Introduction

The continued trend towards digital technologies in educational contexts not only applies to classroom teaching but also to teacher learning. In particular, many teachers engage in online professional development (OPD) activities on social media platforms to share and access resources, connect with other teachers, and increase their knowledge and skills (Bruguera et al., 2019; Fütterer et al., 2021; Greenhow et al., 2020).

Within the social media landscape (e.g., Facebook, Instagram, Pinterest, LinkedIn), Twitter (now known as “X”) is one of the most used microblogging platforms in the United States (Pew Research Center, 2021; see also Aguilar et al., 2021) that allows users to send short messages (up to 280 characters at the time of the study; called tweets in Twitter and now posts in X) to an audience of followers. Whereas research has gained insight how teachers use Twitter and other social media platforms for their professional development (PD), questions about the value of teachers’ participation in social media-based communities remain insufficiently answered. In particular, questions about how social media-based OPD opportunities share characteristics of high-quality PD opportunities remain. A key reason for this relates to the primarily employed descriptive methodologies. For example, early Twitter studies used surveys and interviews to find that teachers utilize Twitter because of its affordances for community and relationship building and opportunity for collaboration with colleagues (Carpenter & Krutka, 2015; Trust et al., 2016; Wesely, 2013). However, the nature of these collaborations on Twitter is understudied, although teacher collaboration has often been viewed as an important factor for improving job satisfaction, teaching practices, and student performance (Methlagl, 2022; Vangrieken et al., 2015; Vescio et al., 2008).

Methodologically, social media allows us to gain insights in the “black box” of teacher collaborations through the collection of time-stamped interaction data (tweets), which is not typically captured through traditional survey-based research methods. Epistemic network analysis (ENA) can provide insights into the relationships between tweet content and examine relationships across conversations (Shaffer, 2017). These relationships can be visualized by the frequency of co-occurrence of different content themes within the same tweet and across multiple tweets in the same thematic conversations, thereby showing which themes teachers frequently discuss together (e.g., discussing labs and sharing resources they use). Furthermore, these content co-occurrences can be compared based on teachers’ activity on social media. Thus, differences between what themes are discussed by active teachers in the platform can be examined (e.g., discussing assessments in their specific teaching context). Notably, research in contexts outside of social media also indicates the potential benefits of using ENA to investigate teacher collaboration during PD activities by comparing the co-occurrences of collaborative problem-solving dimensions between high- and low-performing groups (Zhang et al., 2022). In the context of collaboration, an advantage of ENA over traditional methods that use the isolated frequencies of thematic codes is modeling their co-occurrences and analyzing the interaction between different individuals (Zhang et al., 2022).

Similarly, research on teacher participation on social media seldomly examines the potential effectiveness of teachers’ social media engagement, for instance, by measuring how these platforms have characteristics of high-quality PD. Such high-quality PD characteristics may include the content focus of the PD activity and its coherence with teachers’ personal and contextual needs (Desimone, 2009).

Therefore, in this study, we investigate the use of Twitter as a platform for teacher OPD by analyzing engagement, content, and collaboration patterns within Advanced Placement (AP) Biology communities. Utilizing advanced conversational and epistemic network analyses, the research seeks to determine if Twitter can offer a high-quality learning environment for teachers by fostering meaningful exchanges and supporting a community of practice.

Theoretical Background

Features of Effective Professional Development

Decades of research on the effectiveness of teacher PD led to a consensus on important features for successful PD activities (Borko et al., 2010; Darling-Hammond et al., 2017; Desimone, 2009). For instance, such high-quality features include that PD affords active learning, collective participation, and a sufficient overall duration, as well as coherence with teachers’ individual contexts and a content focus of the PD activity (Desimone, 2009). Active learning describes opportunities for teachers to participate in active engagements, for instance, to lead discussions, review and reflect on instructional materials or student work. Collective participation describes how teachers from the same grade-level, subject area, or school engage collaboratively in learning activities. Duration refers to both the frequency and the overall timespan of engagement in the learning activity. Coherence refers to how PD activities relate to curricular standards, individual contexts, and teachers’ prior knowledge and beliefs. Content focus describes how PD activities enable teachers to increase expertise in relevant knowledge domains.

Whereas the importance of these high-quality PD features are well documented for traditional PD activities such as face-to-face workshops or online courses (e.g., Desimone & Garet, 2015; Fischer et al., 2018; Fütterer et al., 2024), the research base examining how online communities may afford these high-quality PD features is limited. An exception is initial work by Fischer et al. (2019) indicating that Twitter communities may indeed fulfill the high-quality PD features of collective participation and duration for some users. This study extends this earlier work, particularly focusing on how participation on Twitter may have the characteristics of coherence and content focus.

In online learning, content focus often describes activities such as discussions about how students can learn best in a subject or questions about appropriate teaching materials for a particular lesson. For example, Fütterer et al. (2021) found that teachers on Twitter discuss how high-quality, technology-enhanced teaching can be achieved during the COVID-19-induced school closures in a way that is effective for student learning. Similarly, coherence describes activities like discussions about what teachers think about their role in their profession or workplace. For instance, Davis (2015) found that teachers discuss teaching philosophy, reflecting their role as a teacher, and decisions by policy makers on Twitter.

In this study, we focus on these two characteristics (i.e., consent focus and coherence) as they align closely with the nature of professional exchanges and learning opportunities available on social media platforms like Twitter.

Online Teacher Professional Development

Online teacher professional development (OPD) comes in many forms and shapes, often varying in their degrees of formality and agency. To classify OPD activities, Dede and Eisenkraft (2016) suggested a continuous scale. One end of this scale is represented by formal and structured online activities such as courses that offer Continuing Education Units like the Massively Open Online Courses for Educators (MOOC-Ed) initiative (Akoglu et al., 2019; Kleiman & Wolf, 2016) or micro-credentials programs that acknowledge teachers for their participation in self-directed learning (Fishman et al., 2018). Given their informal and unstructured nature, the other end is represented by online teacher communities. In this study, online teacher communities are defined as digital platforms or spaces where educators convene virtually to engage in professional discussions, share resources, collaborate on projects, and seek support from peers, often facilitated through dedicated websites, forums, social media groups, or virtual spaces designed for teachers.

Notably, there is substantial heterogeneity within online communities as some are tied to prominent educational institutions like the NSTA Learning Center community (Brunsell & Horejsi, 2012) or the College Board’s Advanced Placement Teacher Community (Frumin et al., 2018), whereas others organically evolve on social media platforms such as Reddit, Facebook, or Twitter (Fischer et al., 2019; Ranieri et al., 2012; Staudt Willet, 2019; Staudt Willet & Carpenter, 2021). This study is situated in the latter end by examining teacher participation on Twitter—an online community that is informal and less-structured than many other OPD opportunities.

Affordances to participate in informal online teacher communities are manifold. For instance, they extend temporal and geographic boundaries allowing teachers to interact with others at any time, from anywhere, and for a flexible duration given their personal needs and preferences. That said, some Twitter communities host synchronous chat sessions (e.g., #NGSSchat) for teachers to virtually meet up at the same time to discuss topics of interest (Rosenberg et al., 2020). Interactions within these online communities are usually stored on the platforms meaning that users may also revisit them at any time necessary, for instance, teachers catching up on a #NGSSchat session they previously missed. This also leads to a greater ability to personalize learning. Teachers may choose to deliberately participate in topics of interest or use certain features of these communities. For instance, some teachers may choose to download some teaching resources without any further personal interaction whereas others frequently engage in conversations about teaching practices and curricular standards. Notably, the sentiment of tweets can offer insights in the importance and impact of topics on teachers’ beliefs and practices (Coburn, 2001; Rosenberg et al., 2021). For example, Rosenberg et al. (2021) examined teachers’ sentiment towards the NGSS reform and explored factors explaining variance in the public opinion regarding the reform.

Additionally, Twitter offers personalized home feeds so that the initial information displayed to teachers is personalized based on their prior usage history. These opportunities for personalized content are also supported by the high breadth and depth of information available as diverse user groups participate in these communities (Rosenberg et al., 2020), which may contrast more homogenous participation groups in traditional face-to-face activities that may, for instance, be stratified by school districts. Similarly, most online communities are free to use, contrasting PD programs costing hundreds to thousands of dollars in participation, travel, and lodging fees, which may limit teachers’ ability to access these learning activities potentially widening already existing equity gaps. Furthermore, online communities offer just-in-time PD allowing teachers to access information at the time they need them—such as resources about preparing a particular experiment for a class they need to teach the day after their post, which contrasts most traditional just-in-case PD activities that are often offered in the summer before the school year (Greenhalgh & Koehler, 2017). Therefore, it is no surprise that teachers are increasingly integrating participation in online communities to their regular portfolio of professional learning activities.

Research investigating the effectiveness of OPD has indicated their potential to enhance teacher knowledge (e.g., Reeves & Chiang, 2019), influence teaching practices (e.g., Marquez et al., 2016), and improve student achievement (e.g., Fisher et al., 2010; Fishman et al., 2013; for a meta-analysis regarding the effectiveness of OPD on all three levels, see Morina et al., 2023). Notably, some research studies also identified positive effects of online teacher community participation on student performance (e.g., Fishman et al., 2014; Frumin et al., 2018).

Collaborative Learning on Social Media

Teacher collaboration has been conceptualized and studied from many different perspectives (Lavié, 2006; Vangrieken et al., 2015). Teacher collaboration may range from simply sharing information to co-constructive processes and joint production of instructional materials (Gräsel et al., 2006; Vangrieken et al., 2015). Notably, teachers often consider collaboration with their colleagues as important (Hartmann et al., 2021; Richter & Pant, 2016; Vangrieken et al., 2015). This is mirrored in findings that suggest that collaborative learning is often viewed critical for high-quality PD (Darling-Hammond et al., 2017; Desimone, 2009) and overall school success (Darling-Hammond et al., 2009).

Learning on social media can relate to situated learning in communities of practice (CoPs; Lave, 1991; Wenger, 1998)—which are often seen as prime examples of successful learning environments. CoPs occur when teachers who share a common concern or topic regularly interact with each other to engage in professional learning (Wenger-Trayner & Wenger-Trayner, 2015). Notably, CoP membership implies explicit commitment to a common concern or topic. Therefore, participants in CoPs are not random groups of people but often a professional network of participants possessing a wealth of distributed expertise. Notably, even teachers who may not actively contribute new content to these online communities but engage by reading and downloading resources (also called lurkers) are also important participants in the overall community landscape (Bozkurt et al., 2020; Edelmann, 2013; Preece et al., 2004). Overall, teachers may utilize online communities, regardless of whether they actively contribute to discussion or more passively consume the content, as part of their professional learning network (PLN; Bruguera et al., 2019; Greenhow et al., 2020). A defining feature of such PLNs are the opportunities for teachers to interact and collaborate with other teachers and educational stakeholders to advance their professional skills. In that, these aspects of collaborative learning can be viewed related to the collective participation PD feature of Desimone’s (2009) model for studying the effectiveness of PD.

A key aspect of teacher collaboration in learning communities are collaborative conversations, which can be described by following characteristics (see Feldman, 1999):

  • At least two persons are involved who exchange words (dialog).

  • A dialog between partners at the same level.

  • Interactions that are dovetailed (i.e., exchange of views, connected remarks, or utterances), meaning that connections of words go beyond simple chit-chat.

These collaborative conversations are related to positive learning effects in the creation of new knowledge (Hartmann et al., 2021; Vangrieken et al., 2015). Whereas the extent of collaboration among colleagues in a school often depends on contextual factors such as school-leader support (Honingh & Hooge, 2014), social media platforms provide native outlets to enable collaborative conversations that may lead to substantial professional growth and collaborations on challenges and goals over longer periods of time (Gröschner et al., 2018; Teräs, 2016). Notably, conversations in online social media environments are different than those that take place in-person. In-person, two people may communicate directly through turn-taking. Through social media platforms, though, conversations are often asynchronous (Chen et al., 2017), as users can respond to messages with a substantial temporal delay.

The availability of digital trace data (Fischer et al., 2020; Welser et al., 2008) capturing these conversations at-scale affords new ways of understanding engagement and (collaborative) learning patterns that were not feasible to examine through traditional survey- or interview-based research methodologies. With this novel data source, new methods such as epistemic network analysis are needed to reconstruct the relation between messages, for instance, via text similarity and latent semantic transferability, which are at the heart of this study.

Whereas all the concepts mentioned above emphasize different aspects of social learning and interaction, we address our research questions in this study in the context of online communities in which conversations can take place.

Brief Introduction to Epistemic Network Analysis

Epistemic network analysis (ENA) models associations between a set of thematically connected codes in discourse (Shaffer, 2017, 2018). ENA draws upon the general set of network analytic methods (e.g., Carolan, 2014). Different from how network analysis has traditionally been used in educational research, ENA does not focus on relationships between individuals, but, instead, the relationships between ideas. The central assumption is that the structure of connections between different text elements (i.e., tweets on Twitter) provides more information than each element individually (Shaffer et al., 2016). In contrast to comparing the isolated frequencies of individual elements, the strength of ENA lays in modeling the frequencies of co-occurrence between all elements. To establish that a relationship between two elements exists, they need to co-occur in the same text window (i.e., conversation). The connections between elements are represented in dynamic network models, which visualize the structure and the strength of the connections between different codes. In the context of our study, the frequencies of co-occurrence between multiple tweet content categories across thematically related tweets (i.e., conversations) can be visualized in the same network model. Furthermore, the strength of these connections can be compared between networks of different people or groups, for example, active and non-active teachers on Twitter or based on other tweet characteristics, like the tweet sentiment.

In prior research, ENA has been used in tandem with sentiment analysis to model political discourse on Twitter (Misiejuk et al., 2021) or in combination with social network analysis to identify associations between teachers’ agency and inclusive pedagogy (Pantić et al., 2022). Furthermore, ENA was used to compare teachers’ technological pedagogical content knowledge (Zhang et al., 2019) and teachers’ online collaborative learning (Zhang et al., 2022). However, the role of teacher’s engagement on social media for their professional learning has not yet been investigated using methods that can quantify the relationships between different content categories. In particular, our study intends to capture collaboration patterns by examining the relationships between tweet content in teachers’ conversations on Twitter and compare them based on the tweet sentiment and teachers’ activity level.

Research Questions

As there are debates about the value of participation in social media-based communities, this study examines how teachers’ Twitter use may align to characteristics of high-quality OPD opportunities—specifically their coherence and content focus. We focus on these two characteristics as they align closely with the nature of professional exchanges and learning opportunities available on social media platforms like Twitter. Coherence, referring to the OPD’s alignment with a teacher’s professional context and prior knowledge, and content focus, emphasizing the relevance and depth of the OPD content, are both readily observable and important in the context of brief, text-based interactions typical of social media. Moreover, given the scope of our study and the constraints associated with analyzing social media data, we aimed to ensure a more targeted lens with focused analyses that allow us to generate nuanced insights into how teachers engage with OPD activities on Twitter.

To do this, we expand on the commonly used descriptive approaches in the literature as coherence requires an understanding of the connections between individual teachers’ local context and their specific needs. Therefore, we apply ENA, a methodology designed to reveal the connections among the content discussed through the hashtags we studied. Specifically, we identify and understand co-occurrences of topics within the conversations on Twitter. Also, given that content focus is essential for effective OPD, we examined the relation of the emotional valence of tweets to the content to understand how teachers related—positively or negatively—to the content discussed to gain insights on potential impacts on teachers’ beliefs and practices. Notably, prior research alluded to the potential of sentiment analysis of social media data to provide indications of successful reform implementations (Rosenberg et al., 2021; Wang & Fikis, 2019). Lastly, as teachers who participate to differing extents through the hashtags may differentially benefit, we also examine how teachers’ frequency of engagement in the communities relate to the content and connections of the conversations. Together, our research questions (RQs) are as follows:

  • RQ 1: How do the connections between the tweet content shared within teachers’ conversations evidence coherence?

  • RQ 2: How do connections between tweet content differ by tweet sentiment?

  • RQ 3: How do connections between tweet content differ based on teachers’ overall engagement?

Method

Study Context and Sample

This study is situated in the context of the Advanced Placement (AP) reform in Biology. AP Biology courses are offered by the College Board and provide high school students with opportunities to engage in college-level biology curriculum. In AP Biology, students explore foundational concepts in biology such as cellular structure and function, genetics, evolution, and ecological systems. The course is designed to be rigorous and comprehensive, preparing students for further study in biology-related fields at the collegiate level. The AP Biology curriculum was revised to shift the curricular focus on algorithmic procedures (routine sequence of steps to solve a given problem; for example, protein synthesis), rote learning and memorization to inquiry learning, higher-order cognitive skills, and deeper understandings of science concepts and practices (The College Board, 2012). Students often consider AP exams as high stakes because of perceived benefits for college admission, possibilities to replace introductory college courses with AP credit, and preparation for college success (Ackerman et al., 2013; Fischer et al., 2022; Sadler & Sonnert, 2010). Thus, teachers and school administrations often place great value on AP courses, promoting a need for professional learning during times of large reforms.

Notably, the AP Biology reform is part of College Board’s three-part AP science reform effort that also includes reforms to the AP Chemistry and AP Physics curriculum with similar changes across subject areas. As the AP Biology reform took place before the AP Chemistry and AP Physics reforms, this study focuses on the AP Biology reform to capture teachers’ initial responses to this reform.

This study mined the full public tweet history from three commonly used Twitter hashtags of AP Biology teachers, #apbiochat, #apbioleaderacademcy, and #apbioleaderacad, from the inception of each hashtag (July 2011) until June 14, 2016 (4 weeks after the 2016 AP Biology exam) using the Twitter API and custom Python scripts. Notably, this study purposefully examines the core time leading up the AP Biology reform (in which instructional materials were already shared by the College Board) and during the first three years of the reform effort. This led to a total of 121 users and 2,253 tweets with 2,040 tweets sent by teachers (Nteachers = 93).

As this study examines teachers’ participation in the communities, we use the teacher data set only (N = 2,040 tweets, N = 93 teachers). Teachers were identified through a qualitative, manual examination of user biographies and their tweet history. The mean number of tweets of teachers in this community was 21.90 tweets (SD = 58.90, Mdn = 3, range = 434). However, about 60% of teachers contributed five or less tweets to this community and 33 teachers posted only one tweet. Teachers were posting an average of 1.11 tweets/day and the average time between the first and last tweet (i.e., lifespan) for all teachers was 143 days (SD = 231.50, Mdn = 4, range = 888), and 222 days (SD = 256.70, Mdn = 118.50, range = 888) for teachers with two or more tweets.

Measures

We created measures on multiple levels to capture different information regarding teacher participation patterns. This includes tweet-level measures (unit of analysis is a single tweet), conversation-level measures (unit of analysis is a sequence of tweets), and user-level measures (unit of analysis is all tweets from a single user).

We created tweet-level measures capturing both tweet content and sentiment. Tweet content provides the more direct source of information to understand conversational patterns. Sentiment can complement this information by capturing more underlying accounts of expressed information including emotionality and valence (Rosenberg et al., 2021; Tausczik & Pennebaker, 2010).

To capture tweet content, we employed a two-cycle coding strategy (Miles et al., 2014). The first cycle involved exploratory descriptive coding of broader tweet content categories (e.g., instruction, resources, school context). The second cycle applied sub-coding to further distinguish between tweet content categories and to develop a code book. This led to seven dichotomous tweet content variables (1: tweet content feature is present, 0: tweet content feature is not present): (a) AP Biology content knowledge, (b) share resources, (c) seek information, (d) organize professional learning, (e) curriculum elements, (f) labs, (g) assessments, and (h) teaching context. Afterwards, three graduate student coders were trained to utilize this code book to code a subset of 225 randomly selected tweets. Through multiple discussion and iterative coding steps that included think-aloud protocols to better understand the coding procedures, which also led to small modifications of the code book, we were able to increase the interrater reliability for tweet content to an average Cohen’s Kappa (κ) of κ = 70 to satisfy traditional benchmarks representing “substantial” agreement (Landis & Koch, 1977). Table 2 in the Appendix shows exemplary tweets for each tweet content category.

Tweet sentiment was deductively coded through the assignment of positive, neutral, and negative sentiment ratings to the tweets. Although tweet sentiment coding could have been automated using natural language processing tools (Thelwall et al., 2010), we opted to use human coders as they are often seen as the gold standard in sentiment coding and our sample size was sufficiently sized to leverage human coders (Borchers et al., 2021). For tweet sentiment, this led to a Cohen’s Kappa of κ = 0.65 between the three raters meeting satisfying benchmarks of “substantial” agreement (Landis & Koch, 1977). Table 3 in the Appendix shows exemplary tweets for each tweet sentiment category.

On the conversation-level, we identified conversations by classifying tweets appearing in specific timeframes by thematic conversations. A thematic conversation was defined through a user introducing a new topic, idea, question, or experience that received interactions from other users within 48 h. This means that wherever users introduced a new topic, idea, question, or experience, we coded this as a new conversation. More detailed descriptions on the coding process and rules are described in the Appendix. All tweets were manually coded by two trained research assistants showing “substantial” agreement with a Cohen’s Kappa of κ = 0.69 (Landis & Koch, 1977).

On the user-level, we created two continuous variables capturing the number of tweets a user posted in the community and the users’ lifespan (i.e., number of days between first and last tweet in communities). Teachers were classified as active teachers or non-active teachers based on the number of tweets they posted (10 or more tweets) and their lifespan (100 or more days).

Analytic Methods

We applied ENA on the identified thematic conversations using the ENA Web Tool (version 1.7.0; Marquart et al., 2018; Shaffer, 2017). In that, the connections between codes are identified based on their co-occurrences within conversations. Using binary summation, the individual tweet networks are aggregated, and each line indicates whether two codes co-occurred in a tweet. These co-occurrences help us to understand thematic connections within tweets. Before dimensional reduction, all networks are normalized to account for the different number of codes among different units of analysis. ENA uses singular value decomposition (SVD) for dimensional reduction, which results in orthogonal dimensions (i.e., X- and Y-axis) that explain the maximum variance (Shaffer et al., 2016). The resulting one or two-dimensional values can be used for quantitative and qualitative comparisons between groups. For example, we provide exemplary tweets to illustrate the relations among content categories. Please note these exemplary tweets are synthetically created (and retain their original meaning) following recommendations for ethical use of qualitative data from social media to protect users’ anonymity (Moreno et al., 2013; Williams et al., 2017). Additionally, we can compare the strength of associations across different groups (tweet sentiment, teacher activity level) by looking at the subtracted networks. This is possible as the position of the nodes is fixed, and their position is determined by minimizing the difference between the nodes and the corresponding network centroids.

To answer RQ1, we modeled the relationships across tweet content categories in teachers’ Twitter conversations. Because we looked at all tweets, regardless of sentiment (or any other classifying variable), we describe how the different tweet content categories (i.e., AP Biology content knowledge, share resources, seek information, organize professional learning, curriculum elements, labs, assessments, and teaching context) were related across all tweets.

To answer RQ2, we examined tweet sentiment (positive, neutral, negative) to compare the strength of the relationships between different tweet content categories. Mann-Whitney U tests compared the position of the centroids depending on tweet sentiment (neutral vs. negative, neutral vs. positive, positive vs. negative). Subtracted networks explained qualitative differences in tweet content categories by tweets sentiment.

To answer RQ3, we examined teachers’ activity to compare the strength of the relationships between tweet content categories. We used mean rotation, which maximizes the difference between active and non-active teachers. Then, we compared the position of the centroids using Mann-Whitney U tests.

Results

Connections Across Tweet Content Categories (RQ1)

Figure 1a provides an overall qualitative depiction of the strength of co-occurrence between tweet content categories across all tweets. Stronger co-occurrences are depicted with thicker lines connecting the different tweet content categories. Notably, we found strong co-occurrences between teaching context (teachcontext) with labs (labs), seeking information (seekinfo), assessments (assessments), and sharing resources (shareresources). This indicates the prominence of teaching context with many other important content categories in teachers’ tweets which also alludes to coherence of teachers’ experiences and discussed topics with their individual school contexts.

Fig. 1
figure 1

The epistemic network of all tweets and the subtracted networks of tweets based on sentiment. Note. a All tweets regardless of sentiment; b neutral (blue) and negative (purple); c neutral (blue) and positive (green); d positive (green) and negative (purple)

Tweet Content and Sentiment (RQ2)

We compared the strength of co-occurrence among tweet content categories between tweets with negative, neutral, and positive sentiment using three comparisons, neutral vs. negative, neutral vs. positive, and positive vs. negative. The frequencies of tweet sentiment across tweet content categories are provided in Table 1.

Table 1 Descriptive information of tweet content categories per tweet sentiment

We found a statistically significant difference comparing neutral to negative tweets along the x-axis (SVD1), U = 72,787, p < .001, r = .16. This difference stems from stronger co-occurrences between teaching context (teachcontext), assessments (assessments), and labs (labs) in the negative tweets’ network. An example tweet with a negative sentiment for teaching context and labs is “@USER Grading lab reports is challenging when students don’t understand them. Wondering what to do to support them better. #apbioleaderacademy.” An example tweet with a negative sentiment for teaching context and assessments is “#apbiochat SBG is hard for me this year. my students also don’t seem to care about ungraded homework assignments.” In the neutral tweets’ network, there were stronger co-occurrences between sharing resources (shareresources) and content knowledge (contentknowlege) and slightly stronger co-occurrences between sharing resources (shareresources) and seeking information (seekinfo; Fig. 1b). An example tweet with a neutral sentiment for sharing resources and content knowledge is “@USER check out this photosynthesis simulation http://LINK.com#apbio.”

Similarly, neutral and positive tweets were statistically significantly different along the x-axis (SVD1), U = 421,491, p < .001, r = − .11. Overall, all co-occurrences were more frequent in the neutral tweets (blue lines, Fig. 1c). In particular, we found co-occurrences between teaching context (teachcontext), seeking information (seekinfo), assessments (assessments), and curriculum (curriculum). An example of a tweet with a neutral sentiment combining teaching context and seeking information is “@USER So you provide them with a info sheet on the key characteristics and then you ask questions? Or do they need to know more? #apbiochat.” No co-occurrences were stronger in the positive tweets’ network (green) compared to the neutral tweet network (blue).

Also, negative and positive tweets were statistically significantly different on the x-axis (SVD1), U = 49,933, p < .001, r = − .29. Negative tweets had more frequent co-occurrences between teaching context (teachcontext) and seeking information (seekinfo), assessments (assessments), labs (labs), and curriculum (curriculum) compared to positive tweets. An example tweet with a negative for seeking information and teaching context is “Someone here who used the Carolina Cell Communication Lab? Looking for advice, this lab did NOT work for my class. #APBioChat.” In contrast, positive tweets showed more frequent co-occurrences between sharing resources (shareresources) and content knowledge (contentknowledge; Fig. 1d). An example tweet with a positive sentiment is “Love when video game animations help students visualize microscopic processes. http://LINK.com#apbiochat.”

Tweet Content and Teacher Engagement (RQ3)

Comparing active and non-active teachers, we found a statistically significant difference along the x-axis (MR1), U = 240,176, p < .001, r = − .19. Active teachers had more frequent co-occurrences between teaching context (teachcontext), seeking information (seekinfo), labs (labs), assessments (assessments), and curriculum (curriculum), as indicated by the green lines in Fig. 2. For example, a conversation among active teachers involving teaching context and labs looked as follows:

Fig. 2
figure 2

The subtracted network of active teachers (green) and non-active teachers (purple)

@USER @USER It depends on the discussion during the lab. They also do group presentations together #apbiochat

@USER #apbiochat I like this lab for inquiry learning and group pres.

@USER Also, check out my students’ presentations at http://LINK.com/ #apbiochat.

That is, more active teacher used Twitter more regarding on key issues related to instructional implementations and their individual school and classroom contexts. For instance, active teachers requested more resources related to student learning, shared more information about specific laboratory, or provided information about summative and formative assessment strategies related to the AP examination compared to non-active teachers. In contrast, non-active teachers had more frequent co-occurrences between organizing professional development activities (organizepd) and sharing resources (shareresources) as indicated by the purple lines in Fig. 2. That is, non-active teachers often take organizational roles, for instance, to moderate and initiate exchanges between teachers. For example, a conversation among non-active teachers illustrating organizing professional development and sharing resources:

@USER @USER #SBG is the future of grading! hearing so much positive about it #apbiochat

@USER You may also want to check #apbiochat at 8PM CST on Wednesdays

Discussion

This study examined conversational structures in informal teacher communities on Twitter. We investigated how teacher participation on Twitter may fulfill characteristics of high-quality PD (i.e., coherence and content focus) to understand the extent to which social media serves as a potentially beneficial avenue for professional learning. Notably, coherence and content focus are important emphases as coherence pertains to how PD activities and teachers’ curricular standards or teaching goals relate, while content focus refers to how the PD enhances teachers’ expertise in their domain (Desimone, 2009). In both cases, relational aspects are key. The pattern of our ENA findings—we observed a co-occurrence of tweets about the PD related to teachers’ teaching contexts with other topics of discussion—suggests to us that the content of these tweets, which often included needs, suggestions, and experiences, was inspired by the individual teaching contexts of the participants. Therefore, regarding coherence, we infer that the epistemic connections between teachers’ talk about the PD and their standards or goals is of interest. For content focus, the valence of teachers’ talk in relation to content can reveal the extent to which they believe the PD adequately demonstrates this characteristic. Approaches that describe teachers’ talk about PD and their standards or goals, for instance, may fail to capture the ways teachers draw connections between these two elements. The three core findings of this study follow.

First, teachers use Twitter as a platform coherent to their individual learning needs. This is evidenced by the way conversations about content related to the AP program often mentioned teachers’ individual contexts. For example, when teachers sought information, talked about laboratory investigations, or discussed assessment strategies like standard-based grading, these conversations referred to teachers’ unique contexts in their schools and classrooms. This finding is in line with prior research suggesting that a core affordance of Twitter is the ability to personalize and support learning (Fischer et al., 2019; Greenhow et al., 2020). This finding is also in accordance with early survey and interview studies that asked teachers about their perceived benefits of Twitter use, which found that teachers noted the differentiation and personalization afforded by the platform as key features (Carpenter & Krutka, 2014, 2015). This study extends these findings in several ways. Fischer et al.’s (2019) prior work provided evidence for how Twitter affords teachers personalization based upon the varied engagement patterns users evidenced and the varied content or purposes of tweets (e.g., sharing resources, seeking information, and organization professional learning). This study provides evidence of personalization in a different way: by examining the connections among the words in tweets about content with other aspects of teachers’ contexts. Furthermore, this work extends the survey- and interview-based research of Carpenter and Krutka (2015), by revealing some of the specifics ways teachers differentiate and personalize their experience. These contributions are important as they suggest coherence in teachers connecting their instructional contexts with their prior knowledge and beliefs when engaging on Twitter. They are also important because they go beyond the self-report approaches that have predominant research on educational social media use (Greenhow et al., 2020; Niu, 2019).

Second, teachers positively view receiving new information—especially related to content knowledge. This is evidenced by conversations that included subject matter knowledge related topics and resources that were associated with a more positive sentiment. This underscores prior research that indicated that teachers self-reported use of Twitter to gather information and expand their knowledge (Carpenter & Krutka, 2014, 2015). Notably, some prior research has demonstrated how the sentiment of tweets differs based upon relevant factors—especially factors that have to do with the structure of individuals’ use of Twitter, such as whether teachers are participating in a chat (e.g., Greenhalgh et al., 2020; Rosenberg et al., 2021). Further, other research has examined how interactions with content differed between different subreddits on the social media platform Reddit (Staudt Willet & Carpenter, 2021). What is different about the present study is how we used a research approach that revealed the epistemic connections between posts about content and words that had a positive valence, showing that maybe not features of the social media platform, but, rather, the meaning of teachers’ posts that can evidence the way teachers discuss the content of their work positively. More specifically, this finding indicates the potential of Twitter offering opportunities for participation that have content focus.

Third, participation and conversational patterns differ by teachers’ activity level on the platform. More active teachers in these communities had more conversations about how their individual contexts related with the laboratory practices and the assessments, alongside particular requests for information specific to their individual contexts. In contrast, less active teachers had more conversations in which they shared resources and organized the Twitter community compared to more active teachers. These teachers may fulfill more expert level roles by providing input for other teachers in this community. These different degrees and ways of participating suggest that Twitter is being used as a community of practice in which knowledge can be generated, shared, and organized among practitioners in the same domain (Wenger-Trayner & Wenger-Trayner, 2015; Wesely, 2013). Nevertheless, it must be pointed out that some features of communities of practice (e.g., mutual engagement, joint enterprise, shared repertoire; Wenger, 1998) are typically identified due to the unique nature of interactions on the platform and the broader, more diverse context in which these interactions occur, when likening Twitter to the concept of a community of practice. For instance, regarding joint enterprise, the goals of Twitter users within a specific community or hashtag can be diverse and not necessarily jointly negotiated in the same manner as in traditional communities of practice.

Limitations, Implications, and Recommendations for Future Research

This study has some limitations that are important to consider when interpreting the findings. A limitation to external validity of this study is its context (AP Biology). AP courses are usually taught by very experienced and knowledgeable teachers. Therefore, the teachers in our study may not be representative of all teachers in the United States (or the world). Also, two of the three hashtags examined in this study relate to a time-intensive 2-year PD program, which may have shaped some of the conversations exhibited in our data. Similarly, the disciplinary focus on Biology may lead to conversational patterns not representative in other teacher communities on Twitter, for instance, on more comprehensive educational hashtags such as #edchat. This study also only examines users who actively post content (i.e., tweets) in the communities. However, in online spaces, lurkers also play an important role in the community in providing an audience and potentially amplifying content, for instance, through favoring tweets (Bozkurt et al., 2020).

Methodologically, it is important to note that we have used data from the period July 2011 to June 2016. The constantly change of algorithms—often aiming to improve users’ engagement and aligning more personalized content with broader discovery of trending topics and viral content—can impact how users navigate the platform and consume information. Therefore, it cannot necessarily be assumed that today’s Twitter experiences correspond exactly to those of the period 2011–2016. Moreover, this study solely used the content of tweets for the qualitative coding. Included media files such as pictures and videos were not considered, which—if included— potentially change tweet code assignments for some tweets. In addition, the statistical analysis did not account for potentially important moderators and mediators of teachers’ interactions such as years of teaching experience, self-efficacy, motivation, attitudes towards OPD, or school support. These factors could be explored in future research. Also, the classification threshold distinguishing between active and non-active teachers is mostly based on the data distribution so that substantially different threshold values may impact corresponding results. Initial robustness checks indicated stability for similar thresholds, though. In addition, our conceptualization of thematic conversations, based on a user introducing a new topic, idea, question, or experience that received interactions from other users within a 48-h window, may not fully capture the dynamic and ephemeral nature of Twitter interactions in larger, more active hashtags. This approach, while effective for analyzing smaller datasets, may not be directly applicable or scalable to larger communities or more frequently discussed topics, potentially overlooking the complexity of social media discourse. Last, the most important limitation related to the ENA is the inability to statistically compare the strength of each connection between groups, as only the networks of the different groups can be statistically tested for differences. Still, it is still viable to qualitatively compare the strength of individual connections between codes by visually inspecting the thickness of the connecting lines (representing stronger connections), as we have done in this study.

Beside these limitations, this study provides insights in how teachers used Twitter as a platform for informal learning solely using tweet data. Our findings suggest future research should further explore the nuanced engagement patterns of teachers in online communities, focusing on the development of methodologies to enhance the quality and impact of professional learning in digital spaces. Practically, educators and PD facilitators can leverage these insights to design targeted interventions that foster more effective and contextually relevant online communities of practice for teachers. For policymakers, these results underscore the need to recognize and support social media platforms like Twitter as valuable venues for teacher PD, potentially guiding the allocation of resources and the development of guidelines to maximize their educational impact.

Future research may extend the data sources, potentially adding another qualitative component to interview teachers about their experiences in participating in social media communities, for instance, by applying an experience sampling approach (Aguilar et al., 2021; Larson & Csikszentmihalyi, 2014) to triangulate research findings. Also, future work is encouraged to examine the impact of teacher participation on Twitter on teachers’ knowledge, changes to instructional practice, and student performance more directly, ideally in the form of a randomized controlled field trial. For instance, studies in which traditional face-to-face PD activities are evaluated—and in which also outcome data on the classroom and student-level is collected—may encourage a random subsample of teachers to also engage in professional learning on Twitter. Similarly, we want to encourage replication studies examining teachers’ conversations and collaborative practices in social media settings, using current data, and not limited to Twitter, but also on other prominent platforms such as Reddit, Facebook, or Pinterest (Aguilar et al., 2021; Greenhow et al., 2020). In particular, research that identifies factors contributing to thriving communities of practices of teachers on social media would help advance and scale this inexpensive and accessible form of informal professional learning to many teachers.