and 20–27 (severe depression). For the purpose of comparing the textual characteristics of moderate and severe depression in this study, scores ranging from 5 to 19 (mild to moderate depression) were combined into a single category labeled “moderate depression.” The reported reliability of the original instrument was Cronbach’s α = .86, and the reliability coefficient for the present sample was Cronbach’s α = .877.

Anxiety levels were assessed using the Generalized Anxiety Disorder Scale-7 (GAD-7) (Spitzer et al., 2006). The scale consists of seven items rated on a 4-point Likert scale (0–3), with total scores categorized as follows: 0–4 (minimal or no anxiety), 5–9 (mild anxiety), 10–14 (moderate anxiety), and 15–21 (severe anxiety). For the purposes of the present analysis, anxiety scores were dichotomized into two groups: “no anxiety” (0–4) and “anxiety present” (≥ 5). The reported reliability of the original instrument was Cronbach’s α = .92, whereas the reliability coefficient for the current sample was Cronbach’s α = .917.

To analyze the linguistic characteristics of participants’ textual responses of their emotional states, this study employed the Data Evaluation Unified System (DEUS), a Korean text analysis platform developed by the Konkuk Research Institute for Multilingualism and Multiculturalism (KRIMM). DEUS quantitatively evaluates a wide range of linguistic features in Korean texts, including surface structure, word frequency, lexical diversity, referential cohesion, syntactic complexity, pronoun usage, conjunctive adverbs, and psycholinguistic lexical indices (Jang and Jeon, 2024). Jang and Jeon (2024) applied DEUS to examine the difficulty and complexity of the reading section in the Test of Proficiency in Korean (TOPIK), demonstrating its utility as an objective tool for evaluating the structural and lexical characteristics of Korean texts. Furthermore, Kim et al. (2025) utilized DEUS to analyze open-ended responses from participants in agro-healing programs, quantitatively examining linguistic indicators such as word count, sentence count, lexical proficiency, syntactic structure, and emotional expression to capture patterns of emotional expression and psychological change.

The collected open-ended responses were preprocessed and converted into text (.txt) files for analysis. Preprocessing involved removing unnecessary elements from the raw survey data, including special characters, duplicate entries, and excessive whitespace. Original wording was preserved without spelling correction, while line breaks and text encoding formats were standardized to ensure consistency. The resulting text files were subsequently analyzed using DEUS to extract linguistic indicators.

The DEUS framework provides measurements across nine major categories (Table 1): text surface structure, word frequency, lexical diversity, referential cohesion, syntactic complexity, pronoun usage, conjunctive adverbs, psycholinguistic lexical indices, and emotional expression intensity. Log-transformed word frequency quantitatively represents the degree to which a text contains high-frequency everyday vocabulary. It is determined by deriving the general usage frequency of each word from the Word Frequency Database developed by the National Institute of Korean Language (NIKL) and applying a logarithmic transformation (Kim, 2005; Graesser et al., 2007). Total vocabulary includes both content words, which convey substantive meaning (e.g., nouns, verbs, adjectives, and adverbs), and function words, which serve grammatical functions and contribute to sentence structure (e.g., particles, endings, conjunctions, pronouns, and auxiliary verbs). Lexical diversity was assessed using the type-token ratio (TTR), calculated by dividing the number of unique words (types) by the total number of words (tokens). TTR serves as an indicator of vocabulary repetition and richness within a text (Jang and Jeon, 2024), with higher values reflecting greater lexical diversity and lower levels of word repetition. Among the indices of syntactic complexity, the constituent ratio is calculated by dividing the number of constituent nodes within a sentence structure by the total number of words and is used to evaluate the hierarchical depth and structural complexity of sentences. Finally, emotional expression intensity was measured using the Korean Emotional Word Database developed by Son et al. (2012) based on the “Modern Korean Word Frequency” corpus. The database contains 504 emotion-related words assigned intensity scores ranging from 1 to 10. Emotional expression intensity was calculated by averaging the intensity values of emotional words appearing in each text.

Table 1

Language feature categories and their corresponding measurement indicators in the DEUS system

Data Analysis

Descriptive statistics, including frequencies and percentages for participants’ demographic characteristics, were calculated using Microsoft Excel (Microsoft Office 2020; Microsoft Corp., Redmond, WA, USA). Statistical analyses of the linguistic features extracted from the textual responses were conducted using IBM SPSS Statistics for Windows (Version 25.0; IBM Corp., Armonk, NY, USA).

To compare linguistic features across different levels of stress, depression, and anxiety, one-tailed independent-samples t-tests were performed. For stress and depression, comparisons were conducted between the moderate- and high-level groups, whereas for anxiety, differences between the “anxiety present” and “no anxiety” groups were examined. One-tailed tests were employed based on the a priori hypothesis that higher levels of stress, depression, and anxiety would be associated with directional changes in specific linguistic indicators. The levels of the three psychological indicators served as the independent variables, while the text-based linguistic measures generated by the DEUS system were treated as the dependent variables.

In addition, two-tailed Pearson correlation analyses were conducted to examine the relationships among stress, depression, and anxiety, as well as the associations between these psychological measures and linguistic indicators. The correlation analyses included all 24 linguistic metrics encompassed within the nine language-feature categories generated by DEUS. However, only linguistic indicators that showed statistically significant correlations were reported in the Results section. For all statistical analyses, the significance level was set at α = .05.

Demographic Characteristics

This study analyzed data collected from a total of 268 adults and older adults, and their demographic characteristics are presented in Table 2. Regarding gender distribution, 46% of the participants were male and 54% were female. In terms of age, adults aged 20–64 years comprised 52% of the sample, whereas older adults aged 65 years and above accounted for 48%.

Table 2

Demographic characteristics and mental health status of respondents (N = 268)

With respect to mental health status, stress levels were categorized as low (35%), moderate (47%), and high (18%), with the moderate-stress group representing the largest proportion of participants. Similarly, depression levels were classified as low (10%), moderate (51%), and high (39%), indicating that moderate depression was the most prevalent category. Regarding anxiety, 31% of the respondents were classified as having anxiety symptoms, whereas 69% were classified as not having anxiety symptoms. These findings indicate that approximately one-third of the respondents reported experiencing symptoms of anxiety.

Comparison of Linguistic Characteristics According to Stress Levels

A comparison of linguistic characteristics between the moderate-stress and high-stress groups revealed that the high-stress group exhibited a significantly shorter average word length (p < .05) and a significantly higher log-transformed word frequency across the entire vocabulary (p < .05) than the moderate-stress group (Table 3). These findings suggest that individuals experiencing higher levels of stress may tend to use vocabulary consisting of relatively short and frequently used everyday words.

Table 3

Comparison of linguistic characteristics according to stress levels

Stress arises when physiological and psychological demands exceed an individual’s capacity to cope (Anisman and Merali, 1999). Previous research has reported that chronic stress can impair cognitive functions associated with higher-order thinking and planned language production (Ghanbari et al., 2020). Moreover, elevated stress responses have been shown to reduce the cognitive complexity of language (Saslow et al., 2013), and studies have suggested potential associations between the quality of linguistic production and emotional stability or cognitive attention (Girotti et al., 2024; Pennebaker et al., 2003). Consistent with these findings, individuals experiencing high levels of stress have been reported to use relatively short and simple expressions in social media posts and everyday written communication (Nijhawan et al., 2022). In the present study, responses such as “I want to rest” and “I don’t want to do anything” were frequently observed among participants with higher stress levels, providing additional support for a potential relationship between stress levels and linguistic simplification.

Meanwhile, exposure to natural environments and participation in agro-healing activities have been reported to be associated with emotional stability and attentional restoration. According to Attention Restoration Theory (Kaplan and Kaplan, 1989), stimuli derived from natural environments promote involuntary attention, thereby facilitating the recovery of directed attention. In addition, previous research has reported that plant-based environments or horticultural activities are positively related to attention-related indicators (Kim et al., 2020; Kim et al., 2021a; Kim et al., 2021b). Participation in vegetable gardening or rural resource-based activities has also been linked to reductions in perceived stress and cortisol levels (Han et al., 2018; Park and Park, 2024; Yun et al., 2024). Furthermore, Kim et al. (2025) employed DEUS analysis to demonstrate that participation in agro-healing programs can shift emotional expression patterns toward a more stable and organized state.

Taken together, these findings suggest that lexical features, such as average word length and word frequency, may provide preliminary reference points for examining stress-related changes in language use in future agro-healing studies.

Comparison of Linguistic Characteristics According to Depression Levels

A comparison between the moderate-depression and high-depression groups revealed that the high-depression group produced significantly fewer words (p < .01) and sentences (p < .01) than the moderate-depression group. In contrast, both the overall type-token ratio (TTR) and the content-word TTR were significantly higher in the high-depression group (p < .01) (Table 4). In the referential cohesion analysis, the high-depression group exhibited significantly lower noun repetition rates both between adjacent sentences (p < .05) and across all sentences (p < .05). Furthermore, this group demonstrated a lower proportion of sentence components reflecting syntactic complexity (p < .05) and a lower frequency of causal and consequential conjunctive adverbs (p < .01). In the sentiment analysis, the intensity of boredom-related expressions was significantly higher in the high-depression group (p < .05).

Table 4

Comparison of linguistic characteristics according to depression levels

These findings suggest that texts produced by the high-depression group were generally shorter and characterized by lower inter-sentential connectivity and reduced structural complexity. The higher TTR observed in this group is therefore more appropriately understood as a TTR increase pattern that may have resulted from reduced word repetition in shorter responses, rather than as evidence of genuinely greater lexical diversity. In addition, the decreases in noun repetition and conjunctive adverbs suggest that logical connections between sentences may have been relatively weakened during text construction. In the actual responses, expressions such as “I don’t want to do anything” and “I feel sad and want to die” were frequently observed, indicating a linguistic pattern characterized by negative emotional expression.

This pattern is partially consistent with previous findings regarding linguistic withdrawal in depressive states. Trifu et al. (2017) reported that individuals with depression tend to use short and simple sentences, while Smirnova et al. (2018) suggested a preference for single-clause structures over more complex, multi-clause constructions.

Cognitive impairments commonly associated with depression, including deficits in attention, executive functioning, and processing speed, may also be related to the reduced response length and cohesion observed in this study (Hammar et al., 2022). In addition, the increased use of boredom-related expressions may suggest that reduced affect or motivational decline among participants with higher depression levels was partly reflected in the emotional tone of their texts.

Previous studies have suggested that agro-healing activities and nature-based experiences may be associated with responsibility, achievement, self-efficacy, and positive changes in emotional regulation (Kim and Park, 2018; Lee et al., 2016; Park and Jung, 2017). Moreover, exposure to natural environments has been linked to enhanced positive affect and reduced depressive symptoms (Kim et al., 2023; Kim et al., 2025; Olszewska-Guizzo et al., 2022). Therefore, analyzing depression-related linguistic characteristics may provide useful reference indicators for future agro-healing program evaluations, particularly when examining changes in response length, cohesion, and emotional expression patterns associated with depressive symptoms.

Comparison of Linguistic Characteristics According to Anxiety Levels

The anxiety group produced significantly fewer sentences (p < .05) than the non-anxiety group, whereas its overall and content-word TTRs were significantly higher (p < .01) (Table 5). In addition, participants with anxiety demonstrated a significantly higher use of modifiers and adverbs (p < .05). In contrast, their use of both overall (p < .01) and causal/consequential conjunctive adverbs (p < .01) was significantly lower. Sentiment analysis further revealed that the intensity of boredom-related expressions was significantly higher in the anxiety group (p < .05).

Table 5

Comparison of linguistic characteristics by anxiety status

These findings suggest that participants with anxiety tended to produce shorter responses with reduced inter-sentential connectivity. However, within these relatively brief narratives, they used modifiers and adverbs more frequently, indicating that descriptive or emotionally emphasized expressions may have been relatively prominent. Thus, anxiety may not be reflected in greater overall text length, but rather in a pattern in which immediate emotional expression is emphasized while the logical organization of information is weakened. In the actual text responses, expressions such as “I feel anxious” and statements such as “I keep worrying, so it is hard to focus” were frequently observed, suggesting that the tension and nervousness associated with anxiety may be reflected in language use.

This pattern is partially consistent with previously reported characteristics of language use in anxious states. Earlier studies have reported that individuals with elevated anxiety levels tend to use more negative words and emotional adverbs (Rook et al., 2022; Teferra et al., 2022), and some research has suggested that expressions of individuals experiencing anxiety may appear verbose or repetitive (Robinson et al., 2019). However, because the number of sentences decreased in the present study, the findings are more appropriately interpreted not as increased verbosity per se, but rather as a pattern in which modifier use and emotional emphasis become relatively enhanced within shorter texts. Anxiety is associated with physiological arousal, such as increased sympathetic nervous system activity (Wenner, 2018), and such arousal may contribute to more immediate and emotionally intensified forms of linguistic expression (Herman et al., 2018).

Meanwhile, agro-healing activities have been suggested to promote a sense of emotional security and environmental control through structured, repetitive, and predictable activity frameworks (Panţiru et al., 2024). Furthermore, previous studies have reported that visual and tactile stimulation derived from plants, soil, and other natural elements is associated with reduced neurophysiological arousal and enhanced emotional stability (Kim et al., 2023; Kim and Park, 2025). Taken together, these findings suggest that linguistic differences associated with anxiety levels may serve as useful reference indicators for examining changes in expression patterns related to tension and nervousness within agro-healing programs.

Correlations between Psychological Characteristics and Linguistic Indicators

The correlations between stress, depression, and anxiety levels and text-based linguistic indicators are presented in Table 6. Significant positive correlations were observed among all three psychological variables, with the strongest correlation identified between depression and anxiety (r = .838, p < .01). Stress also demonstrated significant positive correlations with both depression (r = .524, p < .01) and anxiety (r = .545, p < .01).

Table 6

Correlations among psychological variables and significant linguistic indicators

Regarding linguistic indicators, the number of lexical senses was positively correlated with stress (r = .122, p < .05). This finding suggests that higher levels of stress may be associated with an increased tendency to use words with multiple meanings (i.e., polysemous words), which partially corresponds with the tendency toward high-frequency vocabulary observed in the high-stress group. In other words, elevated stress levels may be associated with greater reliance on more general and commonly used vocabulary.

The use of third-person pronouns was positively correlated with anxiety (r = .162, p < .01), indicating that higher anxiety levels may be associated with an increased tendency to refer to external objects or other individuals. Together with the greater use of modifiers and adverbs and the higher lexical diversity observed in the anxiety group, this finding suggests that linguistic expressions directed toward external entities may become relatively more prominent as anxiety levels increase.

Among the emotional expression intensity variables, boredom showed positive correlations with both stress (r = .134, p < .01) and depression (r = .149, p < .01). These findings suggest that boredom-related emotional expressions tend to increase as levels of stress and depression rise. This result is also partially consistent with the group comparison findings showing greater intensity of boredom- related expressions in the high-depression group, further supporting the possibility that negative emotional states are reflected in the emotional tone of language.

Overall, the three psychological variables demonstrated significant associations with specific linguistic characteristics, suggesting that text-based linguistic analysis could serve as a useful supplementary approach to understanding an individual’s psychological state. The observed correlation patterns generally align with prior findings, providing a helpful reference for interpreting the linguistic characteristics identified in this study.