A Social Network Analysis on the Danmaku of English-Learning Programs

Abstract

This study utilizes the danmaku on the Bilibili platform as the research subject to examine how their characteristics vary according to the nature or focus of English teaching videos. By employing social network analysis, the study reveals distinctive features in danmaku. For videos categorized under linguistic knowledge (phonetics, vocabulary, and grammar), the danmaku comments predominantly center around topics such as phonetics, vocabulary, and grammar. Conversely, in videos categorized under language skills (listening, speaking, reading and writing), the danmaku comments primarily reflect a vocabulary review for three of the four skills, with only the listening skill showing slight deviations. This underscores the centrality of vocabulary in skill-oriented videos. The findings highlight the unique role of danmaku in distinguishing between knowledge and skills within the context of English teaching videos.

1. Introduction

Numerous studies have explored online interactive models in education, ranging from MOOCs with pre-recorded videos to live synchronous classes. Among these, danmaku, a unique form that combines the advantages of pre-recorded content and real-time interaction, represents a third category. Danmaku, when applied to online learning, enables students to actively engage with course content through peer participation in danmaku interactions. This engagement is influenced by the intensity and design of the curriculum, creating a learning atmosphere that fosters immersion, particularly in professional contexts. Such phenomena underscore the contributions of danmaku to online learning environments.

Implemented on Bilibili, one of China’s major video-sharing platforms, danmaku features a live commenting system that allows users to engage with peers asynchronously. This interaction fosters a sense of community and enhances user engagement in learning (e.g., [1] for English; [2] for Mathematics). The sense of synchronicity generated by danmaku not only enriches the content over time but also gives learners the impression of studying collaboratively, thereby boosting motivation and enthusiasm for learning.

Compared to MOOCs, danmaku offers a unique advantage in enhancing participant engagement. As highlighted by [3], a study involving 4466 participants across 10 highly rated MOOCs emphasized the importance of peer interaction in fostering engagement. Unlike structured peer interactions in MOOCs, danmaku is entirely learner-generated, making it an exemplary form of authentic peer interaction in online learning. Jiang et al. (2022) [1] compared the learning experiences provided by MOOCs and Bilibili, concluding that Bilibili offers a superior environment for fostering engagement due to its interactive features. Specifically, danmaku demonstrated significantly higher effectiveness in stimulating learning interest compared to MOOC platforms. While there was no notable difference in grammar acquisition, danmaku was more effective in enhancing vocabulary acquisition, linguistic intuition, and conversational fluency. Furthermore, Zhang et al. (2023) [4] highlighted the use of a series of L2 vlogs on Bilibili for Spanish learning, emphasizing the critical role of interaction-oriented learning.

In a related study, Yang (2020) [5] investigated the influence of danmaku videos on learners’ social interaction and their role in increasing motivation and engagement. The interactive nature of danmaku strengthens the sense of connection among learners, positively impacting participation, comprehension, and learning outcomes. The cumulative nature of danmaku enables the presentation of diverse perspectives, as learners from different backgrounds share insights, enriching the viewing experience and encouraging critical thinking.

Zeng et al. (2024) [6] integrated danmaku into educational data analysis using the TextMind software for psycholinguistic analysis (https://www.researchgate.net/publication/285653495_Developing_Simplified_Chinese_Psychological_Linguistic_Analysis_Dictionary_for_Microblog, accessed on 7 February 2025). Their study examined 58,143 danmaku comments in an online course on the fundamentals of digital electronics. The results demonstrated how danmaku fosters engagement, offering personalized recommendations to students and practical guidance for improving participation in online education platforms. However, contrary findings were reported by [7], who analyzed the use of danmaku in TED-Ed science videos. They found that merely increasing the volume of comments failed to facilitate deep learning. Similarly, Li et al. (2022) [8] observed that danmaku did not meet learners’ expectations when the interaction between learners and teachers remained one-sided, with no feedback from the instructors. In Table 1, we show a summary of prior studies.

Table 1. The summary of prior studies on danmaku.

Authors and Year	Domain and Platform	Investigation	Main Technology
Yang (2020) [5]	General, Bilibili	The influence of danmaku videos on learners’ social interaction and their role in increasing motivation and engagement	The two-pronged model combining a semiotic resource perspective and a social practice perspective
Peng & Wang (2021) [7]	Science, TED-Ed science videos	The impact of danmaku in learning: the correlation between the number of danmaku comments which subjects leave and the test of comprehension	Spearman’s rank correlation analysis
Li et al. (2022) [8]	General, Video lectures	Identifying the interaction mode that danmaku meets	Semi-structured interviews
Jiang et al. (2022) [1]	English, MOOCs and Bilibili	Comparing the learning experiences provided by MOOCs and Bilibili	Interviews and statistics
Zeng et al. (2024) [6]	General, Bilibili	To understand students’ learning patterns and present corresponding intervention strategies for different types of students	TextMind

Previous studies on danmaku have primarily focused on aspects such as the timing of messages, textual content, and emotional expressions, with an emphasis on data mining and text analysis. Most research has concentrated on danmaku in different types of open courses, often using individual videos as the primary research subjects, with limited exploration of systematic learning collections. Moreover, studies specifically targeting systematic English teaching through danmaku remain scarce. To address this gap, the present study focuses on a systematic collection of English learning materials, using learners’ danmaku data as the primary research subject.

In this study, we propose a social network analysis (SNA) approach to visualize the danmaku for better understanding and revealing the interaction. This study diverges from [6] by employing SNA to highlight the relevance of danmaku to various topics in English language instruction. Using Python-based web scraping to collect danmaku data, we aim to examine whether the interactions are related to specific subcategories of English instruction, such as phonetics, vocabulary, and grammar in professional contexts, as well as listening, speaking, reading, and writing in skills-based contexts. This approach seeks to determine whether danmaku is aligned with the two overarching concepts of English teaching—linguistic knowledge and language skills. Moreover, this research intends to explore whether danmaku facilitates a better learning environment through cumulative interactions or, conversely, whether an overload of information leads to distractions.

We hypothesize that while the danmaku mechanism can enhance the interactive experience, excessive engagement might hinder effective cognitive processing, impeding the deep learning emphasized in educational frameworks. In support of this, Li et al. (2022) [8] noted that 40.9% of the interactions in English videos were related to supplemental knowledge and answering queries, highlighting the potential for meaningful engagement. This study seeks to uncover whether similar patterns emerge in our analysis and whether danmaku can indeed provide a conducive environment for deep learning.

The other parts of this paper are composed as follows. In the Related Work Section, we present the use of SNA alongside related technologies. The Methods Section details our research methodology. In Section 4, we showcase the results of our study, followed by a discussion and conclusion in Section 5. The Section 5 summarizes key findings and provides insights for future research directions.

2. Related Work

In this section, we present the related work of SNA and the technologies in our research.

2.1. Social Network Analysis

Social network analysis (SNA) has become an essential area of research, particularly in computer science and social sciences. SNA is defined as the study of social structures through the use of networks and graph theory. Serving as a powerful framework for understanding complex interactions within various fields, SNA examines how relationships between individuals influence the behaviors and outcomes within a network. SNA has evolved from sociological roots to a multidisciplinary approach that integrates insights from computer science, economics, and organizational studies [9]. Some methodologies in SNA are proposed, such as “Centrality Measures” [10] and “Community Detection” [10,11]. SNA can be utilized in various domains, including co-authorship networks, social media analytics and epidemiology [12]. Recent studies indicate that SNA is experiencing rapid growth, particularly with the advent of big data analytics and machine learning techniques. However, challenges persist, such as data privacy concerns and the need for more robust analytical frameworks to assess user influence effectively [13]. The integration of machine learning with SNA is seen as a promising direction for future research [13].

The evaluation metrics of a social network graph are centrality, degree, betweenness, closeness, eigenvector centrality, diameter/radius, average geodesic distance, average degree, reciprocity, density, and global clustering coefficient [14]. In our research, we utilize eigenvector centrality for our evaluation.

Eigenvector centrality [15] evaluates a node’s importance based on the importance of its neighbors, in contrast to degree centrality, which only considers the number of direct connections. As a result, eigenvector centrality provides a more comprehensive assessment of node significance in a network [16], incorporating the influence of well-connected nodes with high centrality [17].

2.2. Pre-Trained Language Model

In recent years, pre-trained language models have become a pivotal technology in the field of natural language processing (NLP) [18]. RoBERTa (robustly optimized BERT pre-training approach) builds upon BERT (bidirectional encoder representations from transformers) [19] with several significant improvements, including the use of larger training datasets, extended training duration, larger batch sizes, and longer input sequences. Additionally, it removes the next sentence prediction (NSP) task and adopts a dynamic masking strategy for the masked language model (MLM) task [20].

The Chinese-RoBERTa-WWM-Ext-Large is a pre-training model of Chinese BERT for its advanced understanding of Chinese language tasks. This improves its ability to capture contextual meaning, making it highly effective for tasks like text classification, sentiment analysis, and question answering in Chinese [20].

2.3. Clustering Algorithms

K-means, first introduced by [21] is a partitional clustering method developed for classifying and analyzing multivariate observational data. The algorithm partitions the data into k clusters by minimizing the average squared distance between points within the same cluster. Its main advantages are simplicity and speed [22]. K-means is a partitional clustering technique within cluster analysis, an unsupervised exploratory method that is generally classified into two categories: hierarchical and partitional clustering. Hierarchical clustering constructs a tree-like structure by iteratively merging or splitting clusters, ultimately forming a complete hierarchical structure. In contrast, partitional clustering methods, such as K-means, divide the data into a predefined number of clusters, with each data point assigned to exactly one cluster, without any hierarchical relationships [23].

The K-means algorithm is one of the most common, unsupervised methods for its simplicity, efficiency, and scalability in clustering tasks. The K-means algorithm performs well when the number of clusters is predefined, and the dataset is structured, making it ideal for segmenting data into distinct groups.

2.4. TF-IDF

TF-IDF is a classic method for measuring term importance, combining term frequency (TF), which reflects the significance of a term within a document, and inverse document frequency (IDF), which gauges its distribution across the entire corpus. Rare terms are assigned higher weights due to their greater discriminative value [24]. Initially proposed by [25] in the field of information retrieval, this concept highlights the importance of both term frequency and specificity for effective retrieval. TF-IDF, combining TF and IDF, has since become a fundamental approach in information retrieval [26]. In Introduction to Modern Information Retrieval [24], cosine similarity is used to compute the similarity between a query and a document by measuring the cosine angle between their respective vectors. This study employs cosine similarity to calculate the similarity between danmaku vectors.

2.5. The Levenshtein Distance-Based Method

The Levenshtein distance-based method uses Levenshtein’s algorithm [27], which measures the minimum number of edit operations (insertion, deletion, substitution) required to transform one string into another, to calculate text similarity. In this study, an undirected, unweighted edge is created between two danmaku nodes if the Levenshtein distance between their texts is 1. Originally developed for error correction in binary data [27], the Levenshtein algorithm has been widely applied in fields such as computational linguistics [28] and bioinformatics [29].

3. Methods

In this section, we present our research methods, including data collection, and social network analysis procedure.

3.1. Data Collection

This study adopted a systematic computational methodology to analyze danmaku data and construct a social network representing user interactions. The workflow, shown in Figure 1, comprises five key stages: web scraping, preprocessing, embedding, clustering, and network construction.

Figure 1. SNA workflow.

3.1.1. Web Scraping

Initially, web scraping was employed to extract data from the Bilibili platform, focusing on relevant videos and their associated danmaku comments. This process ensured the comprehensive collection of the raw textual data necessary for subsequent analytical tasks.

This study was conducted on Bilibili, a comprehensive video-sharing platform, established in 2009. To identify relevant content, searches were performed using keywords such as “English pronunciation”, “English vocabulary”, “English grammar”, “English listening”, “English speaking”, “English reading”, and “English writing”. The primary selection criterion was the quantity of danmaku comments (real-time comments displayed on videos). Secondary factors, including playback counts, coin donations, likes, and shares, were also considered. From the seven identified categories, the three most popular English learning collections were selected for analysis. Each collection comprises a varying number of videos.

Eventually, data for this study were collected between October and November 2023, encompassing 21 English learning collections on Bilibili. These collections comprised a total of 2057 individual videos and generated 1,721,873 danmaku comments. The danmaku data included both the text content (danmaku comments) and the sender’s unique user ID (UID), representing interactions from 331,263 participants.

3.1.2. Preprocessing

We develop a preprocessing pipeline to standardize and prepare the textual data for analysis, involving three primary steps: decoding, text normalization, and deduplication. Initially, HTML-encoded entities in the danmaku text (e.g., <, &) were decoded using the html.unescape function to restore the original user input. Subsequently, text normalization was performed, which included converting Chinese text to simplified characters, transforming English text to lowercase, and converting full-width characters to their half-width equivalents. Lastly, to address the issue of excessive repetition in online content, sequences of more than three consecutive identical characters were truncated to a maximum of three. This step preserved the semantic integrity of expressions such as “好好學習” (study diligently) and internet slang such as “666” (indicating admiration).

3.1.3. Embedding

Following preprocessing, sentence embeddings were generated using the Chinese-Roberta-WWM-Ext-Large model, a pre-trained transformer-based model optimized for capturing nuanced semantic relationships in Chinese text, serving as a foundation for subsequent social network analysis.

3.1.4. Clustering

Sentence embeddings served as the input for the K-means clustering algorithm, which categorized the danmaku comments into distinct clusters. The number of clusters k in this study is determined using an empirical rule. As shown in Equation (1), k is the floor of the square root of n, with n representing the data size.

$$k=⌊\sqrt{n}⌋$$

(1)

3.1.5. Network Construction

The final step before analysis involved constructing a social network model for each collection based on two fundamental concepts: “User behavior” and “textual similarity between danmaku comments”. “User behavior” is related to a danmaku submission, characterized by the content and frequency of danmaku comments, while textual features are analyzed using the K-means clustering algorithm to explore relational patterns in messages.

The network consists of three types of nodes—users, danmaku, and clusters. To facilitate visualization in Gephi [30], node-related information is stored in CSV files, which include the node content and shape (polygon). The node content includes user IDs, danmaku comments, and the cluster assignment of each comment. To distinguish danmaku comments that consist of a single number, cluster nodes are labeled as “# + cluster number”. For instance, #0 and #1 represent the first and second clusters identified by the K-means algorithm, respectively. The “polygon” attribute is used to differentiate node types in Gephi: user nodes are represented as circles (polygon = 1), danmaku nodes as squares (polygon = 4), and cluster nodes as pentagons (polygon = 5).

The network contains two types of undirected edges: user–danmaku edges (U-D) and danmaku–cluster edges (D-C). U-D represents the relationship between users and the danmaku comments they submit, forming a many-to-many relationship where one user can submit multiple danmaku comments, and a single danmaku comment can be submitted by multiple users. The weight of the edge reflects the number of times a user submits the same content, with a minimum weight of 1. For example, if user “U1” submits danmaku comment “D1” three times, the edge weight between U1 and D1 would be 3. This could occur if the user is watching a series and submits the same comment, “D1”, at different times. D-C represents the relationship between danmaku comments and the clusters to which they are assigned, as determined by the K-means algorithm described in the previous section. These edges are unweighted and form a many-to-one relationship, where each danmaku comment belongs to a single cluster, but each cluster may contain multiple danmaku comments. For instance, if danmaku comments “D1” and “D2” are both assigned to cluster 1, each will be connected to node “C1” by an unweighted, undirected edge.

To ensure the reproducibility of the results during the execution of the Python process, a random seed of 42 was set.

Based on these concepts, this study constructs a social network model as Figure 2 with three types of nodes and two types of edges to investigate interaction characteristics between users in different clusters. It analyzes the behavioral patterns of user interactions through danmaku and identifies the key topics that drive user participation in danmaku interactions.

Figure 2. Gephi visualization of social networks: a case study. The detail of danmaku nodes is presented in Appendix A. Different colors in the figure represent different communities.

3.2. Social Network Analysis Procedure

In this phase, the study focuses on analyzing danmaku nodes within the top three subgraphs of each collection, leveraging insights gained from community detection methods.

3.2.1. Community Detection

After constructing the network, community detection was performed. Community detection aims to uncover naturally occurring groups or clusters within a network without prior knowledge of the number or size of these groups [17]. A common approach involves maximizing the modularity score, which evaluates the quality of a particular division of the network into communities [17]. This study follows a similar approach. According to research by [31], the Leiden algorithm outperforms the heuristic Louvain algorithm [32] in terms of both speed and the quality of community connectivity. Therefore, this study applies the Leiden algorithm to analyze 21 social networks, aiming to obtain the modularity and community structure of an integrated network comprising three types of nodes.

Modularity measures the extent of assortative mixing, where nodes with similar attributes tend to form connections within the same community. A higher modularity value, approaching 1, indicates a strong presence of intra-community connections relative to a randomized network, thereby reflecting significant structural properties of the network [17].

The complexity of social networks means that factors such as the diversity and frequency of user-submitted comments can significantly impact the interpretation of interactions. High-centrality danmaku comments, for example, can create distinct communities with the user nodes from which they originate, while separating from other nodes within the same cluster. This separation enhances the understanding of danmaku interactions. Thus, the greater the number of communities identified by a community detection algorithm, the more diverse the underlying topics, reflecting a wider range of user behaviors and textual features. In contrast, a smaller number of communities indicate a more concentrated set of topics.

3.2.2. Subgraph Construction

For the analysis of community nodes, the top three communities from each network are selected based on node count, yielding a total of 63 communities. The next step is to construct a bipartite network containing only user and danmaku nodes. The user and danmaku nodes from the original communities are first identified, along with the edges connecting them, while excluding cluster nodes and their associated edges. In this bipartite network, edges are retained between user nodes and danmaku nodes that belong to the same community, with edge weights representing the frequency of users posting the corresponding danmaku comments. Given the relatively small scale of danmaku within communities, the analysis focuses on character-level relationships.

Spelling and grammatical errors, commonly associated with internet language [33], are prevalent in danmaku comments as a form of online communication. For instance, spelling variations such as “禮貌” (transliteration: “Li Mou”, translation: “manners”) and “禮帽” (transliteration: “Li Mou”, the spelling error case of “manners”) frequently occur. To address this, both cosine similarity and Levenshtein distance are employed to establish direct connections between danmaku nodes, supplementing potential omissions in clustering results generated by pre-trained models.

The cosine similarity method leverages TF-IDF (term frequency-inverse document frequency) weighting to represent the danmaku comments, converting text into vectors and calculating cosine similarity between the danmaku comments. If the similarity exceeds 0.5, an undirected, unweighted edge is created between the corresponding nodes.

Additionally, for certain collections where danmaku comments are predominantly in English, a Levenshtein distance-based method is applied. In this approach, an edge is created between nodes if the Levenshtein distance between their content equals 1. Consequently, edges between danmaku nodes are constructed when either of the two conditions is satisfied: cosine similarity greater than 0.5 or Levenshtein distance equal to 1.

To identify the top ten danmaku nodes in each community network, weighted eigenvector centrality is utilized.

For each community, a visualization ready for Gephi was created by selecting the top 10 danmaku nodes based on weighted eigenvector centrality. The node data was filtered and sorted by centrality values to identify these key nodes. A graph was then constructed to include all nodes directly connected to these top nodes by a single edge, forming a target node set. Both edge and node files were filtered to retain only the data relevant to the target nodes, with edge weights preserved from the original community to represent interaction strength. The processed data was exported for visualization in Gephi, providing an intuitive representation of interaction patterns.

To enhance the readability of the figures, we utilized labels to represent the contents of danmaku nodes and provided a detailed description of each label along with its translation in the tables of the appendix. In addition, we draw user nodes in green and danmaku nodes in red below.

4. Results

We categorized English learning programs into two main groups: linguistic knowledge and language skills. The linguistic knowledge category includes phonetics, vocabulary, and grammar, while the language skills category, following established classifications, encompasses listening, speaking, reading, and writing. Each of these seven categories was analyzed using the top three videos from Bilibili, resulting in a total of 21 videos that are denoted as P1~P3 (phonetics), V1~V3 (vocabulary), G1~G3 (grammar), L1~L3 (listening), S1~S3 (speaking), R1~R3 (reading) and W1~W3 (writing). For each video, we constructed edges based on interactions between users and the danmaku comments they submitted. From this data, we identified the three largest communities within each video (e.g., V1_1~V1_3 in V1) and calculated the centrality of nodes within these communities.

Our analysis revealed distinctive patterns in the linguistic knowledge category. For instance, danmaku comments centrality for phonetics videos (e.g., P1) prominently featured content-specific terms like “一個是捲到齒齦後,一個是捲到硬齶” (One curls toward after the alveolar ridge, while the other curls toward the hard palate.), see Figure 3 (All nodes exemplified in the text will be highlighted in bold in the corresponding table.). In vocabulary videos (e.g., V2, V3), central danmaku comments often combined English and Chinese explanations, such as “overlook忽視” and “beneath在下方”, see Figure 4. Similarly, grammar videos (e.g., G2, G3) highlighted a series of syntactic discussions of subjunctive mood (“虛擬語氣”) or the functions of word class, such as “狀語修飾動詞”(adverbials modify verbs), see Figure 5. These findings illustrate a clear focus on the respective content within the linguistic knowledge’s category, which we grouped into a distinct class.

Figure 3. Visualization of social network in P1. It shows that danmaku comments centrality for phonetics videos prominently featured content-specific terms. The detail of danmaku nodes is presented in Appendix B.

Figure 4. Visualization of social network in V2 (upper panel) and V3 (lower panel). In vocabulary videos, central danmaku comments often combine English and Chinese explanations. The detail of danmaku nodes is presented in Appendix C and Appendix D.

Figure 5. Visualization of social network in G2 (upper panel) and G3 (lower panel). Grammar videos highlighted a series of syntactic discussions of “subjunctive mood” or the functions of word class. The detail of danmaku nodes is presented in Appendix E and Appendix F.

For the language skills category, the analysis reveals nuanced differences in the content and nature of danmaku comments. In L1, although the lack of context makes it difficult to infer precise content, comments from the third community, such that “…沒聽出來…” (…couldn’t hear…) clearly indicates their failure in listening comprehension, see Figure 6. This aligns with previous findings that learners seek to share similar experiences of misunderstanding. In S1 and S3, the danmaku comments reflect the integration of newly learned vocabulary during the learning process. Examples include “quantum, 量子” and “mosquito n.蚊子” with consistent annotation of the part of speech. Similar trends are observed in R3 (e.g., “radical極端的”) and W3 (e.g., “cultivate, foster培養”), where vocabulary translation is emphasized, see Appendix H, Appendix I, Appendix J and Appendix K. However, these comments show little indication of specific skill-focused training, as the primary focus appears to be on vocabulary explanation rather than the underlying skill itself.

Figure 6. Visualization of social network in L1. It shows that kearners seek to share similar experiences of misunderstanding. The detail of danmaku nodes is presented in Appendix G.

Notably, in L2 and across S1 and S3, longer sentences frequently appear in the danmaku comments, suggesting transcription of phrases or sentences introduced during instruction, see Figure 7. This is distinct from the broader trends in the skills category. Among the four skills, listening and speaking stand out as having unique danmaku patterns compared to the others, highlighting their distinctiveness in the learning process.

Figure 7. Visualization of social network in L2. Longer sentences frequently appear in the danmaku comments. The detail of danmaku nodes is presented in Appendix L.

5. Discussion and Conclusions

In language learning environments, viewers often use danmaku to discuss linguistic features of the target language and to address comprehension challenges. For instance, Zhang and Cassany (2019) [34] identified three primary categories of danmaku related to Spanish: (1) content focused on Spanish (61%), (2) learning Spanish as a foreign language, and (3) Spanish–Chinese translation. Building on this, our study further explores the role of danmaku in English learning, categorizing its content into two main areas: knowledge and skills.

In the linguistic knowledge category, danmaku comments reflect various aspects of English learning, including phonetics, vocabulary, and grammar. In the language skills category, however, a unique pattern emerges: only listening displays distinct danmaku comments, primarily involving learners discussing their own mishearings. In contrast, the other three skills—speaking, reading, and writing—primarily feature vocabulary-focused comments, with danmaku used as a tool for recording and memorizing words. This suggests that vocabulary serves as the foundation for language learning; without a solid vocabulary base, developing other skills is as precarious as building a castle in the air. However, the findings also reveal that danmaku, as employed by English learners, predominantly serves a single purpose—vocabulary-focused learning—with little variation in its application to other aspects of language learning.

According to [35], a study of Chinese undergraduate students in the United States during 2011 and 2012 revealed three key skill profiles: (1) speaking was consistently weaker than the other three skills (S < L, R, W); (2) speaking and writing were weaker than listening and reading (SW < LR). These findings suggest that reading is relatively easier for Chinese English learners compared to the other skills, while speaking poses the greatest challenge. Building on these insights, it is worth investigating whether the weaker speaking proficiency among danmaku users influences their behavior—specifically, their tendency to transcribe longer sentences in danmaku as a means of reviewing and reinforcing content. While this hypothesis aligns with the observed usage patterns, further research is needed to confirm whether the reliance on transcription is indeed linked to difficulties in mastering speaking.

Jiang et al. (2022) [1] highlighted that Bilibili, due to its interactive danmaku features, provides a more engaging learning environment compared to MOOCs. Building on this, the present study employs an SNA approach to visualize danmaku interactions, offering visual evidence of its effectiveness in facilitating vocabulary acquisition. The findings demonstrate that interactive learning through danmaku fosters a more vibrant and engaging learning atmosphere.

Furthermore, Zeng et al. (2024) [6] discussed how danmaku enhances student participation. Our study complements their work by providing a visualized understanding, thereby operationalizing Zeng et al. (2024) [6] ’s assertion that danmaku can improve engagement on online education platforms.

In [36], the researchers employed a variety of methods—including social network analysis, surveys, longitudinal designs, and data visualization—to explore the impact of peer interactions on second language (L2) acquisition during study-abroad experiences. Their findings reveal a strong correlation between the diversity of students’ social networks, the frequency of interactions, and improvements in language proficiency, particularly during the initial phase of their study-abroad period. Building on this framework, our study represents a novel approach by employing social network analysis to classify different English instructional videos based on danmaku comments. The findings demonstrate that in the linguistic knowledge category, the danmaku comments closely align with the specific content of the videos, such as phonetics, vocabulary, and grammar. In the skills category, however, distinct patterns emerge only for listening and speaking, where the danmaku exhibits unique characteristics compared to the other skills. This highlights the potential of danmaku as a tool for categorizing instructional content and identifying learner engagement patterns across various aspects of English language learning.

For educators, the feedback provided through danmaku encourages them to not only consider content that students can independently learn online but also design classroom activities that promote discussion and interaction via danmaku. This approach enhances students’ ability for autonomous learning while naturally shifting the traditional teacher-centered teaching model to a more student-centered approach, aligning with contemporary trends in education.

For platform designers, optimizing danmaku functionality could involve leveraging network analysis to highlight significant danmaku content and keywords based on collections, videos, or video timelines. For example, social network graphs or leaderboards displayed alongside the video could provide real-time insights into learners’ discussion hotspots at specific video progress points. This would benefit content creators by identifying potential learning feedback and challenging concepts, improving teaching materials. It would also help learners identify key learning points and resolve their doubts efficiently.

Moreover, if platforms retain students’ danmaku as a form of learning output and feedback, these could serve as valuable data for platform designers when developing diverse English language teaching modules. This collaborative effort between platform designers and educators would facilitate a deeper understanding of students’ online learning behaviors and support the continuous refinement of teaching strategies.

Although we provide a visualization of danmaku, the visualization may not be easy to interpret when the graph is extremely dense. When designing an interactive GUI, some filtering and zoom-in/zoom-out functions are required. In addition, more SNA functions, such as centrality and modularity, could be integrated into the next version of our approach.

As discussed, danmaku in online learning fosters an atmosphere of active participation by engaging students with interactive content. This study provides visual evidence demonstrating that danmaku creates a more engaging learning environment compared to MOOCs. The practical recommendations for educators and platform designers form a key contribution of this research.

However, the integration of an interactive GUI, such as filtering and zoom-in/zoom-out functions, could enhance the immersive quality of SNA visualizations. This would make the visual representations more dynamic and user-friendly. Consequently, the recommendations for educators and platform designers in both teaching and practice would become more actionable and impactful.