Shadows and Light: Unveiling Multifaceted Polarization in Social Media Discourse on Human Trafficking

Online polarization in social media

The rise of social media platforms has fundamentally reshaped the landscape of public discourse, empowering users to generate, disseminate, and interpret content at an unprecedented scale (McNally & Bastos, 2025; Xing & Zhang, 2025). These platforms facilitate community formation by connecting individuals with shared interests or beliefs, thereby reinforcing social cohesion and enabling the expression of group identities (Guess et al., 2023b; Khawar & Boukes, 2024). However, the same affordances that foster collective engagement may simultaneously exacerbate polarization (Bail et al., 2018). Algorithmically curated content, selective exposure, and identity-driven interactions contribute to fragmented information environments and diminished exposure to dissenting views (Powell, 2024; Robertson et al., 2023).

Polarization in digital spaces is broadly characterized by the intensification of attitudinal, ideological, or interpretive divisions between groups, manifesting in both network structures and discursive content (Lee et al., 2014). Scholars have examined this phenomenon from diverse disciplinary perspectives, analyzing factors such as algorithmic bias (Guess et al., 2023a), cognitive heuristics (Buder et al., 2023), and the role of feedback mechanisms—such as likes, shares, and social endorsements—in reinforcing entrenched viewpoints. For example, Messing and Westwood (2012) demonstrated that prominent social endorsements increase users’ likelihood of selecting particular content while simultaneously reducing partisan-selective exposure, thus complicating assumptions about strictly ideological behavior. Complementing this, Nyhan et al. (2023) found that reducing exposure to like-minded sources can enhance openness to diverse perspectives and decrease impolite discourse, even if it does not significantly alter core attitudes. These findings suggest that exposure patterns and social signals interact in complex ways to shape online polarization.

Current studies further reveal that social media-induced polarization is a multifaceted phenomenon characterized by the exacerbation of ideological divisions and the reinforcement of pre-existing beliefs among digital users (Lee et al., 2014). Recent research has expanded the analytical lens to include affective polarization and ideological extremity (Risius et al., 2024), political divisiveness (Masullo, 2023), and even depolarization interventions (Combs et al., 2023), reflecting a growing concern with both the causes and mitigation of polarization in digital environments.

In synthesizing this broad literature, two influential lines of inquiry have emerged. The first focuses on group-level structural dynamics, examining how individuals cluster with like-minded others and form ideologically or emotionally cohesive communities—commonly termed group polarization (Wuestenenk et al., 2025). This research stream investigates the formation of homophilic clusters, the roles of influential users in shaping group norms, and the feedback mechanisms that reinforce alignment within communities over time (Haq & Kwok, 2024; Marchetti et al., 2024).

The second line of research examines content-level divergence, analyzing how communities engaged with the same issue construct increasingly distinct narratives (Overgaard, 2024). This phenomenon, variably termed opinion polarization, topic polarization, or semantic polarization, highlights the fragmentation of interpretive frames, emotional tones, and normative reasoning across groups (Masullo, 2023; Renstrom et al., 2023). Computational methods such as topic modeling, sentiment analysis, and moral framing analysis are widely employed to detect these divergences.

Although these two strands provide complementary insights, they are often examined in isolation. Recent studies increasingly emphasize their interdependence, recognizing that the structural formation of ideologically coherent communities and the semantic divergence of discourse frequently unfold simultaneously (Cinelli et al., 2024; Yang et al., 2021). This integrated perspective conceptualizes polarization as a multifaceted process, encompassing both the social dynamics of group formation and the fragmentation of public discourse. Such a framework advances a more nuanced understanding of how polarization emerges and stabilizes within complex online ecosystems, with significant implications for public opinion, collective decision-making, and the governance of digital communication.

Social identity and self-categorization

Online discourse on human trafficking illustrates how social identities and categorizations shape ideological divides and opinion formation in digital environments. Social identity theory and self-categorization theory provide foundational frameworks for understanding these dynamics (Moran & Prochaska, 2023), clarifying how group affiliations influence individual perceptions, behaviors, and the emergence of polarization.

Social identity theory, introduced by Tajfel and Turner (1979), explains how individuals construct their self-concept based on their affiliations with social groups. The theory posits that people categorize themselves into distinct social groups—based on factors such as race, ethnicity, religion, or shared interests—and integrate these memberships into their overall identity (Li, 2022; Velasquez & Montgomery, 2020). This process establishes psychological boundaries between “in-groups” and “out-groups,” shaping how individuals perceive themselves and others.

Recent studies have illuminated the application of social identity theory in social media research, demonstrating its explanatory power across diverse phenomena. For example, it has been employed to analyze behavioral intentions, social sorting, and the role of identity in influencing user engagement and platform participation (Lane et al., 2023; Qian & Seifried, 2023). Furthermore, research on intergroup dynamics in digital contexts has shown how social identity processes enhance group cohesion while simultaneously fueling intergroup conflict (Burgers et al., 2023). In addition, Ashforth and Mael (1989) emphasized that social identity theory explains how individuals develop a strong sense of belonging to their in-group, often accompanied by preferential treatment of in-group members and antagonism toward out-groups—processes that intensify polarization (Dutot, 2020).

Self-categorization theory, introduced by Hogg and Turner (1987), extends social identity theory by detailing how individuals categorize themselves across varying levels of abstraction—from specific personal identities to broader social identities. This theory emphasizes the situational prominence of particular identities and explains how such prominence shapes cognition and behavior within specific contexts (Koschate et al., 2021). In social media environments, users often conform to the norms and values of their online communities, as deviation may provoke ideological conflict and social sanctions.

Empirical studies applying self-categorization theory have examined phenomena such as information sharing, impression formation, evaluation of social identities, and adherence to social norms in influencer marketing contexts (Qian & Seifried, 2023). These findings demonstrate how salient self-categorizations foster alignment with group norms and strengthen collective identity, thereby influencing how individuals interpret and disseminate information online (Koschate et al., 2021).

Group polarization describes the tendency for group discussions to shift members’ attitudes further toward the group’s prevailing viewpoint. Social identity theory explains this phenomenon as individuals’ motivation to maintain and enhance a positive image of their in-group, reinforcing adherence to group norms and amplifying distinctions between in-group and out-group attitudes (Fritsche et al., 2018; Wakefield & Wakefield, 2023). In contrast, opinion polarization refers to the increasing divergence of perspectives between groups. Self-categorization theory clarifies this process by showing how individuals adopt the dominant narratives of their in-group while dismissing or rejecting out-group views (Flanagin et al., 2014).

In online discussions about human trafficking, identity-driven groups interpret the issue through distinct thematic frames that align with their collective narratives. For example, law enforcement communities predominantly frame trafficking as a criminal justice challenge, advocacy groups emphasize its human rights dimensions, and political factions situate it within broader ideological debates (Tsai & Bagozzi, 2014). Individuals selectively expose themselves to information that resonates with their in-group’s framing, thereby reinforcing both group cohesion and ideological divisions (Chen et al., 2021). Consequently, group polarization (the intensification of attitudes within groups) and opinion polarization (the widening of differences between groups) emerge as interrelated phenomena shaped by the interplay of social identity and self-categorization processes.

Social identity and self-categorization theories account for how in-group affiliation and identity salience shape online polarization. In-group favoritism and norm conformity strengthen internal consensus (group polarization), while intergroup differentiation exacerbates ideological divides (opinion polarization). These mechanisms demonstrate how identity-driven dynamics structure digital discourse, offering insights into managing polarization and fostering more integrative, cross-group communication in online environments.

Data collection and processing

Rebranded as “X” in 2023, Twitter functions as an online social media and networking platform that facilitates user connectivity, communication, and interaction. With features such as likes, retweets, and replies, X promotes user engagement, fostering seamless content interaction and amplification. Consequently, it has become a crucial source for news updates, coverage of significant events, and discussions on trending topics, establishing itself as a pivotal element in contemporary information consumption. Notably, a substantial number of users on X engage with the issue of human trafficking, with thousands of daily tweets dedicated to this topic.

A web crawler was employed to extract data from X using the keyword “human trafficking” via the X API. The X API provides a programmatic gateway for obtaining essential components related to human trafficking, including tweets, retweets, timestamps, text content, and user profiles, while ensuring user anonymity. Data was collected from 1 January 2025 to 7 January 2025.

During this period, we collected user data, which included user IDs, tweets, retweets, comments, interaction dates, and profile information such as occupation, city, place of residence, links to social media profiles, and account creation date. To maintain data integrity, we excluded comments from users with incomplete profile information. In addition, we conducted an additional level of scrutiny on user profiles to identify and eliminate bot accounts, thereby enhancing the overall accuracy of our research findings. It is important to emphasize that the examination of user profiles was conducted solely for bot detection purposes and not for any third-party use.

In the data processing phase, we employed Excel functions to remove invalid characters, including hyperlinks, null values, emoticons, and punctuation marks (e.g. $, &, [], and similar symbols). We also conducted a manual review to ensure spelling accuracy. Following this cleaning process, we retained a total of 8138 tweets for further analysis.

Modeling multifaceted polarization

The multifaceted polarization framework (Figure 1) applies social identity theory and self-categorization theory to uncover how group dynamics and ideological divides shape online discourse on human trafficking. This framework aims to clarifies: (1) how multifaceted polarization is driven by social identity theory and self-categorization processes, (2) how it can be quantified by combining group polarization and opinion polarization, and (3) how the empirical results align with and contribute to these theoretical perspectives. It evaluates two dimensions:

The mechanism of multifaceted polarization is examined through a two-step analytical process. First, the overlap between group structures and thematic divisions is quantified using similarity metrics. Second, the extent to which each group’s discourse concentrates on specific topics is assessed using the Herfindahl Topic Concentration Index (HTCI). This approach not only demonstrates the applicability of the theoretical frameworks but also enhances our understanding of their explanatory power in the context of online discourse polarization.

OPEN IN VIEWER

Figure 1.

Theoretical framework and analytical process of multifaceted polarization.

Group clustering

This study utilizes NodeXL, an integrated SNA tool in Microsoft Excel, to cluster users into diverse groups based on their interactions in online discussions related to human trafficking. We examine network structures using NodeXL’s Clauset-Newman-Moore algorithm (Clauset et al., 2004), which efficiently identifies communities within extensive networks by initially treating each vertex as an individual cluster and then merging them pairwise. Following the clustering process, the Harel-Koren Fast Multiscale layout algorithm is employed to generate a visual representation of the graph. In addition, the Fruchterman-Reingold algorithm, a force-directed layout method, is used to arrange the communities.

In the clustering graph, vertices represent X users, while edges capture user interactions such as retweets, replies, and mentions, which reflect engagement between users (Colleoni et al., 2014). The network’s size, connectivity, and cohesiveness are assessed through graph density (GD) and average path length (APL). In addition, the prestige, prominence, importance, and authority of users are evaluated through degree centrality, betweenness centrality, closeness centrality, eigenvector centrality, and PageRank.

Topic identification

Furthermore, we employ BERTopic, an advanced topic modeling technique, to uncover online public opinion on human trafficking across X. The BERTopic model utilizes BERT (Bidirectional Encoder Representations from Transformers) embeddings and class-based TF-IDF to generate concise and meaningful topic clusters (Xing et al., 2024). Traditional topic modeling approaches, such as Latent Dirichlet Allocation (LDA) and Non-Negative Matrix Factorization (NMF), face challenges in analyzing short and noisy texts (Menezes et al., 2024). These methods rely on word co-occurrence patterns and bag-of-words representations, which often fail to capture contextual nuances, leading to less coherent and more overlapping topics in datasets with sparse textual content (Celikten & Onan, 2025). In our previous experiments, BERTopic consistently demonstrated superior performance compared to LDA and NMF, producing more interpretable and well-separated topics. Therefore, BERTopic was adopted for topic clustering in this study. The workflow is illustrated as follows:

Topic representation

The introduction of a modified Term Frequency-Inverse Document Frequency (TF-IDF) approach enables topic identification based on word distribution within document clusters, assessing word significance within specific topics (Ao et al., 2023). Building upon the core principles of TF-IDF, documents within a cluster are treated as a cohesive unit through concatenation, leading to the creation of a class-based TF-IDF (c-TF-IDF) for topic description.

W x , c = ‖ t f x , c ‖ * l o g ( 1 + A f x )

(1)

t f x , c indicates the frequency of word x in class c ; f x represents the frequency of word x across all classes; and A represents the average number of words per class.

Similarity/dissimilarity

To uncover the mechanisms underlying multifaceted polarization in human trafficking discourse—by integrating both group polarization and opinion polarization—we calculate textual similarity between user-generated content from clustered groups and the topics identified by the BERTopic model. We represent the textual content tweeted by individuals in each clustered group as TF-IDF vectors, where each term’s weight reflects its significance within the respective discourse. These vectors are then compared with keywords in thematic categories, since the topics are represented through structured keyword rankings generated by the c-TF-IDF algorithm, which assigns greater weight to words that differentiate specific topics.

Subsequently, cosine similarity is calculated to quantify the degree of alignment between each group’s discourse and the identified topics. As a widely recognized metric, cosine similarity is widely used in information retrieval, natural language processing, and social sciences (Agrawal et al., 2021; Kasban & Nassar, 2020). The similarity scores range from 0 (indicating complete dissimilarity) to 1 (indicating identical content), allowing us to systematically evaluate textual similarity and thematic alignment in the discourse on human trafficking.

Multifaceted polarization

The HTCI is an adaptation of the Herfindahl-Hirschman Index (HHI) (Song et al., 2013), applied to quantify how narrowly a group’s discussion focuses on specific topics. For a given group G with similarity scores across n topics, the HTCI is calculated as:

H T C I = ∑ i = 1 n ( s i S ) 2

(2)

where s i represents the similarity score of topic i for the group, and S denotes the total similarity across all topics ( ∑ i = 1 n s i ). The HTCI provides a continuous measure of online polarization, ranging from 0 (uniform topic distribution, indicating maximal diversity and minimal polarization) to 1 (complete focus on a single topic, signifying extreme polarization).

Finally, this study quantifies the Weighted Mean Polarization Index (WMPI) of multifaceted polarization by calculating the size-weighted average of group HTCI scores to overall polarization intensity across the entire dataset:

WMPI = ∑ g = 1 N ( HTCI g * Size g ) ∑ g = 1 N Size g

(3)

Clustering network and group polarization

The clustered network depicted in Figure 2 is a directed graph comprising 3846 vertices and 4150 edges, indicating substantial connectivity. This network contains 356 distinct groups, each representing a unique community structure. The interaction dynamics within this network are multifaceted, with 2398 retweets, 315 original tweets, 435 replies, and 99 mentions, reflecting a broad spectrum of communicative activities. In addition, there are 395 self-loops, indicating some level of self-referential behavior. The reciprocated vertex pair ratio is relatively low at 0.0084, indicating limited mutual engagement, while the reciprocated edge ratio is 0.0166. The network consists of 507 connected components, including 125 single-vertex components, revealing the presence of isolated users or sub-networks.

OPEN IN VIEWER

Figure 2.

Clustering network of group polarization on human trafficking.

The APL of 4.9467 indicates a moderately connected network structure, suggesting that information can propagate through the network with reasonable efficiency despite its size. The GD of 0.0004 reveals a sparsely connected network. In addition, the high modularity score of 0.8478 signifies a well-defined community structure, where nodes within the same cluster exhibit significantly higher interconnectivity compared to nodes in different clusters.

Table 1 summarizes structural metrics for the top 10 clusters based on modularity-driven community detection. G1, the largest cluster (223 users), shows low GD (0.004) but a short APL (1.98), suggesting a centralized network with efficient information flow. This structure reflects the presence of key influencers who drive widespread dissemination.

OPEN IN VIEWER

Table 1.

Statistics in the Top 10 Clusters.

Group	Vertices	Total edges	Self-loops	Graph density	Average path length
G1	223	223	1	0.004	1.982
G2	162	275	18	0.008	2.776
G3	142	142	1	0.007	2.053
G4	104	205	5	0.019	1.947
G5	94	94	1	0.011	2.040
G6	91	98	1	0.011	2.160
G7	85	90	5	0.008	2.021
G8	84	137	15	0.017	4.582
G9	56	65	1	0.020	3.224
G10	51	54	1	0.021	2.957

In contrast, G2 demonstrates a high number of self-loops (n = 18) and the longest APL (2.78), indicating insular discourse with limited external connectivity—features consistent with the dynamics of echo chambers.

Clusters such as G4 and G8 display higher GD (0.019 and 0.017, respectively), yet diverge in APL. G4 (APL = 1.95) reflects tightly interconnected users, possibly enabling rapid coordination, whereas G8 (APL = 4.58) may contain dispersed or weakly linked subgroups which may impede efficient exchange.

Other clusters (e.g. G9, G10) exhibit dense but compact structures, marked by repetitive internal interactions and limited structural reach. These variations highlight differences in group cohesion, communication efficiency, and potential influence across the broader network.

Table 2 presents the top 15 users ranked by betweenness centrality—a key indicator of brokerage and gatekeeping in information flow. The vertex labels indicating user IDs have been normalized to anonymize individual accounts while preserving structural integrity. User1 consistently ranks highest across multiple centrality metrics, including in-degree (226), closeness (0.071), eigenvector centrality (0.718), and PageRank (0.012), positioning this account as a dominant opinion leader. Its exceptionally high betweenness score (278,454) suggests a unique role in linking otherwise disconnected communities, thereby facilitating the spread of polarized content across network clusters.

OPEN IN VIEWER

Table 2.

Top Influential Users Based on Betweenness Centrality.

User2 and User3 also occupy structurally influential positions. Although User2 has a moderate in-degree (87), its high betweenness (189,787) and closeness (0.061) indicate extensive reach across diverse subgroups. User3, despite having no inbound ties, ranks high in closeness (0.065) and eigenvector centrality (0.047), implying influence through proximity to prominent users rather than through direct popularity.

Accounts such as User12, User14, and User15 appear structurally peripheral, each with only one or two connections. However, their disproportionately high betweenness scores reveal their function as strategic bridges between otherwise unconnected regions of the network, amplifying their structural importance despite minimal direct interaction.

From an influence diffusion perspective, users like User6, User9, and User13 exhibit high in-degree and PageRank values, reflecting their roles as attention magnets and local information hubs. For example, User9 (in-degree = 114; PageRank = 0.006) may exert concentrated influence within a specific ideological cluster, even if its influence across the broader network is comparatively constrained.

Topic identification and opinion polarization

The BERTopic algorithm identifies 29 topics within our dataset. Figure 3 illustrates the distribution of these topics related to human trafficking on X. Using dimensionality reduction techniques, the spatial distribution chart visualizes topics and documents in a two-dimensional space (see Figure 3). The distance between nodes reflects their semantic similarity, while the colors of the points indicate their subject categories. Outlier topics within primary topics are infrequent, as most documents closely align with the central topic cluster. This indicates that discussions related to human trafficking predominantly focus on the core semantic domain of the primary topic cluster, demonstrating a high level of semantic concentration.

OPEN IN VIEWER

Figure 3.

Distribution of integrated topics on human trafficking.

To ensure reliable theme identification, two experts in deep learning and sociology conducted a systematic, computer-assisted content analysis of topic labels and top-ranked keywords, resulting in the inductive categorization of seven coherent themes: “vulnerable groups,” “persecution experience,” “incident area,” “law and politics,” “public consciousness,” “contraband,” and “case events.”

Table 3 presents the top five keywords ranked by c-TF-IDF scores in each topic. The specific c-TF-IDF values of each keyword are provided in Appendix 1. The theme “vulnerable groups” comprises topics 0, 9, 17, and 18, with representative keywords including “children, women, parents, girls, kids,” and so on. The theme “persecution experience” encompasses topics 4, 6, 13, and 27, featuring prominent keywords such as “pedophilia, rape, harvesting, slavery, endworkness,” and so on. The theme “incident area” is composed of topics 1, 7, 8, and 14, primarily containing location-related keywords like “Israel, Ukraine, Rwanda, Africa, American,” and so on. The theme “law and politics” includes topics 5, 22, 23, and 26, focusing on terms such as “crimes, republicans, federal, illegals, court,” and so on. The theme “public consciousness” involves topics 3, 10, 11, and 19, containing keywords such as “community, awareness, WalkForFreedom, preventing, disgraceful, extremism,” and so on. The theme “contraband” consists of topics 2, 16, 21, and 25, emphasizing “border, drug, earninday, combat, smuggling, migrant,” and so on. Finally, the theme “case events” is composed of topics 12, 15, 20, 24, and 28, represented by keywords such as “Jadi, POTUS, Biden, Joe,” and so on. The identified topics achieve a topic coherence score of 0.236 and a topic diversity score of 0.864, indicating robust model performance and the emergence of semantically distinct and well-structured topics.

OPEN IN VIEWER

Table 3.

Keywords and Distributed Topics in the Themes.

Themes	Topics	Top keywords (ranking by c-TF-IDF score)	Proportion
Vulnerable groups	0	Children, child, women, parents, families	6.98%
	9	Women, girls, rights, people, humantrafficking	4.49%
	17	Ohio, teacher, kids, victims, vulnerable	3.11%
	18	Tate, andrew, brother, slaves, survivors	2.98%
Persecution experience	4	Porn, pedophilia, rape, harvesting, sextrafficking	4.95%
	6	Slavery, endworkness, salve, remorse, satanism	4.88%
	13	Podcast, suffering, tour, torture, trauma	3.30%
	27	Interview, trafficking, abuse, pedophilia, organ	0.84%
Incident area	1	Israel, israeli, Hamas, boundaries, America	6.37%
	7	Ukraine, bio, labs, Kenya, Rwanda	4.62%
	8	Organ, harvesting, China, Arizona, Ukraine	4.50%
	14	Rwanda, refugees, African, headmen, NYC	3.29%
Law and politics	5	Charges, crime, law, commissioner, unvetted	4.90%
	22	Power, executive, ballot, police, operating	1.96%
	23	Republications, voted, Gaetz, federal, illegals	1.89%
	26	Challenges, supreme, court, trade, empire	0.96%
Public consciousness	3	Claytonconway65, elonmusk, Raph, accountability, contract	5.54%
	10	Community, awareness, join, WalkForFreedom, preventing	3.45%
	11	GregAbbott, TX, Texas, disgraceful, truth	3.40%
	19	Fentanyl, Jordan, Denomi, extremism, strength	2.97%
Contraband	2	Border, cartels, drug, videos, indoctrination	5.85%
	16	EarnInDay, EarnIn, combat, drug, operation	3.13%
	21	Smuggling, migrant, trafficking, money, fraud	2.52%
	25	Arrested, biolabs, death, slaves, import	1.24%
Case events	12	Jadi, NG, korban, Tristan, MikeGil	3.40%
	15	POTUS, Trump, Donald, Travis, Diana	3.22%
	20	Biden, cartels, border, boundaries, Hillary	2.67%
	24	Peru, 43, syndicate, YouTube, released	1.75%
	28	Biden, Joe, oversees, Hollywood, accused	0.84%

Similarity analysis and multifaceted polarization

Figure 4 presents the outcomes of the similarity analysis between clustered groups and identified topics. The findings reveal that individuals within the same cluster tend to exhibit stronger alignment with their most semantically relevant topics. In addition, the results indicate that different groups focus intensely on specific thematic areas while paying comparatively less attention to others. For instance, users in clusters G228 to G232 exhibit a pronounced focus on Topic 1, characterized by keywords such as “America, Israel, boundaries,” which aligns with the “incident area” theme. Similarly, individuals in clusters G157 to G161 demonstrate particular attention to Topic 0 and Topic 9, both of which relate to “vulnerable groups,” as evidenced by keywords such as “children, women, people,” and “parents.”

OPEN IN VIEWER

Figure 4.

Similarities between clustered groups and identified topics.

The identified topics are further synthesized to correspond with key themes related to human trafficking within each clustered group. For example, Groups 8, 16, 93, 124, and 269 predominantly address issues concerning vulnerable populations, whereas Groups 4, 21, 53, 174, and 287 focus on experiences of persecution. Groups 44, 68, 249, 251, and 301 emphasize the geographical distribution of trafficking incidents, while Groups 6, 28, 39, 273, and 345 engage in discussions centered on the legal and political dimensions of human trafficking. Conversely, Groups 2, 83, 125, 318, and 321 prioritize public awareness and consciousness, whereas Groups 7, 46, 75, 164, and 193 focus on contraband and illicit trade. Finally, Groups 24, 33, 157, 183, and 253 primarily discuss specific human trafficking cases.

Table 4 highlights a strong inverse relationship between group size and thematic coherence, as measured by the HTCI. The WMPI of 0.75 indicates a moderate yet notable level of polarization across the dataset. Larger groups tend to exhibit lower topic concentration and greater internal variability, as reflected in their higher standard deviations. A robust negative correlation is observed between group size and topic concentration (Pearson’s r = −.72, p < .001), suggesting that as group size increases, thematic alignment decreases. Specifically, for every 10-user increase in group size, HTCI declines by 0.05 (SE = 0.01), reinforcing the idea that larger groups engage in more diverse discursive content.

OPEN IN VIEWER

Table 4.

Group-Level HTCI Scores and Polarization Levels.

Group	Corresponding themes (highest similarity)	HTCI	SD	Polarization level	Interpretation
G1	Vulnerable groups	0.65	0.18	Low	Large, fragmented discussion
G2	Public awareness	0.68	0.15	Low	Broad, diverse engagement
G3	Law and politics	0.72	0.12	Moderate	Emerging thematic focus
. . .	. . .	. . .	. . .	. . .	. . .
G100	Vulnerable groups	0.78	0.09	Moderate-High	Cohesive but not extreme
. . .	. . .	. . .	. . .	. . .	. . .
G178	Persecution experiences	0.83	0.05	High	Tight-knit echo chamber
. . .	. . .	. . .	. . .	. . .	. . .
G356	Incident areas	0.86	0.02	Extreme High	Maximal topic alignment

Theoretical implications

This study makes a significant contribution to the field of online polarization research by integrating perspectives from social identity and self-categorization in examining public perceptions of human trafficking on social media platforms. We propose a conceptual framework that encompasses a multifaceted analysis of polarization.

The modeling of clustering and group polarization provides valuable insights into group dynamics, enabling the identification of subgroups with shared interests and potentially revealing echo chambers within social networks. Network metrics further underscore the role of influential nodes in shaping information diffusion, exerting structural control, and maintaining network cohesion.

Our study advances the literature through the application of BERTopic, an advanced topic modeling method, for identifying opinion polarization patterns related to human trafficking. Although previous studies have utilized BERTopic to detect signatures of problem gambling in online communication data (Smith et al., 2023) and for thematic characterization (Jeon et al., 2023), its potential for examining opinion polarization on social media remains largely unexplored.

Finally, this study contributes to a deeper understanding of multifaceted polarization in online discourse on human trafficking by analyzing both group and opinion polarization. Using TF-IDF vectors and cosine similarity, we assess the alignment between user-generated content within clustered groups and identified topics, revealing patterns of thematic coherence and overlap. The HTCI captures topic concentration within each group, while the WMPI assesses overall polarization on the issue by accounting for both group size and thematic alignment. Collectively, these methods offer a rigorous analytical framework for examining the dynamics of online polarization and its broader implications for public discourse.

Practical implications

This research advances scholarly understanding in the fields of international affairs and crisis management by addressing the complex and pervasive issue of human trafficking, which entails profound societal consequences. Understanding and combating human trafficking is crucial for protecting human rights, strengthening law enforcement and victim support systems, raising public awareness, informing policy development, and addressing the root causes of exploitation.

Our approach to group clustering facilitates the identification of influential individuals, groups, or organizations within the network, thereby amplifying anti-trafficking messages and supporting resource mobilization. Furthermore, the examination of group polarization enables an assessment of the direction and strength of information flow within the network, which is strategically valuable for designing and implementing targeted interventions.

Our BERTopic-based approach to topic identification in the context of human trafficking represents a substantial methodological advancement in automated topic classification. This method enables the detection of emerging trends and salient issues associated with this pressing social problem. By assessing the prevalence of specific topics or terms, researchers can evaluate public awareness and the depth of understanding across multiple dimensions of the issue. Consequently, government agencies and advocacy organizations can allocate resources more effectively by prioritizing areas associated with heightened public engagement.

Finally, the HTCI and WMPI provide actionable metrics to assess the concentration and polarization of discourse, allowing stakeholders to design targeted interventions that foster balanced dialogue and mitigate ideological fragmentation. In addition, these tools can be applied to other contentious social issues, offering a transferable framework for understanding and addressing the dynamics of online polarization in public discourse.

Limitations and future research

This study has several limitations that should be acknowledged. First, the data were drawn solely from the social media platform X, potentially limiting the generalizability of the clustering and topic modeling results. Different platforms, such as Facebook, YouTube, and Instagram, exhibit distinct patterns of information dissemination, particularly in terms of user interactions and the formation of emergent groups. To enhance the validity of our findings, future research should incorporate additional platforms to assess online polarization in a broader context. Second, the BERTopic model used for topic classification has demonstrated efficiency exclusively on our English-language dataset. To evaluate its general applicability, future studies should test the model on non-English texts. Third, while HDBSCAN allows the BERTopic model to approximate topic distribution within a document using the probability matrix, this method only partially resolves the issue and does not adequately account for documents that contain multiple topics. To address this limitation, future work should aim to develop an adapted version of HDBSCAN that can more effectively manage topic modeling in multi-topic scenarios. Finally, although our interpretation draws on social identity theory to explain patterns of group polarization and discourse alignment, it is important to note that identity processes are inferred from observed group structures and communication patterns rather than directly measured. Future research should incorporate direct measures of identity salience and group affiliation to validate and extend these findings.