Sudden Attention Shifts on Wikipedia During the COVID-19 Crisis

• Wikipedia logs offer a large-scale, immediate trace of information seeking, while survey-based pandemic research faces challenges of global scale and immediacy.Source: §1 · pp. 1–2
• The analysis combines Wikipedia access logs, Apple and Google mobility reports, and intervention dates, using difference-in-differences regression to address trends and seasonality.Source: §1 · p. 2
• Up to 40%, Wikipedia usage volume rose after mobility fell, but the increase was transient, with no significant long-run increase or a decrease in 11 of 12 languages.Source: §1 · p. 2
• Information-seeking distance from normality rose after restrictions, then partly declined; some of the increase remained after mobility returned to normality.Source: §1 · p. 2
• BIOLOGY had transient volume changes, while VIDEO GAMES gained and SPORT lost attention persistently; persistent increases often concerned entertainment and self-actualization.Source: §1 · p. 2

2 Related work

• The review covers the COVID-19 infodemic, pandemic-era user behavior, and Wikipedia research on crises and pandemics.Source: §2 · p. 2

2.1 The COVID-19 Infodemic

• Infodemic research measured misinformation and conspiracy content, whereas this study asks whether digital traces can reveal behavioral change.Source: §2.1 · pp. 2–3

2.2 The Web in Times of COVID-19

• Related work documented residential Internet traffic increases and elevated Bing search activity, with basic-need subcategories receiving the largest boost (Suh et al. 2020).Source: §2.2 · p. 3

2.3 Wikipedia in Times of Crisis

• Prior work used Wikipedia pageviews to monitor disease and infectious-disease anxiety (Generous et al 2014a,b; McIver and Brownstein 2014; Tausczik et al. 2012; Tizzoni et al. 2020); this study notes difference-in-differences may help account for seasonality.Source: §2.3 · p. 3

3.1 Data

• Pageviews from 12 language editions (1 January 2018–31 July 2020) were enriched with 57 article topics and COVID-19-related labels.The topic model was developed for English articles and linked across the studied languages; articles could belong to multiple topics.Source: §3.1 · pp. 3–4
• Because country-specific events did not affect movement uniformly and some were absent, Apple and Google mobility series supplied heuristic dates for sharp movement changes.The reports track percentage changes in time spent across location types against a pre-pandemic baseline; multiple changepoint algorithms and hyperparameters were averaged for a robust estimate.Source: §3.1 · p. 4, Figure 2 · p. 4
• For English, German, and French editions, mobility reports were aggregated across countries where each language is official; normality dates mark mobility staying within 10% of baseline.The figure caption describes weighting country series by native-speaker proportions; the same aggregation scheme was used for normality changepoints.Source: §3.1 · p. 4, Figure 2 · p. 4
• Language editions only approximate countries because any edition can be accessed globally, a particularly important qualification for English.Source: §3.1 · p. 4

3.2 Difference-in-differences

• Difference-in-differences compares the 2020 post-minus-pre change with the corresponding 2019 change to separate treatment effects from concurrent changes, assuming 2020 otherwise resembled 2019.The method is presented as a quasi-experimental approach using observational data and treatment and control groups.Source: §3.2 · pp. 4–5
• Four comparisons examine five-week windows around mobility change and return to normality, similar-mobility periods, and other events such as school closures or lockdown declarations.Where fewer than five post-normality weeks were available, the analysis used data through 31 July 2020 and the corresponding 2019 period.Source: §3.2 · p. 5

3.3 Distance from Normality

• Daily article-view distributions are projected with PCA, then compared by Euclidean distance to the average 2019 attention vector as a measure of distance from normality.The vectors are computed separately for each language edition using all its articles.Source: §3.3 · p. 5

4.1 Shifts in Overall Pageview Volume

• COVID-19-related pageview share rose until mobility changepoints, then tended to decline slightly, with no sharp response when mobility returned to normal.Source: §4.1 · p. 5
• After mobility fell, pageviews rose in nearly all editions, most strongly where lockdowns were strict, then generally returned toward prior-year levels as mobility normalized.Source: §4.1 · pp. 5–6
• Serbian Wikipedia accumulated about 57 days’ worth of surplus pageviews by late July, roughly tripling access versus 2019; several other editions also accumulated gains.Source: §4.1 · p. 6
• Most languages showed no long-term pageview effect, while Norwegian and Danish had significant negative effects and Italian had a significant positive effect.Source: §4.1 · p. 8

4.2 Shifts in Information Seeking Patterns

• As mobility normalized, distance from normality fell but remained above baseline in eight languages, and attention shifts also appeared where total pageviews did not rise.Source: §4.2 · pp. 8–9

4.3 Shifts in Topic-specific Pageview Volume

• Entertainment topics including video games, books, and internet culture had persistent increases, while sports had persistent decreases.Source: Figure 5 · p. 9, §4.3 · p. 10
• Biology pageviews rose to around 109% after mobility fell, declined to around 89% after normality, and had no significant long-term effect.Source: §4.3 · p. 10
• Medicine and health showed no increase following mobility changepoints, with interest shifts often occurring beforehand.Source: §4.3 · p. 10

5 Discussion and conclusions

• The volume of Wikipedia information seeking rose as mobility fell, then returned to prior levels; the nature of content sought remained changed after mobility normalized.The volume increase and content shifts included non-COVID-19-related articles; topic patterns varied, with some transient and others persistent.Source: §5 · p. 10
• The authors interpret lasting changes in content sought as evidence that the crisis’s societal impact extended beyond mobility restrictions to human needs, interests, and concerns.This interpretation concerns the nature of information sought, which did not return to its prior pattern when mobility did.Source: §5 · pp. 10–11
• Persistent increases in VIDEO GAMES and INTERNET CULTURE suggest greater demand for entertainment and arguably self-actualization, while basic-needs interest may also have increased.The authors note that their topic findings do not rule out increases in basic-needs interest; previous work found increases in related queries (Suh et al. 2020).Source: §5 · p. 11
• Wikipedia’s crisis role extends beyond epidemic information: browsing also became more widespread for non-essential needs when other activities were hindered.The authors relate this to Wikipedia’s use for entertainment and self-actualization.Source: §5 · pp. 10–11
• Because the analysis is observational, it cannot conclusively attribute access-pattern changes to mobility restrictions rather than other concurrent pandemic-related changes.Uneven internet access and digital literacy also limit generalization to broad populations, potentially excluding groups with the greatest increases in basic needs.Source: §5 · p. 11
• Future causal work could use COVID-19 as an instrumental variable to study activities people undertake when spending more time at home.The proposed examples include what books people read and what dishes they cook.Source: §5 · p. 11

A Obtaining mobility changepoints

• Google reports daily percentage changes from a five-week weekday-specific baseline; Apple reports relative changes from January 13, 2020, across driving, walking, and transit.Together, the reports provide nine mobility time series.Source: A Obtaining mobility changepoints · p. 12
• For English, French, and German editions, mobility data are aggregated across countries weighted by each language’s share of native speakers there.For example, English combines seven countries and French combines seven; German combines Germany, Austria, and Switzerland.Source: A Obtaining mobility changepoints · p. 12
• Changepoint detection identifies state changes in time series using a segmentation cost function and a search procedure; here it estimates effective lockdown timing from mobility shifts.The analysis uses cost functions and algorithms available in the ruptures package.Source: A Obtaining mobility changepoints · pp. 12–13
• For most countries, mobility changes were clear and different detection methods largely agreed, with standard deviations often below one day.Table 2 reports changepoint results averaged over parameter choices and standard deviations in days.Source: A Obtaining mobility changepoints · p. 13, Table 2 · p. 13