Introduction

The social media service , formerly and popularly known as Twitter, is a microblogging and social networking service. Users of the service can post, interact with, and re-post messages generally referred to as tweets. The service was first established in 2006 and as of 2023 there were over 550 million monthly active users worldwide1. Much attention has been given to the capacity for the platform to be leveraged as an outreach and science communication tool, including in the geosciences2,3,4 and more specifically, volcanology5,6. Twitter is oft-lauded as an effective platform for dissemination of geoscience research and information4,7,8—possibly because academics are more likely than average to use the service3. Geohazard discourse analyses tend to be event-specific and defined a posteriori (e.g. ref.4); the same is true for volcanic eruptions, with social sensing applied to Twitter datasets related to the 2018 eruption of Kīlauea, USA9, and eruption sequences of Fagradalsfjall, Iceland10. Nevertheless, the full potential of Twitter-derived datasets remains underexplored, despite repeated calls for its use3,5,11. With 40–50 actively erupting volcanoes at any given time around the globe, there exists enormous potential for longitudinal analysis of Twitter-derived timeseries. Moreover, with a global userbase and multilingual support, the platform represents an invaluable dataset with which to explore potential differences and commonalities of online discourse as a function of language. This contribution addresses both these gaps through a longitudinal study of volcano-centric keywords in 18 languages covering a period of almost four years.

Event spikes correspond to real-world events

Figure 1 presents subsets of the tweet anomaly score α for each language (Methods; full timeseries presented in Supplementary Fig. 1). A peak-matching algorithm (Methods) corroborates that local maxima correspond to detectable peaks above the baseline and represent discrete temporal events. Comparison of peaks in α with Global Volcanism Program bulletin reports12 and news resources identified potential real-world events, confirmed via keyword searches of the text content. This was performed manually to ensure independent verification of eruptions or related activity; however, in the future—coupled with compilation of a database of verifiable sources—a fuzzy temporal and semantic matching technique or event-based information retrieval approach could facilitate automated peak attribution. Selected volcanic eruptions are labelled on Fig. 1, demonstrating that public attention to volcanic hazard often spikes during and immediately after volcanic events. However, spikes in α can also represent non-event-specific uses of geoscience terms and not necessarily a discrete volcanic phenomenon (see Supplementary Text 1), especially when observed in only one language, underscoring the benefit of multilingual analysis. The presence, timing, and relative magnitude of peaks in α can vary substantially across languages. For example, the eruptions of Whakaari (2019; Aotearoa New Zealand) and Nyiragongo (2021; DRC) are both detectable in multiple languages, but dissemination patterns differ. In the former, spikes in α are essentially simultaneous across languages (Fig. 1a), whereas for Nyiragongo a lag of up to three days can be seen (most notably in languages with lower overall post volumes, e.g. Icelandic, Italian, Thai, Arabic, Korean, and Amharic; Fig. 1d). Further interlingual discrepancies are discussed further in 'Linguistic imbalance'.

Fig. 1: Timeseries of heightened Twitter activity with respect to major volcanic events.
figure 1

Each panel comprises 31 days of the anomaly score α for each of the 18 studied strings (see Table 1). Timing of notable volcanic eruptions are highlighted by arrows: a Whakaari (Aotearoa New Zealand); b Taal (Philippines); c Fagradjallsfjall (Iceland); d Nyiragongo (Democratic Republic of the Congo); e Tajogaite (Cumbre Vieja, Spain); f Gunung Semeru (Indonesia); g Hunga Tonga–Hunga Haʻapai (Tonga); h Sakurajima (Japan) and Fagradjallsfjall; i Sheveluch (Russia).

Full size image

Tonga–Hunga Haʻapai: the eruption heard around the world

On 15 January 2022, Hunga volcano—a largely submerged caldera volcano located in the southwest Pacific—underwent the most explosive eruption of the satellite era13. The explosion was the culmination of a sequence of volcanic activity beginning in December 2021, characterised by Surtseyan activity, pyroclastic phenomena, and ash emission. The Hunga Tonga-Hunga Haʻapai (HTHH) explosion was extraordinary in many ways, with myriad global impacts. The eruption plume penetrated the mesosphere, and the explosion was associated with remarkable propagation of tsunami14, sound15,16, and infrasonic atmospheric17 waves. The eruption caused widespread devastation across the nearby Tongan islands, destroying infrastructure, severing international communications with the Kingdom of Tonga14, and causing substantial environmental damage. The extraordinary nature of the eruption is reflected in the tweet anomaly (Fig. 1g) and count data (Methods): for 16/18 of the considered strings, January 2020 appears in the top five SM values. The same is true for 15/18 of the top five MaxM values. For half of the strings (9/18), January 2022 exhibits the highest SM and MaxM across the whole timeseries, with peaks in daily counts on the 15th, 16th, or 17th of that month. Interestingly, the maximum daily count \({{{{\rm{Max}}}}}_{M}\) in 〈eruzione vulcanica〉 [Italian] occurred on 15th January 2023 (i.e. the one-year anniversary of the eruption), thanks in large part to a multi-part summary posted by Il Mondo dei Terremoti, which outlined major scientific discoveries catalysed by the HTHH eruption (Il Mondo dei Terremoti—@mondoterremoti—is an account dedicated to disseminating news and articles concerning earthquakes and volcanoes). In short, the HTHH eruption dominated Twitter discourse across multiple languages. The exceptions are Amharic and Icelandic (Fig. 1g), discussed further in 'Linguistic imbalance'.

Step-changes in background chatter

To further explore how specific eruptions shape longer-term discourse on Twitter, select timeseries are segmented using the Pruned Exact Linear Time algorithm (Methods). Figure 2a shows the segmented timeseries for 〈eldgos〉 [Icelandic]. Over the data collection period, volcanic-tectonic reactivation of the Reykjanes Peninsula (Iceland) commenced December 2019, and two eruption sequences occurred in that region (March 2021 and August 2022). As with the HTHH eruption, the Fagradalsfjall eruption in March 2021 was largely unanticipated, ending ~800 years of regional quiescence18. The unrest period in 2019–2020 and both eruptions were accompanied by peaks in α (Fig. 1c,h); an additional precursory peak in 2021 reflects reports of increased regional seismicity19. The 2021 eruption in particular triggered extensive online public discourse, in large part due to the accessibility of the eruption site and the spectacular nature of the eruptive phenomena19. For comparison, between 2019-08-25 and 2019-12-31, only 18 posts containing 〈eldgos〉 were identified (not all of which were related directly to volcanic activity), whereas during the height of volcanic activity in March 2021, as many as 1331 posts per day contained this string. This is in spite of notable volcanic eruptions occurring elsewhere in the world during late 2019, such as the fatal eruption of Whakaari in December. The 7-day rolling mean count data for 〈eldgos〉 is shown in Fig. 2a. Each shaded window is determined to have internally consistent statistical properties, and each transition between consecutive segments is a PELT-defined changepoint. Also highlighted in Fig. 2a are the 'quiescent' periods, i.e. periods of background Twitter activity not directly associated with volcanic-tectonic activity (labelled A–D). The transitions A | B, B | C, and C | D demarcate the Reykjanes volcanic-tectonic reactivation, the March 2021 eruption, and the August 2022 eruption, respectively. Parametric and non-parametric tests (Methods) confirm that the consecutive quiescent periods (A, B, etc.) are statistically different to each other: there is a detectable change in mean background after each event. The statistically significant increase in the background level of posts containing 〈eldgos〉 seems to be indicative of a heightened level of public awareness of volcanic phenomena in the months following reactivation of the Peninsula and the 2021 eruption. Natural language processing has revealed that related expressed emotions in social media discourse during the pre- and post-eruptive stages were measurably and significantly different10. However, a different pattern is observable for the 2022 eruption: mean background decreases following the event. Tellingly, the bulk of online discourse happened after the eruption in 2021, but before the eruption in the case of the 2022 event (cf. Fig. 1c, h). This likely reflects a shift towards anticipatory discourse in 2022, given how recent the previous eruption was and the fact that the details of the 2021 eruption were still very much in the public consciousness. However, the decreased total volume of posts corresponding to either eruption, coupled with the decrease in background mean following the 2022 event, suggests that the perceived newsworthiness of the Fagradallsfjall eruptions diminished as they became less novel.

Fig. 2: Statistically segmented timeseries of mean scaled post counts.
figure 2

Timeseries for 〈eldgos〉 and 〈erupción volcánica〉 are shown in (a and b), respectively. Each consecutive shaded region has internally consistent statistical properties, and each transition between segments is a statistically determined changepoint. The mean value of each segment is annotated (2 s.f.). Note logarithmic y-axes. 'Quiescent' periods are denoted by A, B, etc. The value n indicates the number of segments determined for a given β value.

Full size image

Step-change patterns of increased background can be identified in other strings as well. The example of 〈erupción volcánica〉 [Spanish] is given in Fig. 2b. In this case, the transition A | B (corresponding to the 2020 Anak Krakatau eruption, Indonesia) initiates a detectable increase in background activity, but this is not statistically significant (Methods). The transition B | C, however—the eruption of Tajogaite—represents a significant step change in online activity, which persists throughout the rest of the timeseries (means of segments C–E are consistently higher than the segments preceding the Tajogaite eruption, Fig. 2b).

Linguistic imbalance

The volume of posts over the studied timeframe varies by several orders of magnitude as a function of language (Table 1). By way of comparison, eruption-related posts in Japanese rarely if ever exceed 100 per month, whereas in English there are rarely if ever fewer than 100 per month. In general, the higher the volume of posts, the more closely the monthly (30-day) running mean count dataset approaches a lognormal distribution, for which the distribution mean \(\left\langle c\right\rangle\) can be determined. The strings 〈éruption volcanique〉 [French], 〈erupción volcánica〉 [Spanish], and 〈volcanic eruption〉 [English] yielded the highest distribution means (\(\left\langle c\right\rangle\) = 22, 56, and 318, respectively), whereas all other strings exhibit \(\left\langle c\right\rangle\) < 10. Distributions are shown in Fig. 3. There is no evidence of strongly right-skewed distributions which might imply a steady, sustained interest in eruption-related topics. In general, the tendency towards log-normal distribution with increasing userbase (i.e. Fig. 3a–r) points to an evolution from a decentralised, community-driven network towards a centralised, 'influencer-driven' network. This has implications for the 'top-down' dissemination of volcano-relevant information—it is likely that most of the online discourse in the larger, more connected communities is controlled by a relatively small number of users20. In this context, the role of 'verified' (i.e. independently authenticated) users in information transfer could be a valuable avenue for further study21.

Fig. 3: Post count distribution broken down by language.
figure 3

Histograms of 30-day mean counts of each timeseries are shown in (a–r). Axes are the same for all panels. Black crosses show mean \(\left\langle c\right\rangle\) and 2 standard deviation range calculated assuming a lognormal distribution (not performed for (a, b), and (f) which have visibly different distributions). The number n of posts in the dataset is also shown.

Full size image
Table 1 List of studied strings in 18 languages
Full size table

Each of the time series is different (Fig. 1); however, some differ more than others. This is shown quantitatively in Fig. 4a—the cosine distance matrix for all timeseries, ordered using a hierarchical clustering approach (Methods). A small distance between two languages means that their respective timeseries are close in terms of the timing and relative magnitude of peaks. A score of zero would imply that every event observed in timeseries X is reflected in timeseries Y, and has the same degree of perceived importance. The higher the distance, the less likely an event (i.e. an eruption or other volcanic event) represented in one timeseries is equally represented in the other. I use the cosine distance data to define four primary relations between timeseries, referred to here as [type 1] full synchronisation; [2] one-way coupling; [3] common mode response; and [4] independence.

Fig. 4: Time series differencing.
figure 4

a Heatmap of pairwise cosine distance dcos for each string. Darker colors indicate high similarity between timeseries, while lighter colours represent low similarity. Grey boxes highlight pairs shown in more detail in (c–g). b Dendrogram plot for (a), ordered so that similar timeseries are clustered. c–g Mirror plots comparing select (X, Y) pairs of strings, typical of the timeseries types 1, 2, 3.1, 3.2, and 4, respectively. Peaks in the normalised weekly average timeseries are highlighted for illustrative purposes, indicating where a given event is reflected in one or both of X and Y. Note that logarithmic y-axis is positive in either direction. The full set of pairwise mirror plots (153 pairs) can be seen in code provided at https://doi.org/10.5281/zenodo.1600858753.

Full size image

Full synchronisation comprises timeseries pairs characterised by very low distance scores (hence high similarity), reflecting synchronised information flow to and between languages and implying a high degree of overlap in culture and social media platform use. A type example is German/Dutch (Fig. 4c).

One-way coupling exists when most or all events observed in one timeseries are reflected in a second timeseries, but not vice versa. The example of English/Icelandic is shown in Fig. 4d: the set of events detectable in the English timeseries is a superset of those detectable in Icelandic (i.e. there are no 〈eldgos〉 events not detected in 〈volcanic eruption〉, but there are many 〈volcanic eruption〉 events not detected in 〈eldgos〉).

Two patterns of common mode response can be discerned, largely depending on the post volume: [type 3.1] nominally independent signals punctuated by common response to external events (i.e. internationally newsworthy eruptions); and [3.2] signals consisting almost solely of responses to external (i.e. international) events. In the case of the former, the similarity between timeseries is not necessarily strong. Tweet strings corresponding to notable 'local' eruptions in the dataset tend to exhibit this characteristic (e.g. 〈erupción volcánica〉 [Spanish], 〈éruption volcanique〉 [French], 〈eldgos〉 [Icelandic], 〈erupsi vulkanik〉 [Indonesian]). Contrastingly, in the case of the latter (response signals only), timeseries tend to exhibit strong similarity, because the common responses comprise a major proportion of all detected events. These strings tend to be characterised both by low post volumes and non-Latin character alphabets: e.g. 〈ภูเขาไฟระเบิด〉 [Thai], 〈ثوران بركاني〉 [Arabic], 〈화산 분출〉 [Korean], 〈火山噴火〉 [Japanese], and 〈火山爆发〉 [Chinese (Simplified)]. Tellingly, these exhibit strong similarity to each other and also to 〈volcanic eruption〉 [English], indicating a leader-follower dynamic. Examples are shown in Fig. 4e, f.

Finally, full independence or decoupling can be observed between strings corresponding to low overall post volumes, reflecting weakly connected networks where couplings are too sparse to sustain full or partial synchronisation (Fig. 4g). Unsurprisingly, with only 6 datapoints የእሳተ ገሞራ ፍንዳታ [Amharic] is not highly similar to any others, although there is slight agreement with 〈éruption volcanique〉 [French] (Fig. 4a), corresponding to the 2021 Nyiragongo eruption in DRC. The highest distance scores are interpreted as representing isolated information ecosystems, due to differences in local event framing and/or data scarcity. This can be due to low Twitter penetration in certain countries or the predominance of other social media platforms22, which can be both a function of internet access and sociopolitical factors. For example, Russian tends to exhibit a high distance score relative to other languages in this study (Fig. 4a), certainly driven to some extent by the fact that Twitter has been blocked in Russia since March 2022, which is manifest in a marked decrease in daily activity from that date onwards (Fig. 1).

As well as spurious peaks in online activity (Supplementary Text 1), there are notable omissions. The deadliest volcanic eruption during the data collection timeframe—the December 2021 eruption of Gunung Semeru in Indonesia—had a death toll of at least 57 (ref. 23), with over a hundred people injured and multiple people still unaccounted for24. The destabilisation of the lava dome, attributed to heavy rainfall25, triggered a series of pyroclastic density currents and lahars which damaged over 5000 homes and other infrastructure. Only in one language—Korean 〈화산 분출〉—was December 2021 in the top 5 months with highest SM and MaxM values (see also Fig. 1f). Perhaps surprisingly, December 2021 does not stand out in the 〈erupsi vulkanik〉 [Indonesian] or 〈letusan gunung berapi〉 [Malay] datasets, whereas months corresponding to the eruptions of HTHH, Fagradalsfjall, and Whakaari do (Fig. 1a, c, g). Keyword text searches of the timeseries do reveal that the eruption prompted online activity, but (as visible in Fig. 1f), this was far lower and the signal less impulsive than other eruptions (e.g. Whakaari, Fig. 1a; Taal, Fig. 1b; Tajogaite, Fig. 1e). Similar discrepancies are observable for the July 2022 eruption of Sakurajima, Japan. This eruption—characterised by a large explosion, ash plume, and ballistic ejecta—prompted an evacuation of local residents, and the volcanic alert level was raised to its maximum (the first time this had been issued for Sakurajima). Spikes in α can be observed in 〈火山噴火〉 [Japanese], 〈火山爆发〉 [Chinese (Simplified)], 〈извержение вулкана〉 [Russian], 〈화산 분출〉 [Korean], and 〈ภูเขาไฟระเบิด〉 [Thai] (see Fig. 1h). However, the event did not trigger a substantial α anomaly further afield (Fig. 1h), nor was it reflected in SM or MaxM data. As a final example, the VEI 4 eruption of Sheveluch, Russia, prompted some online discourse, primarily in English but not particularly in Russian (due to the Twitter ban) or other languages (Fig. 1i). This is despite being the largest Sheveluch eruption in decades26: an ash plume ~20 km high blanketed villages in as much as 85 mm of ash and prompted code-Red aviation warnings, while lavas coursed down the volcano’s flanks and melted large volumes of snow. Local water supplies were polluted, schools were closed, and residents ordered to stay indoors27. While these seeming omissions may in large part be due to geographic differences in social media preferences—other social media platforms may have been used more frequently during eruptive crises—these observations suggest a geographic and linguistic divide in how volcanic eruptions are discussed online. It would be valuable to explore event-response trends in data collected from other social media platforms, such as VK, Facebook, Telegram, or QQ.

Implications and perspectives

This contribution demonstrates that volcanic eruptions can be detected in near-real time by parsing social media data using pre-defined search terms. The anomaly score (Fig. 1) thus provides a near-real-time indicator of notable volcanic activity. Although by no means a perfect proxy for real-world volcanic activity, potentially prone to spatial bias28 and misinformation5, it may have promise as a supplementary monitoring tool for areas where more traditional data sources have low coverage or are subject to transmission delays. While it is important to note that many of the spikes in online discourse observed are triggered by breaking news articles or emergency bulletins, it has been previously noted29 that Twitter is often one of the earliest sources for eruption images, typically shared by the public (rather than news organisations or observatories). Local reports often appear on Twitter prior to traditional news media, which is particularly pronounced if the reports are not in English29. Moreover, it provides a dataset for gauging public perception of and reaction to volcanic events—particularly valuable for crisis response and public communication plans6. In this context, recent work has been done to facilitate the integration of social media data into geohazard crisis management workflows30, an important step for capitalising on such data in a systematic and operational way.

The striking dominance of the HTHH eruption in Twitter discourse relative to other eruptions detected in the datasets (Fig. 1) reveals how unprecedented and unanticipated events can become newsworthy by the very fact of their magnitude and rarity. However, the duration of an event relative to the news cycle is also key31: cataclysmic eruptions that occur at nominally discrete points in time are more likely to be recorded as 'news' than protracted periods of volcanic unrest. This also implies that eruptions take priority in online discourse over other related processes that take place over extended periods of time, such as building resilient communities, developing and maintaining hazard mitigation infrastructure and protocols, or cleaning, rebuilding, and relocation in the aftermath of a damaging eruption.

The increase in background activity following notable eruptions (Fig. 2) may indicate a fertile online environment for disseminating volcano-related information, particularly relevant for increasing awareness about volcanic hazards, improving future eruption preparedness (e.g. by disseminating evacuation procedures), and for strengthening public engagement with academic research. However, the above analysis highlights that the window of increased public engagement can be a limited one—unless actively impacted by an ongoing crisis, attention soon diverts elsewhere. This is a common phenomenon; summarised in 2005 by the then director of Save the Children UK, referring to a series of cascading disasters in Guatemala: 'A country hit by a volcanic eruption, a hurricane, and then devastating mudslides would in any normal week be considered a major emergency. But this is not a normal week…'32. Despite the mudslides displacing 120,000 people, the event was rapidly—within hours—supplanted in the news cycle by another of even more devastating proportions: the 2005 Kashmir earthquake. Thus, an accident of timing can lead to so-called 'hidden disasters,'33 where crises do not capture public attention. This effect of attention economy34 has important ramifications for the allocation of international aid.

Media reports on natural disasters have a first-order systematic impact on the flow of international disaster aid, and—critically—have been found to be subject to a distance bias35,36. It is understandable that in many cases, 'local' eruptions dominate the online discourse in a given language. A clear example is Icelandic, where the top two SM and MaxM values are March 2021 and August 2022—corresponding to the dates of the two Fagradalsfjall eruptions in the studied timeframe (see Fig. 1c,h). Similarly, the eruption of Tajogaite (Spain; Fig. 1e) in September 2021 drove maximum Twitter discourse in Spanish, and the same pattern is observed for the May 2021 eruption of Nyiragongo (DRC; Fig. 1d) in French. However, not every language represented corresponds to a volcanically active country or region within the constraints of the timeseries; nevertheless, 'international' eruptions can gain prominence in online discourse. Developed for the representation of geopolitical events in news media, the theoretical frameworks of Galtung and Ruge31 can provide valuable insight into these observations. Firstly, the concept of cultural proximity: attention is more likely to be afforded to events that occur in countries that are more closely culturally linked than to events that occur elsewhere. This often overlaps with spatial proximity37 (e.g. the eruption of Sakurajima, Japan, is reported in Russian and Korean, but not Italian or Spanish: Fig. 1h), but also includes former colonies (e.g. the eruption of Semeru, Indonesia, drove an α spike in Dutch: Fig. 1f). Closely connected to cultural proximity is event relevance. For example, the 2019 Whakaari (Aotearoa, New Zealand) eruption is detectable in both the Chinese and Malaysian datasets: in both 〈火山爆发〉 [Chinese (Simplified)] and 〈letusan gunung berapi〉 [Malay], December 2019 stands out as one of the top 5 months with highest SM and MaxM values. The eruption caused the death of 22 tourists and a tour guide who were on the island at the time; although New Zealand, Malaysia, and China are not necessarily culturally proximal, there were Chinese and Malaysian nationals among the fatalities.

As a strategy for retaining public interest in volcanic hazards, it is useful to look at the dataset for 〈eruzione vulcanica〉 [Italian]. As noted in 'Event spikes correspond to real-world events', the maximum α, SM, and MaxM for 〈eruzione vulcanica〉 correspond to the one-year anniversary of the HTHH eruption. This enhanced online discourse was precipitated by a series of science outreach posts, underscoring the importance of structured communication in determining long-term engagement. This may be a particularly important strategy for disseminating information related to public awareness of and preparedness for volcanic hazards6, especially regarding processes that occur over longer times (like those mentioned above). This emphasises the importance of official social media accounts with a large follower base providing timely and accurate information to the public5,38,39; not least, information must be easy to understand, a concept well-summarised as: 'The remote and the strange will at least have to be simple if it is to make the news'31. Effective use of imagery19 and data visualisation40 are essential tools here.

Harnessing massive datasets generated through social media offers a potentially invaluable tool for detecting eruptions and gauging public engagement in near-real time. However, its effectiveness in both short-term crisis response and longer-term hazard awareness and preparedness ultimately relies on strategic and structured communication, cross-language outreach, and an understanding of the 'economics of attention'41. This research reveals substantial linguistic imbalance in eruption-related discourse: English dominates, even regarding eruptions in countries where English is not the main language. As a result, not all eruptions result in equal online traffic, even if the characteristics of the eruptions are themselves similar. The huge discrepancies in information transfer between languages highlight an important and evolving role for multilingual information dissemination, potentially including human- or machine-based translation and reposting on social media platforms.

Previous research has analysed Twitter-derived datasets using sentiment analysis (natural language processing) in order to determine the emotional tone of discourse related to specific volcanic events9,10 or related to a specific geoscience user account42. This approach requires the construction of sentiment codebooks tailored to specific datasets (corpora), but bilingual or multilingual approaches remain rare10. The majority of tools and annotated corpora are developed in English, with limited equivalents in other languages43. Machine translation-based methods are commonly employed44, but introduce semantic drift and cultural loss (i.e. change or loss of meaning, nuance, and cultural context due to translation). With either approach, there are challenges due to the wide variability of data availability between languages, compounded by contextual and syntactic differences. Given these methodological barriers, sentiment analysis is not performed in this study; however, the datasets offer clear potential for future investigation. In particular, there is considerable scope for the use of semantic and relational retrieval and analysis methods, including embedding-based search45 and hashtag co-occurrence analysis46.

It is worth noting that any given post could contain text in multiple languages (a concept referred to as code-switching43), meaning that a single post could appear in more than one language dataset. Further, the analyses shown here treat all posts as unique and equal, ignoring the capacity for users to re-post content originally posted by themselves or others (typically referred to as a 'retweet'). This possibility introduces further complexity regarding the flow of information reflected in these data, and suggests that analysing these datasets as digraphs (in a graph theory sense) could provide further valuable insights47.

It is unfortunate that during the course of this research, Twitter underwent notable and often negative changes, both from a technical perspective (e.g. data accessibility has been greatly reduced for researchers) and in terms of the userbase (including a reported 23% decrease in users after its sale and acquisition in 202242 and migration of academic users to other social media platforms48,49). Given the lack of clarity as to whether free API access will be restored for research or operational purposes, it is of increasing importance to explore alternative social media platforms. Both the results and methods outlined in this study are transferable and can be adapted to other platforms to continue to develop a more comprehensive picture regarding the dissemination of volcano-relevant information across different languages and regions. This would also provide an opportunity to capture more detail regarding countries with relatively low Twitter penetration but high usage of alternative platforms.

As a final note, these results serve as a call for volcanologists to navigate beyond the silos of their own languages. This may well include active engagement with diverse linguistic groups through reading and writing multilingual publications50, pursuing international collaboration, or by leveraging machine-translation tools to consume and disseminate information across language boundaries, in a truly global sense. Such efforts would ensure that hazard communication is more inclusive and knowledge more widely distributed, ultimately building towards better community resilience in the face of volcanic risks.

Methods

Data collection

Initial search strings were compiled from twenty languages, chosen based on the most common languages used on Twitter, official languages (where they exist) of the top ten countries by annual tweet volume, and the countries with the most volcanoes. This study focusses on the string 〈volcanic eruption〉 and its translated equivalent in other languages. A combination of manual and machine translation was used to determine the initial keywords, with an initial check to ensure that historical posts existed containing those strings. Not all searches yielded new results over the data collection period (strings in Tok Pisin and Afrikaans): the final list of strings and languages is provided in Table 1. The languages included here are English, Japanese, Spanish, Malay, Portuguese, Arabic, French, Turkish, Thai, Korean, Chinese, German, Dutch, Icelandic, Russian, Italian, Indonesian, and Amharic. Note that there is some natural overlap, in particular for closely related languages such as Malay and Bahasa Indonesian. While Table 1 undoubtedly does not account for all potentially relevant strings, it encompasses a broad range of volcano-focussed Twitter activity, not only from countries and languages with relatively prolific tweet volumes but also from those where Twitter usage is less common.

Using Twitter’s standard search application programming interface (API), tweets containing any of the strings listed in Table 1 were crawled and downloaded by executing a Python script every 24 hours (see Supplementary Information Listing 1). The standard API mechanism is restricted in both the number of tweets that can be crawled and the timespan over which this can be performed, which has imposed limitations in previous studies, which look at timeseries of days to a few months51,52. These roadblocks were circumvented by automatically pausing the execution for 15 minutes whenever the search number threshold was reached. This procedure was iterated every day between 25 August 2019 and 12 June 2023 (3 years, 9 months, 19 days). If for any reason, the script could not be executed with a 24 h period, the searched timeframe was increased in the subsequent run, up to a maximum of the API-imposed limit of seven days. Note that free use of the API was discontinued in June 2023 as a consequence of change of ownership of the social media service. Each tweet object retrieved—in JavaScript Object Notation (JSON) format—has a variety of root-level and child attributes, including a unique identifier, a timestamp, and text content. If users have enabled geotagging, then location information can also be scraped from the JSON data. Only the timestamp, text, and URL data are used in this study.

The primary data type used here is the timeseries of the daily number of tweets containing a given string. In order to compare multilingual strings with greatly differing volumes of data, a tweet anomaly score α is defined, such that \(\alpha ={c}_{i}/(m\bar{c}+b)\), where \({c}_{i}\) is the number of tweets at date index i, and \(\bar{c}=\frac{1}{\left|w\right|}{\sum}_{i\in w}{c}_{i}\) is the running mean such that the number of date indices |w| within the window w is nominally 7 (i.e. one week) unless there are missing or irregular data; accordingly, \(w=[i-6,i]\). Additionally, \(m=1.25\) and \(b=5\) are scaling parameters. The variable α illustrates deviations of the volume of tweets from the running mean of any given timeseries, thereby giving a measure of heightened Twitter activity, which can be cross-referenced against known events. Two additional metrics are also defined here: the monthly sum of counts, \({S}_{M}={\sum}_{i\in M}{c}_{i}\), and the maximum daily count for month M, given by \({{{{\rm{Max}}}}}_{M}={\max }_{i\in M}{c}_{i}\). Data for the top N maximum values are available in code provided at https://doi.org/10.5281/zenodo.1600858753.

Changepoint analysis

To identify changepoints in timeseries, the Pruned Exact Linear Time algorithm (PELT)54 is implemented. PELT minimises a cost function C, combining the goodness of fit for each potential time segment with a penalty parameter β, which penalises for the number of changepoints (to avoid overfitting):

$${{{\rm{C}}}}\left({{{\rm{y}}}},{{{{\rm{t}}}}}_{1},{{{{\rm{t}}}}}_{2},\ldots ,{{{{\rm{t}}}}}_{{{{\rm{k}}}}}\right)={\sum }_{{{{\rm{i}}}}=1}^{{{{\rm{k}}}}+1}{{{\rm{Cost}}}}\left({{{{\rm{Segment}}}}}_{{{{\rm{i}}}}}\right)+{{{\rm{\beta }}}}{{{\rm{k}}}}$$

where Cost(Segmenti) is a measure of how the a model fits each segment, and t1, t2, …, tk are the potential changepoints. In practice, β needs to be uniquely tuned to each dataset as a function of background variability and data volume: there is no single value which allows effective retrieval of changepoints across all strings. In future work, this could be done by using part of the timeseries as a training dataset for validation-based tuning or through unsupervised model selection. Nevertheless, it is illustrative to show some results using boilerplate values of β = 102 and 106 (Fig. 2). Once segments are identified, the mean of each segment is calculated. However, the general patterns described hold for other relevant descriptive statistics (for example variance, standard deviation, median).

Peak matching

A peak matching procedure was carried out on each timeseries using the scipy.signal find_peaks55 algorithm with a prominence parameter of 0.1, after applying a 1-D Gaussian filter and normalising each signal to its maximum across the whole timeframe.

Timeseries differencing

A 1-D Gaussian filter is applied to each of the timeseries of counts per day, followed by z-score normalisation such that each datapoint x is transformed to its z value by z = (x – μ)/σ, where μ and σ are the mean and standard deviation of the smoothed timeseries. The cosine distance dcos(X, Y) is calculated pairwise for each combination of timeseries pairs X and Y according to dcos (X, Y) = 1-(X⋅Y)/‖X‖‖Y‖, where X⋅Y is the dot product, and ‖X‖ and ‖Y‖ are the Euclidean norm of X and Y, respectively. These data define a distance matrix Acos; hierarchical clustering is then performed after ref. 56, computing the pairwise cosine distance between every pair of row vectors from the input matrix, and using this to sort the pairs.

Statistical testing

The data from Fig. 2 are tested for significant differences between consecutive quiescent segments. Both the non-parametric Mann–Whitney U test and a student’s t-test were performed on subsequent segment data. To account for non-normality, the student’s test was performed on the logarithm of the data; further, to handle zero values, an atom value 10−3 was added to the data before transformation. In all cases, statistical significance is ascribed for p-values < 0.01. It is important to note that the data are a 7-day rolling mean which inherently introduces temporal dependence between consecutive data points. This means that the data do not satisfy the independence criterion for either test used; while the Central Limit Theorem provides some robustness given the sample sizes, the lack of independence should nevertheless be considered when interpreting results. Future work could benefit from using a statistical approach explicitly developed for timeseries data.