How South Korean Internet users experienced the impacts of the COVID-19 pandemic: discourse on Instagram

The study investigates how South Korean Instagram users expressed and coped with emotional and behavioral responses to COVID-19 during sustained social distancing (the “everyday-life quarantine” period starting May 6, 2020). Motivated by the widespread shift of social sharing to online platforms amid stay-at-home measures and the recognized psychological impacts of pandemics (fear, anxiety, loneliness), the authors aim to identify the co-constructed public discourse on Instagram: What messages were negotiated, whose voices were amplified or suppressed, and how individuals managed daily life under ongoing threat. The work is situated within discourse analysis, viewing online posts as performative acts shaping social meanings, and leverages text mining to handle large-scale SNS data.

The paper situates discourse analysis as the study of language as social practice that both reproduces and transforms social relations. With the rise of social media, users co-construct meanings and regulate daily life through performative posting and interaction. Traditional discourse study sampling is challenged by the scale and dispersion of SNS data, prompting integration with text mining. Prior work shows text mining (e.g., co-word analysis and topic modeling such as DMR) can uncover latent topics and co-occurrence structures in unstructured data, enabling interpretation of public perceptions during crises. The authors also reference scholarship on social sharing of emotion, noting that online networks may facilitate community knowledge circulation, emotional regulation, and identity/cohesion but can also amplify fear, blame, and hate frames toward minorities.

Design: Mixed-method text analytics combining co-word analysis and Dirichlet Multinomial Regression (DMR) topic modeling to examine Instagram discourse during sustained social distancing in South Korea. Data collection: Instagram posts in Korean from May 6, 2020 to June 16, 2020, corresponding to the easing from intensive measures to “everyday-life quarantine.” Posts were collected via a Python 3.6 crawler using hashtag queries matching the Korean equivalents of #social_distancing and/or #everyday_distancing. Collected fields included post date, main text body, and image descriptions; emoticons and uninterpretable unicodes were excluded. No personal identifiers (e.g., usernames, locations) were collected. Initial pool: 27,362 posts (mean ≈ 651.48 posts/day). Screening and inclusion: Excluded (1) regrammed/reposted content; (2) posts lacking the terms “Corona” and/or “COVID” in the main text. Final dataset: 8,241 posts. Qualitative sorting yielded three subsets by post intent: ANN (official announcements/news; n=287), BIZ (commercial/promotion; n=2,799), and Everyday-Life (individual experiences and emotions; n=5,155). Analyses focused primarily on the Everyday-Life subset, with ANN/BIZ used contextually. Pre-processing: Tokenization to sentences and unigrams; morphological parsing with KOMORAN to extract meaningful tokens and remove stopwords (Korean stopword list from ranks.nl). Spacing and spelling inconsistencies typical of SNS were addressed to improve analytic reliability. Co-word analysis: For each subset, computed bigram co-occurrence matrices based on adjacent token pairs within sentences. Visualized networks in Gephi using Force Atlas layout. Nodes represent tokens; edges represent co-occurrence strength. Applied frequency thresholds to enhance interpretability following prior large-scale visualization practice. For Everyday-Life, threshold = 69 co-occurrences, resulting network: 247 nodes (2.54% of pre-threshold nodes) and 1,014 edges (0.24% of links). Nodes were translated to English for presentation. The network modularity revealed two main clusters. DMR topic modeling: Applied MALLET to the Everyday-Life subset, conditioning on posted date as metadata to assess temporal variation in topic proportions. Compared models with 10–15 topics using interpretability and perplexity; selected a 13-topic model grouped into five higher-order themes based on thematic commonalities. ANN and BIZ subsets were not modeled due to smaller size relative to typical DMR applications. Temporal analyses examined daily topic proportion trajectories for selected themes (e.g., fear vs. self-care; fear vs. recognition of external support).

Temporal posting patterns: Daily posts exhibited two spikes aligned with outbreak events: May 10–13 (n=3,403 posts) coinciding with the Itaewon nightclub cluster; May 28–30 (n=2,650 posts) related to infections among couriers at a large delivery company. Posts declined after May 30 as daily new cases fell below 30. This suggests increased posting activity as a means to regulate heightened fear and anxiety. Image descriptors: Across ANN, BIZ, and Everyday-Life, “outdoor” was most frequent, followed by nature-related terms (tree, sky, nature). “Indoor” with eating-related terms (food, drink, table) was also frequent. “Child” appeared among top descriptors in ANN and Everyday-Life, reflecting salience of child-rearing and education concerns during school closures. Co-word network (Everyday-Life): Two major clusters emerged. - Cluster 1 (centered on COVID-19/social distancing/mask; daily life/today): Tokens reflected daily life management and resilience—maintaining routines (working out, play, school opening), staying connected (communication, gatherings), following safety measures (prevention, overcome), family/parenting and concerns of working mothers, compassion for others (friends, medical staff). Emotional expressions skewed positive (happiness, love, healing, gratitude). Time/location tokens referenced May family events (weekends, Parents’ Day) and venues (famous restaurants, cafés). - Cluster 2 (centered on Itaewon, nightclub, Gangnam, Hongdae): Tokens reflected fear of spread, with strong references to entertainment venues (nightclub n≈1,070; Gangnam n≈674; Hongdae n≈578; bars/cafés/pubs/karaoke combined n≈1,630) and prevention measures/damage control (quarantine, testing, epidemiology, disinfection, prohibitions). Mentions of “identity” linked the nightclub to LGBTQ patrons, indicating stigmatization within discourse. DMR topic modeling (Everyday-Life): Thirteen topics grouped into five themes with example topic proportions: - Theme 1 Self-care strategies (e.g., prevention/disinfection, psychological well-being resources, health/diet; proportions approx. 0.14, 0.08, 0.05, respectively). - Theme 2 Fear of re-proliferation and worries about children’s education (child-rearing/education concerns, blame of misbehaviors incl. Itaewon, karaoke, Shincheonji, LGBTQ references; proportions approx. 0.10 and 0.06). - Theme 3 Recognition of external support (medical personnel commitment ~0.09; responses from government/experts ~0.06). - Theme 4 Recuperation from intense distancing (mood-improving activities ~0.08). - Theme 5 Daily life management during sustained distancing (partial recuperation ~0.06; daily lives ~0.08; family events/travel ~0.07; weekends ~0.061; anniversaries ~0.059). Temporal dynamics of themes: - Fear (Theme 2) peaked around May 10–13 (Itaewon case). Self-care (Theme 1) tended to increase after fear subsided, suggesting reactive coping. - Early period (May 6–26): Recognition of external support (Theme 3) tracked with fear. Later period (May 27–June 16): Recognition decoupled from fear; discourse shifted toward self-reliance, recuperation, and rebuilding daily life (Themes 1, 4, 5) despite persistent worries. Social implications: Discourse sometimes reproduced stigmatizing frames, notably blaming LGBTQ communities associated with the Itaewon cluster, reflecting and amplifying existing intolerance. Counter-voices raised human rights concerns and potential deterrent effects of identity disclosure on testing. Overall, positive emotion sharing and self-care strategies coexisted with fear and occasional othering.

The findings address the research questions by showing that during sustained social distancing, South Korean Instagram users co-constructed two dominant discourses: (1) resilience and daily life management through self-care, family engagement, and positive reappraisal; and (2) fear of infection spikes that sometimes translated into blame and stigmatization of specific groups and venues. Temporal analyses reveal that heightened fear coincided with outbreak news, spurring social sharing; as fear abated, posts increasingly emphasized self-care and rebuilding routines, while reliance on institutional support diminished over time. This underscores SNS’s dual role: a venue for emotional regulation and community cohesion via positive sharing, and a conduit for the circulation of hate frames and othering during heightened threat. The results are relevant to public health communication and social policy, suggesting the need to harness SNS for prosocial messaging and to mitigate discriminatory discourse, while supporting individuals’ long-term psychological needs as the crisis persists.

By combining co-word analysis and DMR topic modeling of 8,241 Instagram posts, the study shows that South Koreans sustained daily routines and regulated emotions during COVID-19 through self-care and positive sharing, yet fear spikes led to increased posting and occasional stigmatization of minority groups (e.g., LGBTQ persons) linked to outbreak venues. As the pandemic continued, discourse shifted toward self-directed recuperation and daily-life rebuilding with reduced emphasis on institutional support. The authors recommend deliberate social measures to promote inclusion, diversity, and community, counter hateful frames on SNS, and implement long-term, large-scale psychological support systems (e.g., expanded mental health services, online/visiting counseling, community-based programs). Future work should triangulate SNS analyses with qualitative data capturing participants’ voices, focus on specific populations (e.g., medical workers, working mothers, students), and examine cross-cultural differences in responses and coping.

Findings may not generalize to all South Koreans due to platform and sampling constraints: Instagram users (e.g., elderly populations) may be underrepresented; posts lacking the target hashtags were excluded; and SNS-derived perceptions are anchored to specific time points and events. Analyses relied on textual content without direct participant interviews, limiting depth on intentions and context. ANN and BIZ subsets were not topic-modeled due to size constraints. Future research should integrate SNS data with qualitative methods (interviews, focus groups), target specific demographic/occupational groups for tailored insights, and conduct cross-cultural comparisons to assess variability in discourse and coping strategies.