Content area
X (formerly Twitter) provides a social media platform for teacher collaboration. This study comprehensively maps the educational Twittersphere—sets of education-related communities—in Germany (N = 2,761,579 tweets, N = 143,004 users) between 2007 and 2022, serving as a blueprint for researchers seeking to study informal teacher learning at scale in other national contexts. We introduce a novel, reproducible method to discover, label, and examine national-level teacher communities using large-scale Twitter data. In doing so, we address a significant gap in the literature, as most prior research on online teacher learning has focused on English-speaking contexts, especially the United States. We describe both quantitative and qualitative interaction patterns across subcommunities, including participation and engagement trends of teachers and non-teachers over time. Focusing on Germany’s largest education-related community—Twitter’s Teacher Lounge (#twlz)—we examine how it reflects the characteristics of a Community of Practice (CoP), a model of informal learning with well-established links to effective teacher development. Our findings suggest that teacher participation in online CoPs is widespread, sustained, and characterized by shared domain language, community interaction, and collaborative practice. This evidence contributes novel understanding to why online teacher professional development is effective for teacher learning.
Introduction
The rise of digital technologies led to many changes in education, not only for classroom instruction but also for teacher education. Online professional development enables novel forms of teacher learning activities. A recent meta-analysis suggests that these forms of professional development have medium-sized effects on teacher learning outcomes and small but tangible effects on student outcomes (Morina et al., 2025). Teachers have been increasingly using social media to participate in online professional development activities, especially as early-career teachers build professional networks (Staudt Willet, 2024). To date, many teachers engage in social networking sites (SNS) such as X (formerly known as Twitter), Facebook, Pinterest, or discussion boards.
Twitter is a microblogging service that allows the public broadcasting of short messages (tweets) with up to 280 characters to an audience of followers. After its acquisition by Elon Musk, it expanded its character limit to 4,000 for selected users. These messages may include pictures, videos, polls, and links to external websites. Original tweets can either be further broadcasted to the poster’s followers as is (retweets) or with additional comments (quotes). Furthermore, users can reply to original tweets, retweets, or quotes. Its public and open nature distinguishes Twitter from other SNS, such as Facebook, where teachers typically connect in more closed groups (Bergviken et al., 2018; Bissessar, 2014).
Prior studies indicated that teachers use Twitter to share and access links, resources, and content relevant to their teaching (Aguilar et al., 2021; Bruguera et al., 2019; Greenhow et al., 2020). Teachers also discuss current popular education topics and trends, collaborate with their colleagues, or form new relationships with educators beyond their school or districts (e.g., Carpenter, 2015; Carpenter & Krutka, 2014, 2015a). Teachers often praise Twitters’ real-time and on-demand accessibility and its personalization. Prior research has shown that teachers view Twitter as a supportive environment for professional learning and a form of personalized professional development (Fischer et al., 2019).
Online activities on public social media platforms also have particular affordances for researchers to understand learning and interaction processes that are unreasonably expensive to collect at scale (Fischer et al., 2020). For example, tracing all conversations of teachers participating in week-long face-to-face workshops is neither feasible nor desirable. In contrast, each interaction from teachers on online platforms is time-stamped and traceable, generating millions of potential data points. For Twitter, this includes text data from each tweet with corresponding interaction data on direct communication with users (signaled with the @-sign) and communities (signaled with the #-sign), and the reach of each tweet (i.e., received likes, mentions, and retweets). This big data can be used to understand learning and interaction processes to inform educational policy (Fischer et al., 2020; Siemens & Baker, 2012). For example, early studies that utilized hundreds of thousands of tweets examined conversational patterns of users in teacher communities and the public opinion of teachers regarding large curriculum reforms, such as the Next Generation Science Standards and the Common Core State Standards (Rosenberg et al., 2020, 2021a; Wang & Fikis, 2019).
Although teachers’ participation in Twitter communities is a worldwide phenomenon (Carpenter et al., 2022), much of this work has been situated in the U.S. For non-English speaking countries, research is limited with only a few peer-reviewed articles examining how teachers use Twitter (or other social media platforms) for their professional learning. For example, Rehm and Notten (2016) investigated German teachers’ interaction in the hashtag #EdchatDE, a weekly chat-based community in which teachers discuss the latest trends and developments and share resources for lesson plans. Another study investigated the German Twitterlehrerzimmer (TWLZ; German for Twitter’s teacher lounge) community during the COVID-19-induced shift to emergency distance education (Fütterer et al., 2021). Another exception is a study by Mosquera-Gende et al. (2024) who studied the social dynamics of a Twitter hashtag community related to a Spanish education project using graph analysis methods. While the study confirmed that the community exhibits characteristics typical for affinity spaces, it did not analyze teachers and represents a single example of a teacher Twitter community outside the United States. Finally, Sörensen et al. (2023) found that universities of teacher education in Switzerland are underrepresented on social media compared to other higher education institutions in Switzerland. There is a lack of research that studies how, when, and why teachers using social media for professional development participate in online communities outside the United States, and how researchers can discover those communities at scale. Specifically, past research has not examined how participation on social media is organized in distinct learning communities outside the United States (e.g., Wesely, 2013).
Therefore, this paper attempts to advance research in two directions: First, we provide a blueprint to comprehensively map a national Twittersphere (i.e., sets of education-related communities) to advance research on informal learning spaces. This includes generating country- and language-specific variables not provided by the Twitter API and sharing its reproducible code. Second, we exemplify how such data may be used by investigating how teacher participation on Twitter fulfills criteria of a Community of Practice (CoP; Wenger-Trayner & Wenger-Trayner, 2015), a theoretical model of informal learning that has been empirically related to effective learning processes (Amin & Roberts, 2008; Mittendorff et al., 2006). This may encourage educational stakeholders to recognize participation in social media communities as a viable form of informal online professional development (OPD).
Research questions
This study applies a data science perspective to identify CoP characteristics at scale. We hope to increase the generalizability of past findings due to the substantially larger sample sizes, while also increasing its validity of past findings, for instance, by using behavioral trace data instead of potentially biased self-reports in interviews or surveys. This study examines the following two research questions (RQs):
(RQ1) What are prominent communities in the German educational Twittersphere, as well as corresponding participation and engagement patterns of teachers and non-teachers?
(RQ2) How does the TWLZ community fulfill the characteristics of an online community of practice (i.e., domain, community, practice)?
Theoretical background
Online teacher professional development
OPD programs gained substantial popularity in recent years – even before the COVID-19-induced shift to distance education – as teachers seek a variety of more personalized and convenient settings for their PD. Importantly, online formats might vary in their specific activities and routines, depending on the level of formality (Dede & Eisenkraft, 2016; Kyndt et al., 2016; Morina et al., 2025). For example, more formal OPD programs often imitate traditional seminar-like programs emphasizing common learning goals, time schedules, and support systems. Typical activities can be synchronous or asynchronous, often including video- and audio-conferencing, text chats, and simultaneous work on teaching materials. Informal activities are characterized by less rigid structures, higher autonomy, and access to more personalized information (Bruguera et al., 2019; Richter et al., 2011; Kyndt et al., 2016; Staudt Willet, 2024). Informal OPD, often initiated by teachers as a voluntary activity, is mainly used to collaborate with colleagues, learn from one another, and share materials and resources (Kyndt et al., 2016; Ramlo, 2012; Recker et al., 2007).
These collaborative sharing activities offer benefits for teachers. For example, it affords teachers a sense of agency and independence in terms of teachers’ learning processes due to fewer hierarchical structures (Fischer et al., 2019). This supports the notion of personalization instead of a one-size-fits-all approach of traditional PD programs (Carpenter & Krutka, 2015a). Furthermore, the absence of geographical restrictions allows participation of teachers in more remote areas or teachers with little-to-no colleagues teaching the same subject at their school (Powell & Bodur, 2019). Early studies examining online communities’ effectiveness could identify improvement in students' achievement when their teachers engaged in online teacher communities for their professional learning (Fishman et al., 2014; Frumin et al., 2018). More recent studies have noted an increasing use of informal professional learning activity due to the COVID-19 pandemic (Avidov-Ungar et al., 2023) and confirmed the effectiveness of these activities for teacher learning and, to lesser extent, student learning (Morina et al., 2025).
Teachers in online communities on twitter
Online communities are defined as groups of people who gather in online spaces to interact with others over their topics of interest. Continuous interaction creates a sense of community and mutual commitment (Scheckler, 2003). Research showed that many teachers participate in online communities almost daily (Carpenter & Krutka, 2014; Fischer et al., 2019; Visser et al., 2014). For early-career teachers, these communities can be crucial to build an early professional network for resources and teaching support (Staudt Willet, 2024). In addition to networking, teachers often turn to highly prolific teacher influencers on social media for professional development content (Carpenter et al., 2023).
Twitter is a prime example enabling online community formation in which collaborative participation and informal learning can occur with teachers explicitly acknowledging the platform as part of their PD (Aguilar et al., 2021). Teachers often mention the uncomplicated accessibility and the limited time commitment as reasons for their continued participation (Fischer et al., 2019; Staudt Willet, 2019). For example, Carpenter and Krutka (2014) found educators highlighting the opportunities for accessing novel ideas and trends, particularly regarding educational technology. Other studies investigated teacher’s participation patterns, social roles, and benefits of Twitter use for classroom practice and student achievement (Carpenter & Krutka, 2015b; Lord & Lomicka, 2014; Risser, 2013; Rosenberg et al., 2016). For instance, Rosenberg et al. (2020) examined more than 7,000 unique tweets of about 250 individuals in the synchronous #NGSSchat (Next Generation Science Standards) community. They found that over half of conversations were transactional including affirmations of shared content, reframing of previously shared content, and invitations to further discussions.
Recent research also found that these hashtags help parents of children with special needs organize into groups and support each other (Khasawneh, 2024) and that these communities can mirror real-world educational initiatives and policies outside the United States (Mosquera-Gende et al., 2024). However, what is less well known is how automated methods can help create maps or overviews of existing social media communities related to teacher professional development. The present study addresses this gap by proposing and applying a novel method for discovering education-related communities to German Twitter data.
Online communities of practice
Due to the interactions of teachers with and on educational content, such online communities may be described as communities of practice (CoPs). This stems from the social learning theory of Lave and Wenger (1991), who initially investigated several forms of apprenticeship programs (e.g., tailors, midwives) to develop models of informal learning. Notably, a CoP describes groups of people improving their skills and drawing expertise from social interactions fulfilling the following three key characteristics: domain, community, and practice (Wenger-Trayner & Wenger-Trayner, 2015).
Domain refers to the shared identity and competence among community members that lead to differentiation from other groups. Importantly, knowledge in the domain may not be viewed as expert knowledge outside this community. For example, teachers may use specific terms that identify them as teachers as they have unequivocal meanings in the context of teaching (e.g., 1:1 [one device for each student in a class], IEP [individualized education program for students with special education needs], title 1 [schools that receive federal funding due to high percentages of enrolled low-income students]), whereas their meaning may differ in other contexts (Freeman, 1993).
Community refers to interactions among community members and their shared learning and relationship-building processes. These interactions can take multiple forms, including joint discussions and activities that help create a sense of belonging and commitment to the group. While interactions do not need to occur regularly for each member, the community needs to afford opportunities that enable members to engage with each other in the context of the shared domain and identity. For example, teachers’ lounges in schools provide physical spaces for formal and informal opportunities for teachers to discuss teaching ideas and to share information (Ben-Peretz & Schonmann, 2000; Mawhinney, 2010).
Practice refers to the nature of community members as practitioners who develop a joint knowledge base and shared skill sets through interaction in the community. This repertoire might include anecdotes, stories, experiences, and problem-solving strategies. For example, teachers might share how specific classroom management techniques enable certain instructional techniques to better engage students in cognitive thinking processes (Bell, 1994; Gage et al., 2018; Said, 2013; Skiba et al., 2016).
Beyond these shared characteristics, Communities of Practice also involve dynamic participation structures in which members differ in their level of engagement and influence. Lave and Wenger (1991) describe this as a trajectory from legitimate peripheral participation to fuller, more central participation as individuals gain experience and recognition. In online CoPs, such as the online education communities studied here, this often manifests as a core group of highly active contributors who sustain conversations and a larger periphery of less frequent participants who nonetheless learn through observation and occasional interaction (Wesely, 2013). Recognizing these patterns is crucial for understanding how central and long-term users emerge and sustain teacher communities on platforms such as Twitter to improve their professional practice, a point we return to in the Discussion when interpreting user hierarchies in our data.
Prior research on Twitter as a CoP primarily utilized qualitative research. For example, Wesely (2013) interviewed nine world language (WL) teachers on Twitter and organized the findings according to the three CoP characteristics. These teachers shared a common domain, as they identified themselves as WL experts and shared resources and teaching instructions with each other using the #langchat hashtag, which allowed them to connect with each other to form a community. Interestingly, Wesely (2013) noted that these teachers frequently engaged in shared debates on their classroom innovations and strategies for assessing student learning proficiencies, corresponding to a shared practice. Similarly, research outside of Twitter also primarily takes qualitative perspectives. For example, Booth (2012) interviewed eight teachers and two moderators engaging in either the community of the U.S. National Education Leaders Network or the English Teachers’ Online Community. In particular, Booth (2012) discussed how teachers found these online communities to foster knowledge sharing and relationship building.
Methods
Study setting and data collection
This study uses a multi-step procedure to comprehensively map the German educational Twittersphere. Downloads of publicly available tweets used the Twitter Academic API 2.1 and took place between April 8–25, 2022. The data set covered tweets from the inception of Twitter in March 2006 to the download dates. We note that the first tweet in our data set is dated February 17, 2007. First, we downloaded all tweets associated with Germany’s most popular community, the Twitterlehrerzimmer (TWLZ; Fütterer et al., 2021), through the hashtag-based communities “#twlz”, “#twitterlehrerzimmer”, “#tlz”, “lehrerzimmer”, “#twitterlz,” and “#twitterkollegium”. Then, we examined co-occurring hashtags within TWLZ to systematically download other large educational communities. Besides a large chat-based community (#EdchatDE), we generated hashtags associated with (a) education more broadly (indicated through “-edu” suffixes), (b) disciplinary subjects (e.g., “English-”), and (c) German federal states using typical pre- and suffixes (e.g., “Bayern-” for Bavaria). This novel hashtag-selection process resulted in a comprehensive list of 21,893 potential hashtags resulting in 1,074 hashtags-based communities, as many of these potential hashtags did not yield any posts.
Next, we applied our exclusion criteria to these 1,074 communities so that communities with (a) a tweet volume below 100 tweets (excluding retweets); (b) an apparent participation of minors (i.e., students; based on manual and ad hoc inspection of a few hundred tweets for each considered hashtag) who need special protection due to ethical considerations; (c) an apparent majority of tweets with non-educational conversations; and (d) a majority of non-German tweets were removed from the database.
Notably, downloads of hashtag-based communities via the Twitter API only collect data from tweets that include the hashtag as part of their tweet. However, many tweets elicit replies and quotes which do not include the initial hashtag (and are thus not natively part of the data). This conversation data has also been downloaded for tweets with at least one reply. To assign hashtags to these additional tweets, the initial tweet of a conversation was tagged as a sparking tweet. Overall, the final study sample included 100 communities and 2,761,579 tweets.
Data anonymization and user protection
Given the breadth and depth of social media data, it is important to protect individuals’ privacy. For instance, interactions with tweets increase their public visibility and might be used to reconstruct individuals’ identities even after the original tweet has been deleted (Keküllüoglu et al., 2020). Similarly, postings on social media may reveal personally identifiable information (PII), such as student names and faces. For instance, schools may proudly share on their social media platforms that one of their students won a regional athletic championship revealing that student’s PII (Burchfield et al., 2021; Rosenberg et al., 2021b).
We applied several data protection procedures: First, we replaced all Twitter-assigned tweet, user, and conversation identifiers with random, numeric, and anonymous identifiers. Second, we replaced email addresses and phone numbers extracted from text variables with placeholders. Third, we replaced all mentions of usernames with their new anonymous user identifier. Additionally, we never directly quote tweet text as these could be used to potentially identify users via search engines. Instead, we share synthetic tweets that preserve the original meaning but use different wording (Moreno et al., 2013; Williams et al., 2017). Lastly, we purposefully omitted privacy-sensitive variables from our data not relevant to the present study (e.g., media-related information on attached images).
Measures
Tweet-level variables
We classify tweets into original tweets (i.e., unique tweets from a user), reposts (i.e., retweets or quotes of an original tweet), and replies (i.e., comments on an original tweet or a repost). For each tweet, continuous variables represent the engagement with the tweet through the number of likes, reposts, and replies. Moreover, we use the tweet timestamp to examine temporal engagement patterns. We base these variables based on standard descriptive statistics for Twitter/X data from past research (e.g., Fischer et al., 2019).
Tweet sentiment
Sentiment analysis is the automated extraction of emotional valence from text (Yue et al., 2019). We sampled 1,000 unique tweets to employ a human training data coding procedure. Two trained human coders reached almost perfect consensus of κ = 0.84 (Landis & Koch, 1977) for a trinary sentiment classification into the categories positive, neutral, and negative on a sample of 100 tweets. Two raters split the remaining 900 tweets. Afterwards, we evaluated classification accuracy measures for the (German) sentiment dictionaries of SentiStrength (Thelwall et al., 2010), the German LIWC dictionary (v.2015; Meier et al., 2019), and the German sentiment dictionary SentimentWortschatz (Version 2.0; Remus et al., 2010). With a distribution of 100 positive, 873 neutral, and 27 negative tweets, we found trinary, compared to binary classification merging positive and neutral tweets, to have unsatisfactory classification accuracy, similar to previous validation studies (Borchers et al., 2021). Weighted F1 scores also found LIWC (F1 = 0.93) and SentimentWortschatz (F1 = 0.91) exhibiting superior binary classification performance compared to SentiStrength (F1 = 0.79). SentimentWortschatz also had a higher recall rate of (0.59) for negative tweets than LIWC (0.37). Therefore, we employed SentimentWortschatz with a binary classification (positive/neutral vs. negative). Afterwards, we computed the sentiment ratio (Wang & Fikis, 2019) defined as the ratio of negative over positive/neutral tweets.
Tweet classification
As a novel method, we identified eight categories to identify the CoP features domain, community, and practice. Afterward, two trained human coders coded the 1,000 tweets with the most engagement for all eight categories. The coders reached satisfactory agreement on all variables (κ = 0.66–1.00, M = 0.84). Notably, a single tweet can be assigned multiple categories.
Domain includes two categories: teaching profession-specific and providing information. Teaching profession-specific tweets refer to lesson designs, the preparation of the content, or the discussion of teacher-specific situations and problems. Providing information tweets contain teaching-specific information, materials, resources, or links that are shared with the community. Community includes three categories: network, search for information/questions, and emotional support. Network tweets refer to networking between users or facilitation of joint work and discussion. Search for information/questions tweets contain questions users pose to the community, for instance, regarding lesson designs. Emotional support tweets express users’ emotional states and reactions to shared emotional states. Practice includes three categories: COVID-19, educational policy, and anecdotes. COVID-19 tweets discuss the COVID-19 situation in users’ school contexts or in their teaching. “Educational policy” tweets discuss political issues related to education (outside of COVID-19), such as regulations or statements by politicians. Anecdotes tweets include teachers sharing personal stories within an educational context. Appendix A (Table 9) includes exemplary synthetic tweets for each category.
User-level variables
On the user-level, we calculated a user’s lifespan (i.e., difference between their join date and their last interaction in the data) and their frequency of participation (i.e., the total number of tweets divided by the lifespan). These variables were taken from past research on Twitter community participation patterns (Borchers et al., 2023).
User classification
Users are classified into three categories: teachers, bots, and non-teachers/non-bots. Non-teachers refer to all users that are not teachers, while non-bots refer to all users that are not bots. We randomly sampled 1,000 profiles to generate training data for subsequent automated user classification (Pennacchiotti & Popescu, 2021). Two trained human coders evaluated users’ Twitter bio and up to 50 randomly sampled tweets posted by each profile to assign the users as teacher or non-teacher. The coders reached a substantial agreement, κ = 0.77 (Cohen, 1960; Landis & Koch, 1977), after independently coding 250 user profiles. The remaining 750 user profiles were split between the two coders. Next, a classifier based on textual features of user bios and tweets predicted whether users were teachers. Initial model explorations resulted in logistic regression and random forest as viable models based on AUC values. Grid search for hyperparameter tuning curated two candidate models for each model architecture. Optimizing for AUC, we used tenfold cross-validation for hyperparameters tuned on a training dataset (N = 750). Based on test set performance (N = 250), we chose the logistic regression model for classifying teachers because permutation tests indicated significantly better F1 (0.53), precision (0.93), and AUC scores (0.79) based on 100,000 bootstrapped permutations.
As bots are present on Twitter, these bot accounts need to be detected to get an unbiased estimate of “human” community sizes. For example, a TWLZ bot automatically retweets all tweets including the #TWLZ hashtag. However, while automatic bot detection algorithms have already been used on German Twitter data (Keller & Klinger, 2019), recent studies suggest that popular algorithms may not exhibit satisfactory accuracies (Rauchfleisch & Kaiser, 2020). Therefore, we employed a conservative, deterministic algorithm that matches bot-relevant keywords (i.e., “bot”, “bot-account” and “community-bot”) against Twitter bios to tag bots. Afterwards, a human coder performed spot-checks on potential bot accounts to identify false positives. The led to the identification of N = 187 bots posting a total of N = 261,160 tweets (99.8% reposts) across all communities.
Conversation variables
We define conversations as a thread that involves at least two tweets (including the sparking tweet) from at least two users. For each conversation, we calculate the total number of tweets (i.e., length of the conversation), users, and mentions in the thread. Additionally, we computed the half-life of a conversation, which indicates the time passed between the sparking tweet and the median tweet in the conversation.
Community variables
Aggregate tweet- and user-level variables within a community include the number of users, tweets, reposts, replies, and conversations. We then calculate the ratio of posts per user to better gauge individual user’s average contribution to the community.
Analytical methods
To identify prominent communities and to examine participation and engagement patterns in the German educational Twittersphere (RQ1), we provide descriptive information on key community metrics for each of four core types of educational communities: TWLZ, state-specific communities, discipline-specific communities, and EdchatDE. The level of engagement of tweets is measured through the number of reposts and likes. As most measures are not normally-distributed and heavily zero-inflated, we fitted negative binomial models (McElduff et al. 2010). To compare the sentiment ratios of tweets by teachers and non-teachers, we used a χ2 test to check relations between user type and sentiment with the Phi coefficient measuring the strength of the association.
To examine how participation in Twitter communities may fulfill key CoP characteristics (RQ2), we focus our analyses on the most prominent community, the TWLZ.
For domain, we counted the occurrence of teacher-specific language in tweets as defined by 47 pre-selected domain-specific words that were coded by two independent coders (κ = 0.79) across the entire community.
For community, we examined social network characteristics of user engagement alongside the frequency of the lifespan of user engagement and user-level teacher classifications. We generated directed graphs of user interactions (i.e., mentions, replies, retweets, or quotes) and classified users by whether they were part of the top 5% most influential users based on betweenness centrality (confirming robustness for 1%, 10%, and 15% thresholds). Betweenness centrality represents the degree to which users in our interaction networks can bridge connections with peripheral users in our network (Riquelme & González-Cantergiani, 2016), effectively being able to build new community connections. We study how these highly influential users interacted with other users and how teachers interacted with non-teachers. Following prior definitions of community membership based on two required interactions with the community (Borchers et al., 2023), we also analyzed how the 50% oldest TWLZ members interacted with newer members.
For practice, an exploratory analysis of selected dimensions of the German LIWC dictionary (Meier et al., 2019) examined tweet ratings on certainty, tentativeness, and social following prior work relating LIWC to cognitive presence in communities of inquiry (Joksimovic et al., 2014). We investigate these LIWC dimensions aggregated by year and across teachers and non-teachers. Similarly, LDA topic-modeling algorithms manually investigate relationships of practice with the top ten most weighted (i.e., most distinct and representative) words across the top 12 topics in conversations of teachers and non-teachers, featuring documents spanning multiple tweets and a total of minimum 50 words each, following commonly used best practices (Tang et al., 2014). We also employed standard preprocessing techniques for tokenization, including stopword removal, stemming, and removing user mentions and URLs. We determined the number of topics through manually determining the elbow point of the distribution (Arun et al., 2010).
Notably, we conducted an additional qualitative content analysis of the 1,000 tweets with the most engagement in the TWLZ community to identify topics related to all three CoP characteristics.
Results
Participation patterns in German twitter communities (RQ1)
The entire German educational Twittersphere includes N = 2,761,579 tweets from N = 143,004 users who posted an average of 19.31 tweets (SD = 667.83). The largest community in the data set is the TWLZ community (2,039,402 tweets, 98,312 users), followed by subject-specific communities (374,892 tweets, 51,991 users), state-specific communities (329,743 tweets, 21,859 users), and EdChatDE (261,748 tweets, 11,260 users; Table 1).
Table 1. Descriptive information on community types
Category | Number of tweets | Number of reposts | Number of replies | Number of users | Posts per user | Number of conversations |
|---|---|---|---|---|---|---|
TWLZ | 2,039,402 | 875,138 | 954,784 | 98,312 | 20.74 | 99,988 |
Subjects | 374,892 | 124,642 | 121,715 | 51,991 | 7.21 | 21,835 |
States | 329,743 | 139,872 | 151,773 | 21,859 | 15.08 | 19,093 |
EdChatDE | 261,748 | 75,639 | 84,447 | 11,260 | 23.25 | 25,097 |
Overall | 2,761,579 | 1,108,709 | 1,201,577 | 143,004 | 19.31 | 150,857 |
A tweet may be assigned to more than one community due to hashtag co-occurrences
Comparing the two largest communities, TWLZ had around eight times as many tweets as the second-largest community (i.e., the EdChatDE community). Yet, TWLZ had only around four times as many conversations. Notably, users posted an average of around 12.1% more tweets in EdChatDE compared to TWLZ. Finally, users engaging in state-specific communities posted more than twice as many tweets on average (M = 15.08) than in subject-specific communities (M = 7.21). Yet, users in subject- and state-related communities still had fewer tweets per user than TWLZ and EdChatDE.
Both subject-level and state-level communities were characterized by a few dominant communities. For subjects, the largest communities relate to German (146,546 tweets, 22,175 users), followed by Religious Studies (85,516 tweets, 8,368 users), and Computer Science (39,342 tweets, 4,816 users). For states, the largest communities are located in Bavaria (208,259 tweets, 11,779 users), North-Rhine Westphalia (58,071 tweets, 10,581 users), and Baden-Württemberg (55,563 tweets, 6,088 users). Notably, users engaging in Bavaria-related communities also exhibited the most number of tweets per user on average compared to all other states (M = 17.68). Tables in Appendix A (Tables 10–13) show the descriptive statistics of the individual communities.
User classes
Examining user classes in the communities, we distinguish between teachers (N = 10,517), bots (N = 187), and non-teachers (i.e., non-teachers, non-bots, N = 132,300; Table 2). Notably, across all communities, teachers posted almost as many tweets (N = 1,212,473) as non-teachers (N = 1,287,946) despite the considerably lower number of teachers (N = 10,517 vs. N = 132,300; Table 2). Notably, teachers account for most tweets in the TWLZ (47.6%) and the state-wide communities (49.5%). Meanwhile, subject-specific communities had the lowest percentage of tweets for teachers across community types (25.3%; Fig. 1).
Table 2. User class counts by community type
User class | Overall | TWLZ | EdChatDE | Subjects | States |
|---|---|---|---|---|---|
Number of users | |||||
Teacher | 10,517 | 10,151 | 1,769 | 4,406 | 5,694 |
Non-teacher | 132,300 | 88,010 | 9,467 | 47,509 | 16,115 |
Bot | 187 | 151 | 24 | 76 | 50 |
Number of tweets | |||||
Teacher | 1,212,473 | 971,502 | 111,342 | 94,830 | 163,340 |
Non-teacher | 1,287,946 | 812,967 | 149,957 | 270,384 | 137,414 |
Bot | 261,160 | 254,933 | 449 | 9,678 | 28,989 |
A tweet may be assigned to more than one community due to hashtag co-occurrences
[See PDF for image]
Fig. 1
Total tweet counts by user type across communities (left) and total tweet count by user type and tweet type (right)
Temporal participation patterns
Figure 2 describes temporal participation patterns of all users (excluding bots) and tweets across the four community types. We find three notable trends. First, TWLZ tops all other communities. It increased its user numbers from 2016 to 2020 by 12.7% (2016: 176 users, 2020: 22,524 users) alongside a 95.6% increase in tweets (2016: 635 tweets, 2020: 607,855 tweets). Second, with the introduction of new digitization initiatives in Germany (BMBF, DigitalPakt Schule, 2023), the number of users and tweets in state-specific communities plateaued from 2011–2016 at around 31 users and 153 tweets per year before showing a 2.3% increase in users and a 32.5% increase in tweets in 2017 (N = 582 users, N = 24,755 tweets). Mostly due to the growth of the #bayernedu and #nrwedu communities. Third, EdChatDE showed a decline in users and tweets after a peak at around 60,000 tweets and 2,200 users in 2015 and 2018, respectively. This decline in EdChatDE users and tweets coincided with the exponential growth of TWLZ users and tweets.
[See PDF for image]
Fig. 2
Yearly temporal participation patterns of all users and tweets across community types of teachers (left) and non-teachers (right)
Figure 3 shows the time of day of posting across community types for teachers and non-teachers. Interestingly, there are no large differences in the time of posting between teachers and non-teachers across all community types. Whereas the posting hours of EdChatDE coincide with 8-9PM, the postings across the other community types are mostly distributed during 7AM-11PM. TWLZ has the highest participation during 5–10 PM (36.7%). This is similar to state-based communities (38.3%) but contrasting subject-based communities that have their highest posting frequencies during 11–12 AM and 2–3 PM for non-teachers (13.8%) and 10-11AM and 4-7PM for teachers (26.2%). Notably, teachers and non-teachers show relatively stable degrees of engagement during regular working hours (8AM-4PM) with 8-9AM being on the lower end of engagement. That said, there is considerably more activity from teachers across all community types after their regular working hours.
[See PDF for image]
Fig. 3
Time of day posting percentages for all community types in local time in Germany (GMT + 1). To improve readability, we removed #relichat from the subjects-based communities because of recurring chat hours (Wednesday, 8–9 PM)
Teachers exhibited more extended engagement on weekends (23.6%) than non-teachers (21.6%; Fig. 4). Comparing weekdays, Saturday was the day with the lowest activity across teachers (11.5%) and Sunday for non-teachers (10.7%). Conversely, Tuesday was the day with the highest activity (20.3% and 21.8%, respectively). Unsurprisingly, EdChatDE had the highest daily post count on the designated chat day. However, teachers and non-teachers also contributed to this hashtag on other days of the week (26.3% and 29.3%).
[See PDF for image]
Fig. 4
Weekday posting percentages for all community types. To improve readability, we removed #relichat from the subjects-based communities because of recurring chat hours (Wednesday, 8–9 PM)
Figure 5 describes monthly posting percentages across community types and teachers and non-teachers. Overall, monthly activity patterns are similar for teachers and non-teachers across all community types. Interestingly, the overall activity pattern roughly follows the school holidays with June to August being the months with the lowest monthly activity. The months with the highest monthly activity are February/March and October/November.
[See PDF for image]
Fig. 5
Monthly posting percentages for all community types
Figure 6 shows the lifespan and participation frequency of a random sample of 100,000 users of the dataset across all community types. Notably, this plot indicates that many users participate in the communities for a long time (500 or more days, N = 34,529 users in the entire dataset). Similarly, many users with lifespans of 500 + days also post on average more than once a day. Overall, this indicates that many users do not use Twitter as a one-off to post content but as a sustainable resource and communication tool.
[See PDF for image]
Fig. 6
Distribution of lifespan and tweet frequency of users for all community types. Note.N = 100,000 random users from the dataset, only showing tweets/day < Q_75 of participation frequency
Examining user engagement and conversations
In line with its distinct temporal activity patterns, EdChatDE had lower engagement (i.e., mean number of reposts per tweet and the mean number of likes per tweet) compared to other community types (Table 3). Conversely, TWLZ has more engagement on almost all metrics compared to the other community types (reposts: M = 0.81, SD = 8.39; likes: M = 4.16, SD = 36.83).
Table 3. Descriptives of engagement for all community types for all users, excluding retweets and bots
Category | M reposts per tweet (SD) | M(Q95) reposts per tweet | M likes per tweet (SD) | MQ95) likes per tweet |
|---|---|---|---|---|
TWLZ | 0.81 (8.39) | 0.40 | 4.16 (36.83) | 2.39 |
EdChatDE | 0.40 (1.41) | 0.26 | 1.31 (3.01) | 1.07 |
Subjects | 0.51 (3.06) | 0.30 | 2.45 (14.49) | 1.62 |
States | 0.78 (3.52) | 0.44 | 3.34 (12.37) | 2.33 |
Overall | 0.71 (7.19) | 0.36 | 3.57 (31.56) | 2.09 |
Comparing the differences in the engagement metrics for teachers and non-teachers, we found significant differences across almost all engagement metrics and community types (Table 4). For example, in the entire dataset, teachers had significantly fewer average reposts per tweet than non-teachers (RR = 0.883, p < 0.001). This pattern is replicated across all community types. The differences in likes between teachers and non-teachers was more nuanced. Teachers in EdChatDE received more likes (M = 1.42, SD = 3.66) compared to non-teachers (M = 1.22, SD = 2.35, RR = 1.160, p < 0.001). However, there was no significant difference between teachers and non-teachers for subject- (RR = 0.986, p = 0.072) and state-communities (RR = 0.992, p = 0.226).
Table 4. Engagement comparison of teachers and non-teacher tweets
Category | Teachers M Reposts per Tweet (SD) | Non-Teachers M Reposts per Tweet (SD) | Reposts Rate Ratioa | Teachers M Likes per Tweet (SD) | Non-Teachers M Likes per Tweet (SD) | LikesRate Ratioa |
|---|---|---|---|---|---|---|
TWLZ | 0.73 (6.86) | 0.91 (10.02) | 0.803*** | 4.22 (32.78) | 4.10 (41.47) | 1.028*** |
EdChatDE | 0.40 (1.41) | 0.41 (1.40) | 0.966** | 1.42 (3.66) | 1.22 (2.35) | 1.160*** |
Subjects | 0.47 (1.86) | 0.53 (3.39) | 0.892*** | 2.42 (6.16) | 2.45 (16.46) | 0.986 |
States | 0.70 (3.41) | 0.87 (3.64) | 0.805*** | 3.33 (13.08) | 3.35 (11.51) | 0.992 |
Overall | 0.67 (6.18) | 0.76 (8.06) | 0.883*** | 3.79 (29.5) | 3.36 (33.47) | 1.128*** |
aRate Ratio of negative binomial regression comparing teachers to non-teachers
*p < 0.05, **p < 0.01, ***p < 0.001
Examining conversations (Table 5), we find that both TWLZ (M = 10.66, SD = 25.56) and the state-based communities (M = 10.39, SD = 23.77) have more than twice the mean conversation length compared to EdChatDE (M = 4.49, SD = 6.05). There are different patterns regarding the mean half-life of conversations, as well as the first response time to conversations. For example, the half-life is the lowest for conversations of EdChatDE (M = 14.82 h, SD = 463.72 h) compared to all other asynchronous communities. Notably, subject-specific communities have the longest half-life (M = 52.45 h, SD = 760.69 h), which is substantially longer compared to state-based communities (M = 31.73 h, SD = 368.60 h) and the TWLZ (M = 23.26 h, SD = 307.47 h), showing that temporal patterns of conversations are more spaced out.
Table 5. Conversation statistics across community types across all users
Category | M lengths of conversation (SD) | M number of users (SD) | M number of mentions (SD) | M half-life of conversation in hours (SD) |
|---|---|---|---|---|
TWLZ | 10.66 (25.56) | 5.78 (11.99) | 21.64 (107.26) | 23.26 (307.47) |
EdChatDE | 4.49 (6.05) | 2.80 (2.12) | 6.42 (22.26) | 14.82 (463.72) |
Subjects | 7.05 (16.26) | 4.25 (8.00) | 11.37 (85.57) | 52.45 (760.69) |
States | 10.39 (23.77) | 5.48 (10.49) | 24.25 (113.76) | 31.73 (368.60) |
Overall | 8.95 (21.86) | 4.99 (10.33) | 17.23 (89.31) | 25.04 (413.94) |
Examining user sentiment
When examining tweet sentiment, we found across all community types for every 100 tweets with positive and neutral sentiment, 32 tweets were posted with a negative sentiment (SR = 0.32). Notably, TWLZ had the most negative tweets (SR = 0.37), followed by the state-based communities (SR = 0.28), the subject-based communities (SR = 0.22), and EdChatDE (SR = 0.18). This speaks to a mostly positive environment in the German Twittersphere.
Comparing the sentiment of teachers to non-teachers (Table 6), we find that teachers’ post on average less positive tweets compared to non-teachers, χ2 = 134.43, p < 0.001. Notably, the direction of the differences is different across community types with TWLZ and state-based communities having more negative posts for teachers, and EdChatDE and the subject-based communities having more positive posts.
Table 6. Sentiment ratio comparisons of teachers and non-teachers
Category | SR Teachers | SR Non-Teachers | ||
|---|---|---|---|---|
TWLZ | 0.35 | 0.39 | 598.31*** | .022 |
EdChatDE | 0.21 | 0.16 | 553.76*** | .054 |
Subjects | 0.23 | 0.21 | 85.14*** | .018 |
States | 0.27 | 0.29 | 55.27*** | .016 |
Overall | 0.32 | 0.31 | 134.43*** | .009 |
*p<0.05, **p<0.01, ***p<0.001
Twitter as an online community of practice (RQ2)
Domain CoP characteristic
To examine whether TWLZ fulfills the domain CoP characteristic, we conducted a quantitative text analysis of the usage of common educational words across all TWLZ tweets. We found that 42.3% of all TWLZ tweets included one of the 47 pre-selected domain-specific words (N = 755,616 tweets). Notably, teacher tweets included domain-specific words more frequently (43.6%) compared to non-teachers (40.8%).
The qualitative analysis of the 1,000 tweets with the most engagement in the TWLZ indicated that 69.6% discussed topics that we identified within the domain COP feature. These tweets focused on teacher and teaching specific situations (58.3%) and providing information for teaching related topics (11.3%). For example, users discuss characteristics of good video-based teaching.
Community CoP characteristic
To examine the community CoP characteristic, we examined the frequency of user interactions based on their influence in the community as measured by their betweenness centrality. We examined how highly influential users interacted with less influential users, how teachers interacted with non-teachers, and how older TWLZ members (i.e., veterans; operationalized as the first 50% of members to have joined the community) interacted with newer members. Classifying a total of N = 3,823,974 interactions, we found many interactions within all three dimensions of users (centrality, teachers, veterans; Table 7). Highly influential users, teachers, and veterans primarily interacted within their peer-group. Notably, veterans interacted with more than 50% of users that joined TWLZ later than them, pointing to the prolonged engagement of early joiners. While non-central and non-veteran users tended to interact more with central and veteran users than with their peers, non-teachers showed similar interaction patterns with both teachers and non-teachers.
Table 7. Number of directed interactions (i.e., mentions, replies, retweets, quotes) across different hierarchies of users
Centrality Status | To Non-Central User | To Central User |
From Non-Central User | 112,294 | 749,651 |
From Central User | 383,426 | 2,578,603 |
Teacher Status | To Non-Teacher | To Teacher |
From Non-Teacher | 800,184 | 820,637 |
From Teacher | 825,140 | 1,378,013 |
Veteran Status | To Non-Veteran | To Veteran |
From Non-Veteran | 120,348 | 371,634 |
From Veteran | 250,356 | 3,029,323 |
Community membership required at least two community interactions leading to a reduced sample size of transactions between veterans and non-veterans
Additionally, we investigated the distribution of teachers and non-teachers across highly central and veteran users (Table 8). We find that teachers are both overrepresented among the top 5% most influential user group, as well as for veteran user status. This overrepresentation was particularly the case for central users. All associations are significant according to χ2 tests of independence, including the interaction distribution across central and non-central users χ2(1) = 4.10, w = 0.00 (null effect), p =.043, teachers and non-teachers χ2(1) = 54,260, w = 0.12 (small effect), p <.001, and veteran and non-veteran users χ2(1) = 136,679, w = 0.18 (small-to-medium effect), p <.001.
Table 8. Distribution of teachers and non-teachers across centrality and veteran-status dimensions
Centrality Status | Non-Teacher | Teacher |
Non-Central User | 83,921 | 7,368 |
Central User | 3,549 | 7,368 |
Veteran Status | Non-Teacher | Teacher |
Non-Veteran | 32,839 | 2,157 |
Veteran | 27,897 | 7,099 |
Community membership required at least two community interactions leading to a reduced sample size of transactions between veterans and non-veterans
The qualitative analysis of the 1,000 tweets with the most engagement in the TWLZ identified 69.5% of tweets related to community. In particular, the majority of tweets (51.8%) were written for emotional support, 14.7% of tweets included calls for active networking and collaboration, and only 3.0% of tweets requested information. For example, a teacher asked for emotional support after learning that one of their students is going to be deported to Afghanistan.
Practice CoP characteristic
To examine the practice CoP characteristic, we investigated qualitative data on the 1,000 tweets with the highest engagement in the TWLZ. We triangulated the results with selected LIWC dimensions and LDA-derived topics across teachers and non-teachers in TWLZ tweets (Fig. 7). LIWC analysis indicated that between 2012 and 2016, the degree of social references and certainty in teacher tweets was consistently higher than for non-teachers. This gap subsequently narrowed, while social references starkly increased in 2017 which coincided with the exponential growth of the TWLZ (Fig. 2). This suggests that concurrent with TWLZ’s growth, social references, particularly those expressed by non-teachers, became more common.
[See PDF for image]
Fig. 7
Average level of certainty, social references, and tentativeness by year across teachers (right) and non-teachers (left). Social references are scaled by 1/10 for readability
Additionally, LDA identified 12 topics in the tweets of teachers and non-teachers, which were all shared across groups but entailed different connotations. In particular, we find two shared themes between teachers and non-teachers. First, digital innovation, with non-teachers focusing on “moodle,” “apps,” “iPad,” and “iserv” while teachers additionally focused on “data protection,” “tablets,” “laptops,” and “beamer.” Second, we found a shared topic on the COVID-19 pandemic. However, while non-teachers tended to focus on “mask requirements”, “letting fresh air in,” and “in-person instruction,” teachers tended to focus on “vaccinations,” “incidence rates,” “FFP masks”, and the “RKI”, the German government’s central scientific institution in the field of biomedicine. Appendix B includes supplemental material that shows all topics and words of the topic modeling.
These findings are mirrored in the qualitative analysis of the 1,000 tweets with the most engagement in the TWLZ, which identified 95.7% of tweets related to practice. Notably, these tweets related to topics discussing teaching and teacher specific topics related to the COVID-19 pandemic (52.8%), policies and political decisions concerning education (36.3%), and teacher and teaching related anecdotes (6.6%). For example, a tweet informed about the opportunity to a free download of a world atlas for students to use during the COVID-19-induced school closures.
Discussion
This study provides the first large-scale, computational analysis of the German educational Twittersphere, advancing prior research that has largely relied on the study of large United States Twitter communities (Rosenberg et al., 2021a, b; Staudt Willet, 2024) or individual hashtag communities in international contexts (Mosquera-Gende et al., 2024). By examining 2.7 million tweets and 143,000 users, we uncover the structure and dynamics of online teacher communities—insights that were previously inaccessible—offering a novel, data-driven perspective on informal teacher learning. As a secondary contribution, we describe novel methods to detect communities on Twitter related to educational communities with teachers. This study comprehensively maps the entire German educational Twittersphere and provides an array of aggregated measures representing participation and engagement patterns (RQ1), while also identifying characteristics of a CoP of education-related communities (RQ2). This study advances the field of educational research with social media data both methodologically and theoretically.
First, our study provides a principled and reproducible way of sampling Twitter communities in a specified domain and locality. This advances past research that has exclusively relied on user networks (Staudt Willet, 2024) or highly influential users (Carpenter et al., 2023) to detect educational communities on social media. Potential applications of our methods are manifold. For example, future research may reproduce our methodology to gather records of Twitter communities in other countries or domains that can also be regularly updated with new data. Effectively, researchers need only to specify a set of predominant or archetypal Twitter communities to sample co-occurring hashtags from and a set of pre- and suffixes for community discovery to reproduce our sampling procedure. Notably, we include data anonymization and aggregation procedures, which are important adhering to ethics and privacy concerns (Naomi et al., 2021; Rosenberg et al., 2022a, b). This opens novel research directions, for example, the real-time investigation of topical trends across communities or switching user bases over time.
Second, while CoP research is widely established and has been replicated in many fields, prior CoP research related to OPD primarily utilized qualitative methodologies (Booth, 2012; Wesely, 2013). To our knowledge, this study represents the first attempt at merging educational data science perspectives with qualitative analysis to provide a more generalizable understanding of online CoPs for teachers on social media platforms.
Characteristics of the German educational twittersphere (RQ1)
Our descriptive analyses uncovered several interesting patterns of user engagement and participation. For example, we find a strong representation of German, religion, computer science, music, and history-related hashtags. Especially, the existence of a large religious education chat-based community is surprising and encourages further quantitative explorations (Peböck, 2020). Similarly, the computer science-related communities included hashtags that call for computer science to be a required course in the German K-12 curriculum (currently, computer science constitutes an elective in most German states). These communities included hashtags related to future workplace skills (e.g., #pflichtfachinformatik), which are themes regularly invoked by stakeholders from the private sector (Waters & Williams, 2011). Future work may investigate how different stakeholders differ and shape the discourse around curricular changes. Interaction networks among these stakeholders induced by replies, mentions, retweets, and quotes may reveal differences in the influence of these stakeholder groups on the engagement patterns of users (see Rosenberg et al., 2020 for U.S. examples).
Regarding state-based communities, we find differences in the composition and representation of different administrative regions (with independent educational systems). Most notably, Bavaria showed more activity in their state-specific educational hashtags than all other regions of Germany. North Rhine-Westphalia and Baden-Württemberg followed Bavaria, which also ranks among Germany's three most populous regions. Northern and eastern German regions were underrepresented. However, when accounting for state-wide population size, city-states like Berlin, Bremen, and Hamburg showed the largest proportional user count and tweet activity. Different interpretations of this finding come to mind. First, activity distributions might be skewed due to the general "winner-take-all" structure of Twitter, where few communities (e.g., TWLZ and EdChatDE) account for most activity (Borchers et al., 2023). An alternative explanation is that the degree of activity is confounded with the German regions' general economic and technological development. For example, a community in Baden-Württemberg, #schuledigitalbw, relates to a focused effort by the regional government to foster digital infrastructure in education. Notably, both explanations are not mutually exclusive. Future research may investigate whether policy actors or policy differences may explain activity differences in state-based hashtags.
In contrast to prior research on the U.S. educational Twittersphere, which has characterized teacher participation as centered around a few large, recurring hashtag communities such as #EdChat or #NGSSchat (Carpenter & Krutka, 2015a; Rosenberg et al., 2020), our findings reveal a more decentralized network of teacher Communities of Practice in Germany. While the TWLZ functions as a dominant hub, we identified numerous regional and subject-specific communities, including discipline-based spaces such as #relichat, that foster sustained participation and knowledge exchange alongside national-level discussions. From a CoP perspective, this suggests that professional learning in the German educational Twittersphere occurs through multiple, partially overlapping communities that cultivate localized domains of shared practice rather than through a single, centralized space. Future research could explore whether similar patterns of distributed or nested CoPs exist in the U.S. context but have remained underexamined due to the visibility of large, centralized hashtags.
When comparing user types, we find that while teachers only represent a minority of all users (7.4%), they still account for almost half the tweets across all communities (43.9%). Notably, teachers account for most tweets in the TWLZ (47.6%) and the state-wide communities (49.5%). This finding is novel as past research studying national communities similar to ours has not classified users into teachers (Mosquera-Gende et al., 2024). The high activity of teachers could be related to “teacher influencers” which have been noted in United States social media data (Carpenter et al., 2023), though that hypothesis is subject to future research. Notably, teachers and non-teachers did not substantially differ in the engagement related to time of day, day of week, or month of year. However, teachers also often tweet outside their regular working hours. This finding aligns with prior work arguing for the suitability of educational Twitter communities to foster teacher PD at times convenient to teachers (Fischer et al., 2019). Still, more research could help understand teacher retention and engagement patterns. These may include the role of engaging with central users in the network (Rosenberg et al., 2020), sentiment (Borchers et al., 2021; Rosenberg et al., 2021b), or subcommunities on user engagement (Carpenter & Krutka, 2015a). In addition, offline analyses may use mixed methods to investigate how teachers use the TWLZ to complement and advance their teaching practices outside the Twitter CoP and in the classroom.
From a Communities of Practice perspective, these patterns suggest that teachers often occupy more central or core roles within the TWLZ, shaping discussions and maintaining community continuity. Their disproportionate contribution to the overall volume of tweets mirrors the participation gradients typical of CoPs, where a small subset of highly active members sustain the domain and practice through ongoing knowledge exchange and modeling of expertise (Lave & Wenger, 1991). In contrast, less active or newer participants may remain in more peripheral roles, observing and selectively engaging with these central figures. Future work could examine how such trajectories, from peripheral to central participation, unfold over time, and how they influence both the sustainability of the community and the professional learning opportunities it affords to educators.
Educational twitter as a community of practice (RQ2)
Results indicate that TWLZ fulfills the characteristics of a CoP. First, TWLZ tweets show that teachers share a common domain, by demonstrating that teachers use more words that are substantially linked to their teaching practice than non-teachers. A shared identity among teachers may be fostered among teachers via social interactions in the community. First, most tweets focused on teachers and teaching specific situations, demonstrating that the community is a venue for expression and exchange about their profession. Second, social references in the TWLZ substantially increased concurrent with its exponential community growth around 2017. These findings align with the interpretation that a shared community identity can arise from interactions about one’s profession and identity with individuals sharing that identity (Akkerman & Meijer, 2011). Next, analyzing social network features we found that TWLZ to exhibit the community COP feature. We found substantial transactions across our user classifications (i.e., novices, teachers, and highly influential users). Yet, our results highlight hierarchies in user interaction preferences corresponding to Lave and Wenger (1991). Their theory states that communities change and grow by novices turning into experts by interaction and through acquiring more knowledge with time. Our findings highlight that novice users usually interact more with veterans since they may be perceived as more knowledgeable within the community. Similarly, non-teachers preferentially communicated with teachers as they may be perceived as more versatile in topics discussed in the TWLZ (e.g., teaching practices). Our qualitative analysis further informs transactions between users of different positions in user hierarchies. For example, the high number of tweets that were written with the intention of emotional support indicates that TWLZ members may view other members as capable in understanding their struggles and capable of providing the tweeter with appropriate advice. Overall, our study suggests that users may leverage the TWLZ to fulfill their professional learning needs and goals. Given the recent meta-analytic evidence that online teacher professional development improves teacher learning and, to lesser extent, student learning (Morina et al., 2025), this finding opens up new opportunities to study how and why online communities like TWLZ are effective for teacher learning. Future research may further study how community members progress from more peripheral to central roles and correlate those roles to their learning.
Limitations and future work
The nature of the data set and our analytical methods lead to three major limitations. First, every Twitter data set is bound to its time of download. For instance, the number of likes and engagements with a post are collected at the download time. However, as user accounts and tweets are deleted over time, representations and comparisons of older and new communities might introduce some biases. Future work may explore the robustness of different statistical tools to control the time of download, for example, by adjusting for the running average of community engagement across time.
Second, as Twitter is only one platform for teachers’ interactions on social media, our data are limited in speaking to distinct participation patterns of communities in other social media. For example, researchers may want to survey Twitter users included in our sample outside of Twitter regarding their professional Twitter use and compare differences in social media use across platforms (Aguilar et al., 2021). Furthermore, the COVID-19 pandemic induced a massive shift in user growth and engagement. It is worth studying discontinuities in social media use before and after the pandemic via appropriate statistical models (e.g., interrupted time series) in future work to better understand its impact on education-related communities online. These models could also be used to study user composition changes in more depth. Here, we only considered teachers vs. broad categories of non-teachers. However, it is possible that distinct engagement patterns of parents and other stakeholders that are not teachers could change over time, which is subject to future research.
Third, as our work spearheads new methods regarding domain-specific sampling and the discovery of Twitter hashtag communities, we acknowledge that additional methods to validate further and refine our definition community are needed. For instance, we employed a cutoff of 100 tweets in a hashtagged community to establish a minimally relevant size. However, a more conservative community definition may have additional prerequisites such as a minimal degree of between-user interactions. Future work could also define different types and classes of Twitter communities beyond those encoded in community names (e.g., “chats” compared to presumably less cohesive “Twitter teacher rooms”). These community types may also be detected via bottom-up network clustering. For community detection, modes other than co-occurring hashtags could be explored, for example, user interaction patterns and engagement network clusters. In all cases, future work could consider establishing ground-truth criteria for community detection, for example by surveying the self-perception users engaged in specific hashtags, to empirically evaluate the accuracy of the enumerated potential methods for community detection.
Conclusion
This study presents the first large-scale map of the German educational Twittersphere and a reproducible approach for identifying and analyzing education-related communities nationwide. Drawing on 2.76 million tweets from 143,004 users, we identified key communities, most notably the #twlz, and traced engagement patterns among teachers and non-teachers. The #twlz shows several characteristics of a Community of Practice, with shared language, sustained interaction, and collaboratively developed practices that suggest informal participation on Twitter/X supports professional learning. At the same time, our method enabled the discovery of several subcommunities covering regional and subject-specific interest groups.
We provide a transparent pipeline for community discovery, anonymization, and mixed-methods analysis that can be applied across contexts. Our findings show that teachers, though fewer in number, generate much of the activity, engage beyond school hours, and emphasize distinct topics across stakeholder groups, offering insights for online professional development. Future work should extend this analysis across platforms and time, link online engagement to offline practice, and trace how members move from peripheral to central participation.
Acknowledgements
None.
Authors’ contributions
CB performed data analysis, data curation and acquisition, visualization, and contributed to conceptualization as well as writing of the original draft and editing. FM contributed to conceptualization, writing of the original draft, and editing. LK performed data analysis, data curation and acquisition, visualization, and contributed to conceptualization as well as writing of the original draft and editing. FM contributed to conceptualization, writing of the original draft, and editing. CF provided supervision, project administration, project resources, and contributed to conceptualization as well as writing of the original draft, and editing.
Funding
Open Access funding enabled and organized by Projekt DEAL. The authors have no funding to disclose.
Data availability
The datasets generated and/or analyzed during the current study contain data collected from X (formerly Twitter) and cannot be shared publicly due to the platform's terms of service. Aggregated statistics and code used to process and analyze the data are available from the corresponding author upon reasonable request.
Declarations
Ethics approval and consent to participate
All data were obtained from publicly available posts on X (formerly Twitter) and analyzed in accordance with the platform’s terms of service and applicable privacy standards.
Consent for publication
All authors have read and approved the final version of the manuscript. The authors affirm that this work has not been published previously and is not under consideration for publication elsewhere. If accepted, the authors agree to the terms and conditions of publication as required by the journal. A preprint of this manuscript has been made available using the OSF website. (https://osf.io/preprints/xu8gb/). A previous iteration of this work has been presented as a poster at the 2023 EARLI conference.
Competing interests
The authors have no known conflict of interest to disclose. The authors have no relevant financial or non-financial interests to disclose.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
Aguilar, SJ; Rosenberg, JM; Greenhalgh, SP; Fütterer, T; Lishinski, A; Fischer, C. A Different Experience in a Different Moment? Teachers’ Social Media Use Before and During the COVID-19 Pandemic. AERA Open; 2021; 7,
Akkerman, SF; Meijer, PC. A dialogical approach to conceptualizing teacher identity. Teaching and Teacher Education; 2011; 27,
Amin, A; Roberts, J. Knowing in action: Beyond communities of practice. Research Policy; 2008; 37,
Arun, R., Suresh, V., Veni Madhavan, C. E., & Murthy, M. N. (2010). On finding the natural number of topics with latent dirichlet allocation: Some observations. Advances in Knowledge Discovery and Data Mining, 391–402
Avidov-Ungar, O; Hadad, S; Shamir-Inbal, T; Blau, I. Formal and informal professional development during different Covid-19 periods: The role of teachers’ career stages. Professional Development in Education; 2023; [DOI: https://dx.doi.org/10.1080/19415257.2023.2174163]
Bell, B. Using anecdotes in teacher development. International Journal of Science Education; 1994; 16,
Ben-Peretz, M., & Schonmann, S. (2000). Behind closed doors: Teachers and the role of the teachers’ lounge. SUNY Press
Bergviken Rensfeldt, A; Hillman, T; Selwyn, N. Teachers ‘liking’ their work? Exploring the realities of teacher Facebook groups. British Educational Research Journal; 2018; 44,
Bissessar, C. S. (2014). Facebook as an Informal Teacher Professional Development Tool. Australian Journal of Teacher Education, 39(2). https://doi.org/10.14221/ajte.2014v39n2.9
BMBF (2023). DigitalPakt Schule. https://www.digitalpaktschule.de/index.html
Booth, SE. Cultivating knowledge sharing and trust in online communities for educators. Journal of Educational Computing Research; 2012; 47,
Borchers, C., Klein, L., Johnson, H., & Fischer, C. (2023). Timing Matters: Inferring Educational Twitter Community Switching from Membership Characteristics. Proceedings of the 15th International Conference on Educational Data Mining (EDM). International Conference on Educational Data Mining (EDM), Bengaluru, India
Borchers, C., Rosenberg, J. M., Gibbons, B., Burchfield, M. A., & Fischer, C. (2021). To Scale or Not to Scale: Comparing Popular Sentiment Analysis Dictionaries on Educational Twitter Data. Proceedings of the 14th International Conference on Educational Data Mining (EDM). International Conference on Educational Data Mining (EDM), Paris, France
Bruguera, C., Guitert, M., & Romeu, T. (2019). Social media and professional development: A systematic review. Research in Learning Technology, 27, 1–18. https://doi.org/10.25304/rlt.v27.2286
Burchfield, M., Rosenberg, J., Borchers, C., Thomas, T., Gibbons, B., & Fischer, C. (2021). Are Violations of Student Privacy “Quick and Easy”? Investigating the Privacy of Students’ Images and Names in the Context of K-12 Educational Institution’s Posts on Facebook. Proceedings of the 14th International Conference on Educational Data Mining (EDM). Paris, France
Carpenter, J. Preservice teachers’ microblogging: Professional development via Twitter. Contemporary Issues in Technology and Teacher Education; 2015; 15,
Carpenter, J; Krutka, DG. How and why educators use Twitter: A survey of the field. Journal of Research on Technology in Education; 2014; 46,
Carpenter, J; Krutka, DG. Engagement through microblogging: Educator professional development via Twitter. Professional Development in Education; 2015; 41,
Carpenter, JP; Krutka, DG. Learning in 140 characters: English teachers’ educational uses of Twitter. International Journal of English and Education; 2015; 4,
Carpenter, J; Tani, T; Morrison, S; Keane, J. Exploring the landscape of educator professional activity on Twitter: An analysis of 16 education-related Twitter hashtags. Professional Development in Education; 2022; 48,
Carpenter, JP; Shelton, CC; Schroeder, SE. The education influencer: A new player in the educator professional landscape. Journal of Research on Technology in Education; 2023; 55,
Cohen, J. A coefficient of agreement for nominal scales. Educational and Psychological Measurement; 1960; 20,
Dede, C; Eisenkraft, A. Dede, C; Eisenkraft, A; Frumin, K; Hartley, A. Online and blended teacher learning and professional development. Teacher learning in the digital age: Online professional development in STEM education; 2016; Harvard Education Press: pp. 1-12.
Fischer, C; Fishman, B; Schoenebeck, SY. New Contexts for Professional Learning: Analyzing High School Science Teachers’ Engagement on Twitter. AERA Open; 2019; 5,
Fischer, C; Pardos, ZA; Baker, RS; Williams, JJ; Smyth, P; Yu, R; Slater, S; Baker, R; Warschauer, M. Mining big data in education: Affordances and challenges. Review of Research in Education; 2020; 44,
Fishman, B., Fischer, C., Kook, J., Levy, A., Jia, Y., Eisenkraft, A., McCoy, A., Lawrenz, F., Dede, C., & Frumin, K. (2014). Professional development for the redesigned AP Biology exam: Teacher participation patterns and student outcomes. 2014 Annual Meeting of the American Educational Research Association, Philadelphia, PA
Freeman, D. Renaming experience/reconstructing practice: Developing new understanding of teaching. Teaching and Teacher Education; 1993; 9,
Frumin, K; Dede, C; Fischer, C; Foster, B; Lawrenz, F; Eisenkraft, A; Fishman, B; Jurist Levy, A; McCoy, A. Adapting to large-scale changes in Advanced Placement biology, chemistry, and physics: The impact of online teacher communities. International Journal of Science Education; 2018; 40,
Fütterer, T; Hoch, E; Stürmer, K; Lachner, A; Fischer, C; Scheiter, K. Was bewegt Lehrpersonen während der Schulschließungen? – Eine Analyse der Kommunikation im Twitter-Lehrerzimmer über Chancen und Herausforderungen digitalen Unterrichts. Zeitschrift Für Erziehungswissenschaft; 2021; 24,
Gage, NA; Scott, T; Hirn, R; MacSuga-Gage, AS. The relationship between teachers’ implementation of classroom management practices and student behavior in elementary school. Behavioral Disorders; 2018; 43,
Greenhow, C., Galvin, S. M., Brandon, D. L., & Askari, E. (2020). A Decade of Research on K–12 Teaching and Teacher Learning with Social Media: Insights on the State of the Field. Teachers College Record: The Voice of Scholarship in Education, 122(6), 1–72. https://doi.org/10.1177/016146812012200602
Joksimovic, S; Gasevic, D; Kovanovic, V; Adesope, O; Hatala, M. Psychological characteristics in cognitive presence of communities of inquiry: A linguistic analysis of online discussions. The Internet and Higher Education; 2014; 22, pp. 1-10. [DOI: https://dx.doi.org/10.1016/j.iheduc.2014.03.001]
Keküllüoglu, D., Magdy, W., & Vaniea, K. (2020). Analysing Privacy Leakage of Life Events on Twitter. 12th ACM Conference on Web Science, 287–294. https://doi.org/10.1145/3394231.3397919
Keller, TR; Klinger, U. Social bots in election campaigns: Theoretical, empirical, and methodological implications. Political Communication; 2019; 36,
Khasawneh, M. The influence of Twitter hashtags in building supportive online communities for parents of children with special needs. Journal of Technology and Science Education; 2024; 14,
Kyndt, E; Gijbels, D; Grosemans, I; Donche, V. Teachers’ everyday professional development mapping informal learning activities, antecedents, and learning outcomes. Review of Educational Research; 2016; 86,
Landis, JR; Koch, GG. The measurement of observer agreement for categorical data. Biometrics; 1977; 33,
Lave, J; Wenger, E. Situated learning; 1991; Cambridge University Press: [DOI: https://dx.doi.org/10.1017/CBO9780511815355]
Lord, G; Lomicka, L. Twitter as a tool to promote community among language teachers. Journal of Technology and Teacher Education; 2014; 22,
Mawhinney, L. Let’s lunch and learn: Professional knowledge sharing in teachers’ lounges and other congregational spaces. Teaching and Teacher Education; 2010; 26,
McElduff, F., Cortina-Borja, M., Chan, S. K., & Wade, A. (2010). When t-tests or Wilcoxon-Mann-Whitney tests won't do. Advances in physiology education, 34(3), 128–133.
Meier, T., Boyd, R. L., Pennebaker, J. W., Mehl, M. R., Martin, M., Wolf, M., & Horn, A. B. (2019). “LIWC auf Deutsch”: The Development, Psychometrics, and Introduction of DE- LIWC2015 [2019]. PsyArXiv. https://doi.org/10.31234/osf.io/uq8zt
Mittendorff, K; Geijsel, F; Hoeve, A; De Laat, M; Nieuwenhuis, L. Communities of practice as stimulating forces for collective learning. Journal of Workplace Learning; 2006; 18,
Moreno, MA; Goniu, N; Moreno, PS; Diekema, D. Ethics of social media research: Common concerns and practical considerations. Cyberpsychology, Behavior, and Social Networking; 2013; 16,
Morina, F; Fütterer, T; Hübner, N; Zitzmann, S; Fischer, C. Effects of online teacher professional development on teacher-, classroom-, and student-level outcomes: A meta-analysis. Computers & Education; 2025; 228, [DOI: https://dx.doi.org/10.1016/j.compedu.2025.105247] 105247.
Mosquera-Gende, I., Marcelo-Martínez, P., Postigo-Fuentes, A. Y., & Fernández-Navas, M. (2024). The Hashtag #CharlasEducativas as a Teacher Affinity Space on Twitter. Comunicar, 32(78), 222–233. https://doi.org/10.58262/V32I78.18
Naomi, J. F., Vasanthageethan, A., Roshini, G., & Kumar, J. S. (2021). Data Privacy Preserving Recommendations for Social Media. 2021 7th International Conference on Advanced Computing and Communication Systems (ICACCS
Peböck, K. (2020).Peböck, K. (2020). #relichat-informelles Lernen mit Twitter: Religionslehrer* innenfortbildung durch sozial-konstruktivistische Vernetzung in Communities of Practice [Doctoral dissertation, University of Passau]. https://opus4.kobv.de/opus4-uni-passau/frontdoor/index/index/docId/828
Pennacchiotti, M; Popescu, A-M. A machine learning approach to Twitter user classification. Proceedings of the International AAAI Conference on Web and Social Media; 2021; 5,
Powell, CG; Bodur, Y. Teachers’ perceptions of an online professional development experience: Implications for a design and implementation framework. Teaching and Teacher Education; 2019; 77, pp. 19-30. [DOI: https://dx.doi.org/10.1016/j.tate.2018.09.004]
Ramlo, S. Inservice science teachers’ views of a professional development workshop and their learning of force and motion concepts. Teaching and Teacher Education; 2012; 28,
Rauchfleisch, A; Kaiser, J. The false positive problem of automatic bot detection in social science research. PLoS ONE; 2020; 15,
Recker, M; Walker, A; Giersch, S; Mao, X; Halioris, S; Palmer, B; Johnson, D; Leary, H; Robertshaw, MB. A study of teachers’ use of online learning resources to design classroom activities. New Review of Hypermedia and Multimedia; 2007; 13,
Rehm, M; Notten, A. Twitter as an informal learning space for teachers!? The role of social capital in Twitter conversations among teachers. Teaching and Teacher Education; 2016; 60, pp. 215-223. [DOI: https://dx.doi.org/10.1016/j.tate.2016.08.015]
Remus, R., Quasthoff, U., & Heyer, G. (2010). SentiWS – a Publicly Available German-language Resource for Sentiment Analysis. In LREC
Richter, D., Kunter, M., Klusmann, U., Lüdtke, O., & Baumert, J. (2011). Professional development across the teaching career: Teachers’ uptake of formal and informal learning opportunities. Teaching and teacher education, 27(1), 116–126.
Riquelme, F; González-Cantergiani, P. Measuring user influence on Twitter: A survey. Information Processing & Management; 2016; 52,
Risser, SH. Virtual induction: A novice teacher’s use of Twitter to form an informal mentoring network. Teaching and Teacher Education; 2013; 35, pp. 25-33. [DOI: https://dx.doi.org/10.1016/j.tate.2013.05.001]
Rosenberg, JM; Greenhalgh, SP; Koehler, MJ; Hamilton, ER; Akcaoglu, M. An investigation of state educational Twitter hashtags (SETHs) as affinity spaces. E-Learning and Digital Media; 2016; 13,
Rosenberg, JM; Reid, JW; Dyer, EB; Koehler, M; Fischer, C; McKenna, TJ. Idle chatter or compelling conversation? The potential of the social media-based #NGSSchat network for supporting science education reform efforts. Journal of Research in Science Teaching; 2020; 57,
Rosenberg, JM; Borchers, C; Dyer, EB; Anderson, D; Fischer, C. Understanding Public Sentiment About Educational Reforms: The Next Generation Science Standards on Twitter. AERA Open; 2021; 7,
Rosenberg, J; Burchfield, M; Borchers, C; Gibbons, B; Anderson, D; Fischer, C. Social media and students’ privacy: What schools and districts should know. Phi Delta Kappan; 2021; 103,
Rosenberg, JM; Borchers, C; Burchfield, MA; Anderson, D; Stegenga, SM; Fischer, C. Posts about students on Facebook: A data ethics perspective. Educational Researcher; 2022; 51,
Rosenberg, JM; Borchers, C; Stegenga, SM; Burchfield, MA; Anderson, D; Fischer, C. How educational institutions reveal students’ personally identifiable information on Facebook. Learning, Media and Technology; 2022; [DOI: https://dx.doi.org/10.1080/17439884.2022.2140672]
Said, SB. A lighthouse on the beach’ stories, metaphors, and anecdotes in teachers’ personal recounts. Bellaterra Journal of Teaching & Learning Language & Literature; 2013; 6,
Siemens, G., & Baker, R. (2012). Learning analytics and educational data mining: Towards communication and collaboration. Proceedings of the 2nd International Conference on Learning Analytics and Knowledge, 252–254. http://dl.acm.org/citation.cfm?id=2330661
Skiba, R; Ormiston, H; Martinez, S; Cummings, J. Teaching the social curriculum: Classroom management as behavioral instruction. Theory into Practice; 2016; 55,
Sörensen, I; Fürst, S; Vogler, D; Schäfer, MS. Higher education institutions on Facebook, Instagram, and Twitter: Comparing Swiss universities' social media communication. Media and Communication; 2023; 11,
Staudt Willet, KB. Revisiting how and why educators use Twitter: Tweet types and purposes in #Edchat. Journal of Research on Technology in Education; 2019; 51,
Staudt Willet, KB. Early career teachers’ expansion of professional learning networks with social media. Professional Development in Education; 2024; 50,
Tang, J., Meng, Z., Nguyen, X., Mei, Q., & Zhang, M. (2014). Understanding the Limiting Factors of Topic Modeling via Posterior Contraction Analysis. International Conference on Machine Learning
Thelwall, M; Buckley, K; Paltoglou, G; Cai, D; Kappas, A. Sentiment strength detection in short informal text. Journal of the American Society for Information Science and Technology; 2010; 61,
Visser, RD; Evering, LC; Barrett, DE. #TwitterforTeachers: The implications of Twitter as a self-directed professional development tool for K–12 teachers. Journal of Research on Technology in Education; 2014; 46,
Wang, Y; Fikis, DJ. Common core state standards on Twitter: Public sentiment and opinion leaders. Educational Policy; 2019; 33,
Waters, RD; Williams, JM. Squawking, tweeting, cooing, and hooting: Analyzing the communication patterns of government agencies on Twitter. Journal of Public Affairs; 2011; 11,
Wenger-Trayner, E., & Wenger-Trayner, B. (2015). Communities of practice: A brief introduction. https://wenger-trayner.com/wp-content/uploads/2015/04/07-Brief-introduction-to-communities-of-practice.pdf
Wesely, PM. Investigating the community of practice of world language educators on Twitter. Journal of Teacher Education; 2013; 64,
Williams, ML; Burnap, P; Sloan, L. Towards an ethical framework for publishing Twitter data in social research: Taking into account users’ views, online context and algorithmic estimation. Sociology; 2017; 51,
Yue, L; Chen, W; Li, X; Zuo, W; Yin, M. A survey of sentiment analysis in social media. Knowledge and Information Systems; 2019; 60,
© The Author(s) 2025. This work is published under http://creativecommons.org/licenses/by/4.0/ (the “License”). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.