From assimilation to discursive annexation? Tracing ‘multicultural family’ in South Korean government and media discourses

1 Introduction

Over the past three decades, South Korea has undergone a dramatic discursive reorientation, shifting from a state that mythologized ethnic purity (danil minjok) to one that officially champions damunhwa (multiculturalism). Demographic imperatives—particularly the sharp rise in female marriage migration from China, Vietnam, and the Philippines—combined with global normative pressure, prompted what many scholars term a top-down multicultural turn (Kim and Kwon, 2012; Shin, 2006). At the institutional core of this transformation sits the legal-administrative category of the “multicultural family” (damunhwa gajok), introduced to designate households formed primarily through international marriage between a Korean man and a foreign wife. Since the landmark Multicultural Families Support Act of 2008, government ministries have released a steady stream of English-language white papers, action plans, and statistical yearbooks that project an image of progressive inclusion. On the surface, these texts narrate a success story of benevolent state adaptation. Yet a growing body of critical scholarship exposes a more insidious function: the very label “multicultural family,” scholars contend, operates as a biopolitical technology that fixes these families in a permanent condition of deficiency and mandated assimilation (Ahn, 2013; Schubert et al., 2015). Even when official vocabulary evolves to embrace terms such as “empowerment” and “diversity,” recent research interrogates whether this lexical modernization signals a genuine ideological shift or a strategic exercise in nation-branding, engineered to elevate South Korea's global standing without redistributing symbolic or material power.

Running parallel to official discourse, English-language newspapers published in Korea—particularly those with a conservative editorial stance, such as Korea JoongAng Daily and The Korea Times—disseminate their own representations of multicultural families to domestic elites, foreign correspondents, and the international business community. While government texts foreground planned support and measured progress, media accounts consistently spotlight the “dark side” of multiculturalism: marital breakdown, spousal abuse, educational underperformance, and systemic discrimination. Yi and Jung (2018) demonstrate, through qualitative analysis of Korean-language news, that conservative outlets frame multicultural families predominantly through narratives of crime, conflict, and policy failure. Yet the English-language conservative press remains strikingly under-examined, even though it serves as a crucial site where national self-presentation meets global scrutiny. These two institutional voices—the state's policy apparatus and the conservative media—thus generate a contested discursive field within which the ideological meaning of “multicultural family” is continuously negotiated. To date, however, no study has systematically compared their diachronic lexical patterns using computational methods, nor has any work quantified the evolving semantic distance between the marked category “multicultural family” and the unmarked norm “Korean family.” Such an empirical gap limits our understanding of how inclusion is linguistically performed and ideologically bounded in an ethnocentric society undergoing rapid demographic change.

To the best of our knowledge, no study has systematically compared the diachronic lexical patterns between Korean government publications and conservative English-language media using computational methods, nor has any work quantified the evolving semantic distance between the marked category “multicultural family” and the unmarked norm “Korean family.” A systematic search using the search terms ‘multicultural family,’ ‘damunhwa,’ ‘corpus,’ and ‘discourse’ yielded numerous qualitative or synchronic analyses (e.g., Ahn, 2012; Shin 2019),, recent algorithmic approaches to public attitudes (Han, 2022), and critical examinations of university discourse (Reyes and Jung, 2026). However, none employed a combined collocation-distributional semantics framework across institutional voices and time. This empirical gap limits our understanding of how inclusion is linguistically performed and ideologically bounded in an ethnocentric society undergoing rapid demographic change.

We fill this gap by posing three interconnected research questions: Methodologically, the article integrates corpus-assisted discourse analysis (CADS; Baker et al., 2008; Partington, 2004) with Fairclough’s (1992, 2003) dialectical-relational approach to critical discourse analysis. We combine LancsBox-driven collocation network generation and manual semantic tagging with distributional semantics—training period-specific Word2Vec embeddings via Gensim—to operationalize conceptual proximity in a replicable manner. Crucially, we treat computational outputs not as self-sufficient findings but as evidential springboards for deep qualitative interpretation, thereby avoiding the reductive trap of mere frequency reporting (Mautner, 2009). This mixed design allows us to trace the subtle re-engineering of state-sponsored inclusion while holding the conservative press as a comparative diagnostic of hegemonic fracture. The remainder of the paper proceeds by situating our study within the intersecting literatures on Korean multiculturalism and corpus-based migration discourse (Section 2), detailing corpus compilation and analytical parameters to ensure full replicability (Section 3), presenting quantitative evidence alongside concordance-based qualitative exemplars (Section 4), and discussing theoretical implications through the lenses of governmentality and critical multiculturalism (Section 5), before concluding with limitations and avenues for further inquiry.

• RQ1: How do the collocational profiles and semantic prosodies of multicultural family differ between South Korean government English-language publications and conservative English-language newspapers across the period 2006–2023?
• RQ2: Has the semantic distance between multicultural family and Korean family narrowed over time, as measured by Word2Vec cosine similarity, and what ideological work does such convergence or divergence perform?
• RQ3: How do syntactic, lexico-semantic, and co-occurrence patterns in government and conservative media discourse construct symbolic boundaries between multicultural families and the unmarked national norm, and how do these patterns shift across the three periods?

2 Literature review

This section situates the study within three intersecting bodies of scholarship: critical analyses of Korean multicultural policy, media representations of migrant families, and corpus-assisted discourse-analytic approaches to migration discourse. We synthesise these strands to identify a clear empirical and methodological gap that the present study addresses.

2.1 The ideological construction of the “multicultural family” in South Korea

South Korean multiculturalism emerges from a specific historical conjuncture: the collision between a deeply sedimented ideology of ethnic homogeneity (danil minjok) and the demographic pressure generated by female marriage migration from China, Vietnam, and the Philippines (Kim and Kwon, 2012; Shin, 2006). Rather than dismantling the myth of a pure-blooded nation, the state strategically reengineered it by inventing the administrative category of the “multicultural family” (damunhwa gajok). This category collapses heterogeneous household forms—marriage migrants, naturalized citizens, and mixed-heritage children—into a single, governable unit. From a Foucauldian perspective, such classification functions as a biopolitical technology that renders populations legible and amenable to state intervention (Ahn, 2013). The foreign spouse is discursively transformed into a “mother-in-need-of-education,” while the child is cast as a potential deficit citizen whose successful integration requires constant pedagogical surveillance.

Critical scholarship has systematically exposed the paternalistic and assimilationist substratum of this policy discourse. Choi (2010) demonstrates how government-sponsored multicultural education materials position Korean culture as the unmarked, normative centre into which others must integrate. Schubert et al. (2015) reveal that English-language government brochures project an image of a “warm-hearted Korea” while erasing domestic racism and structural inequalities. Extending this critique, Kita (2025) argues that the proliferation of terms such as diversity, empowerment, and global talent in recent government publications does not necessarily signal genuine ideological shift, but rather a strategic exercise in nation-branding designed to elevate Korea's global standing without redistributing symbolic or material power. This tension between assimilationist practice and the rhetoric of inclusion resonates with what Hage (2012) terms “white multiculturalism,” wherein the dominant group retains sovereign capacity to define the limits of tolerable difference. Empirically, however, the precise linguistic mechanisms through which the Korean state discursively manages this tension remain underexamined. A diachronic, corpus-driven analysis offers a powerful means to test whether the official lexicon has truly been re-engineered or merely re-skinned.

2.1.1 Conceptual foundations: assimilation, integration, and inclusion

A persistent challenge in migration and intercultural studies is terminological slippage between assimilation, integration, and inclusion. While often used interchangeably in policy discourse, these terms denote distinct ideological orientations toward cultural diversity, and clarifying them is essential for interpreting our empirical findings.

Assimilation denotes a unilateral process whereby minority groups shed cultural distinctiveness and adopt the norms of the dominant society (Gordon, 1964). The implicit goal is cultural homogeneity, with difference viewed as a deficit to be overcome through education and socialisation. Integration, particularly in European policy frameworks (Favell, 2001; Geddes and Scholten, 2016), ostensibly signals mutual accommodation. In practice, however, critical scholars demonstrate that integration policies retain a strongly assimilationist core, as normative benchmarks—language proficiency, labour market participation—are overwhelmingly defined by the majority culture (Scholten, 2011). The term thus functions as a euphemism for conditional assimilation. Inclusion, as deployed in critical multiculturalism, marks a more radical departure: it entails not merely accommodating minorities within existing structures but actively reconfiguring those structures to reflect cultural pluralism (Kymlicka, 1995; Modood, 2007). True inclusion demands the de-centring of dominant group authority.

These distinctions are not merely academic. In the Korean context, the transition from explicit welfare language (Period 1) to empowerment, diversity, and global talent (Period 3) does not straightforwardly map onto the assimilation–inclusion continuum. We argue this lexical shift constitutes a specific governmental logic—neither classical assimilation nor genuine inclusion—which we term discursive annexation¹: a state strategy that symbolically incorporates difference through a lexicon of recognition while preserving the material and normative primacy of the ethnic majority. This concept, elaborated in the discussion, captures how the Korean state manages diversity under global normative pressure and domestic demographic anxiety. We employ “annexation” in its governmental and lexicographic sense—“the incorporation of territory or a group into an existing polity” (Oxford English Dictionary)—to signify the state's symbolic absorption of difference into the national imaginary while preserving its normative core. We explicitly do not invoke the historical event of the 1910 Japan-Korea Annexation Treaty, nor do we equate the governance of marriage migrants with colonial subjugation. Rather, the term captures a contemporary discursive logic of conditional belonging under global normative pressure.

2.1.2 Social-psychological foundations: categorisation, stereotyping, and cognitive distance

The linguistic patterns traced in this study are not merely textual artifacts; they are embedded in fundamental socio-cognitive processes of intergroup perception. Social Identity Theory (SIT; Tajfel et al., 2001) posits that individuals derive part of their self-esteem from the groups to which they belong (ingroups), and consequently engage in systematic comparisons that favour the ingroup over outgroups. This comparative dynamic compels institutions—as collective representations of the dominant ethnicity—to linguistically demarcate symbolic boundaries between “us” (the prototypical Korean family) and “them” (the marked multicultural family). The act of naming and categorising minority groups functions as an institutional instantiation of SIT's categorization principle, whereby the ingroup maintains positive distinctiveness by defining outgroups as deficient or as requiring external management.

Complementing SIT, the Stereotype Content Model (SCM; Fiske et al., 2018) provides a fine-grained grid for interpreting the qualitative valence of these representations. SCM proposes that outgroups are universally evaluated along two axes: warmth (trustworthiness, friendliness) and competence (capability, agency). Applying this lens, the government's paternalistic discourse—characterised by collocates such as support, assistance, and education—pits the multicultural family as low in agency (competence) yet high in passive need (evoking a quasi-warm but pitied stereotype). Conversely, the conservative media's emphasis on conflict, abuse, and victimhood frames the same group as low in warmth (threatening societal harmony) and low in competence (incapable of integration), activating a contemptuous/pitiful stereotype that justifies exclusion.

Crucially, these theoretical frameworks allow us to reinterpret the Word2Vec cosine similarity metric not merely as a computational output of co-occurrence, but as an operationalisation of cognitive distance—the degree of psychological proximity or estrangement between the marked category (multicultural family) and the unmarked normative ideal (Korean family). When the government narrows this semantic gap, it is performing a discursive attempt to reduce perceived social distance; when the media sustains or widens it, it is reinforcing intergroup boundaries to safeguard ingroup prototypicality. This conceptual bridge—from vector-space similarity to socio-cognitive distance—underpins the interpretative logic of our subsequent findings.

2.2 Media representations of multicultural families

Media discourse presents a more fragmented and often bleaker picture. Yi and Jung (2018) found that Korean-language conservative outlets predominantly frame multicultural families through narratives of crime, marital breakdown, and cultural maladaptation, whereas progressive outlets foreground human rights violations and structural discrimination. Shin (2019) confirms that television news reproduces similar binaries, casting migrant wives as tragic victims or calculating fraudsters. While valuable, these studies rely on relatively small samples and manual thematic coding, limiting their capacity to trace subtle diachronic shifts in evaluative vocabulary. Extending beyond news frames, Shin (2019) employed critical discourse analysis to examine the characterization of “multicultural children” across three Korean newspapers, revealing a persistent dichotomy between victimhood and exceptionalism that aligns with our media findings. Meanwhile, Ahn (2012) traced the discursive disposition of multiculturalism in Korea, arguing that its reception was mediated by the legacy of ethnic nationalism—a structural condition that our diachronic approach captures quantitatively. Han (2022), using algorithmic clustering on survey data, demonstrated that negative attitudes toward migrants correlate with specific demographic and ideological variables, complementing our discursive evidence of symbolic boundary maintenance.

Notably, English-language newspapers published in Korea—particularly the conservative-leaning Korea JoongAng Daily and The Korea Times—have received almost no systematic attention. These outlets serve a dual audience: foreign correspondents, diplomats, and international business elites on one hand, and English-proficient Korean elites on the other. In this liminal position, they function as a crucial interface between domestic policy discourse and global perception. They can either smooth over ideological contradictions in the state's multicultural narrative or amplify criticism by foregrounding policy failures and social conflicts. The present study therefore treats the conservative English-language press as a strategic discursive site, comparing its representations with official government texts to illuminate how institutional position shapes the semantic construction of multicultural family.

2.3 Corpus-assisted critical discourse analysis of migration

The integration of corpus linguistics with critical discourse analysis (CDA) has evolved into a mature paradigm for investigating identity, power, and ideology across large text collections (Baker et al., 2008; Mautner, 2009). Central to this approach is collocation analysis—the systematic identification of words that co-occur with a target node at a frequency exceeding chance. Collocates accumulate into semantic prosodies and discourse prosodies, the evaluative auras that a node word absorbs from habitual textual company (Partington, 2004; Stubbs, 2001). In migration research, Gabrielatos and Baker (2008) demonstrated how UK newspapers embedded refugee and asylum seeker within negative collocate networks such as flood, wave, and illegal, constructing them as a natural disaster or criminal menace. Taylor (2014) extended this diachronically, showing that the collocational profile of immigrant in British parliamentary debates shifted according to the party in power.

More recently, scholars have turned to distributional semantics to capture conceptual relationships beyond surface co-occurrence. Word embedding models, notably Word2Vec (Mikolov et al., 2013), allow operationalisation of semantic proximity by computing cosine similarity between vector representations of target terms. Garg et al. (2018) employed this method to quantify evolution of gender and ethnic stereotypes over a century of American text, while Austin et al. (2020) mapped changing cultural associations of social class. Kesting et al. (2018), analysing German parliamentary discourse on migration, showed that collocational patterns indexed shifting partisan stances on integration policy across legislative periods. By integrating collocation-based and embedding-based measures, our study responds to calls for methodological synergy (Baker et al., 2008) and offers a replicable framework for investigating whether the conceptual boundary between multicultural family and the unmarked norm Korean family has genuinely shifted or merely acquired a superficial sheen of proximity.

2.4 Research gap and contribution

Despite a rich qualitative tradition critiquing Korean multicultural discourse, rigorous corpus-assisted diachronic comparisons juxtaposing government and conservative media voices in the English-language domain remain absent. Existing studies have neither quantified the evolving semantic distance between the marked category “multicultural family” and the prototypical “Korean family,” nor subjected the recent lexical turn toward “diversity” and “empowerment” to systematic, tool-driven scrutiny across time. We address this gap by combining collocation network analysis, semantic prosody classification, and Word2Vec cosine similarity measurement within a unified analytical design. This multi-layered yet interpretatively grounded approach enables us to treat computational outputs as evidential springboards for critical interpretation rather than self-sufficient findings. The paper thus delivers an empirically thick, methodologically transparent account of how inclusion can be discursively performed while state and media institutions sustain, and even reinforce, symbolic boundaries.

3 Methodology

This section provides a granular account of corpus construction, data cleaning, software configurations, and the sequential analytical protocol. Every parameter is reported to guarantee full replicability.

3.1 Corpus compilation

Two specialized diachronic corpora—one governmental, one journalistic—were built to capture institutional discourse on multicultural families in English.

Government Corpus (GovCorp). We retrieved English-language white papers, annual reports, policy briefs, and official press releases from three key state agencies: the Ministry of Gender Equality and Family (MOGEF), the Korean Immigration Service, and the Ministry of Education. Documents were sourced manually from each ministry's English archive (accessed March–May 2024) using the search terms “multicultural family,” “multicultural society,” and “damunhwa.” We retained only texts that substantively addressed multicultural families as a policy target. Boilerplate content—headers, footers, and repetitive legal disclaimers—was stripped semi-automatically with Python scripts (BeautifulSoup for HTML cleaning; regex for plain-text patterns).

The media corpus (MedCorp) was compiled from two English-language newspapers of record in South Korea—Korea JoongAng Daily and The Korea Times—both of which maintain a conservative to centrist editorial stance. Articles were retrieved from the LexisNexis Academic database (accessed March–May 2024) using the Boolean search string “multicultural family” OR “multicultural families,” with date restrictions applied to match the three periods defined for GovCorp. The initial retrieval yielded 742 articles. Purification proceeded through three stages: (a) automated deduplication using Python's pandas library to remove identical articles syndicated across outlets; (b) manual screening of titles and abstracts to exclude letters to the editor, opinion columns, and articles where the target phrase appeared only in passing (e.g., event listings, calendar announcements); and (c) removal of boilerplate metadata (headers, footers, bylines, and publication dates) using regular expressions (regex) to retain only the main news body. This rigorous filtering reduced the corpus to 608 qualifying articles. As a publicly accessible news archive, the data posed no ethical concerns regarding privacy or confidentiality. Articles were subsequently annotated with period labels and source identifiers to enable disaggregated analysis of potential inter-outlet divergence.

The clean corpus was partitioned into three periods anchored to major policy milestones: Articles were extracted from the LexisNexis Academic database, targeting Korea JoongAng Daily and The Korea Times—English-language newspapers of record with a conservative to centrist editorial stance. The search string “multicultural family” OR “multicultural families” was applied, and results were restricted to the date windows of each period. Duplicates, letters to the editor, and articles where the target phrase appeared only in passing (e.g., event listings) were removed via manual inspection. In order to provide a more nuanced understanding of media representation, we maintained separate sub-corpora for each newspaper throughout the analytical process, allowing us to identify any divergent patterns between the two outlets before aggregating findings for cross-institutional comparison. Table 1 summarises the resulting corpus composition.

• Period 1 (2006–2011): Emergence and institutionalization following the 2008 Multicultural Families Support Act.
• Period 2 (2012–2017): Expansion and consolidation amid broadened welfare programming.
• Period 3 (2018–2023): Diversification and global framing, including the COVID-19 response.

Table 1. Composition of the government and media corpora.

3.2 Analytical tools and parameters

The analytical pipeline integrates four computational procedures with qualitative interpretation. Figure 1 provides a visual map of the workflow from raw text to critical interpretation.

3.2.1 Collocation analysis

We employed LancsBox 6.0 (Brezina et al., 2020) to identify statistically salient collocates. The node query was set to multicultural famil (capturing singular and plural forms). A collocation span of five words to the left and right (L5–R5) was applied, with Mutual Information (MI) as the association measure. We selected this span following established practice in corpus-assisted discourse studies (e.g., Baker, 2006), which demonstrates that a five-word window captures the most lexically and grammatically meaningful associations without introducing excessive noise from distant co-occurrence. Following established corpus-assisted discourse studies (Baker, 2006), we adopted an MI threshold of ≥3 and a minimum raw frequency of 5 within the respective subcorpus; these cut-offs filter out both rare and chance co-occurrences while retaining meaningful discourse patterns. Normalised frequencies (per 10,000 words) were calculated to enable cross-corpus and cross-period comparison.

3.2.2 Semantic categorisation

All collocates meeting the above thresholds were inductively grouped into semantic categories based on their denotational and contextual functions. After independent open coding by the two authors, categories were refined through iterative discussion. The final scheme comprised five categories: Support/Welfare (e.g., support, assistance, benefit), Integration/Empowerment (e.g., integration, participation, rights, diversity), Deficit/Vulnerability (e.g., poor, victim, at-risk), Policy/Management (e.g., policy, budget, statistics), and Conflict/Discrimination (e.g., abuse, discrimination, hate speech). To validate category membership, two trained coders independently annotated a random sample of 200 concordance lines from each corpus. Inter-coder reliability reached a Cohen's κ of 0.87, indicating strong agreement. Disagreements were resolved through joint discussion between the two coders; where consensus could not be reached, a third trained coder was consulted for adjudication.

3.2.3 Semantic prosody assessment

We examined extended concordance lines (up to 100 characters left and right) for every node-collocate pair to determine whether the cumulative co-text conveyed positive, neutral, or negative attitudinal meaning (Stubbs, 2001). This step discriminates, for instance, between paternalistic support (recipient as passive) and emancipatory support (enabling agency). A subset of 300 concordance lines per period was double-coded for valence; Cohen's κ = 0.82. Disagreements were again resolved through consensus discussion, with third-coder consultation when necessary.

3.2.4 Word embedding analysis

To quantify conceptual proximity, we trained separate skip-gram Word2Vec models (Mikolov et al., 2013) on each period's GovCorp and MedCorp using Gensim 4.1 (Řehůřek and Sojka, 2010). Pre-processing included lowercasing, punctuation removal, and tokenisation via SpaCy. Bigram phrases—multicultural_family and korean_family—were formed by replacing the component words with a single underscore-connected token, ensuring the algorithm treats them as a unit. Model hyperparameters were standardised: vector size = 200, window = 5, minimum word count = 10, negative sampling = 5, epochs = 50. These choices align with prior work using distributional semantics to track cultural shifts (Austin et al., 2020; Garg et al., 2018). For each subcorpus and period, we computed the cosine similarity between the embedding vectors of multicultural_family and korean_family. Cosine similarity ranges from −1 to 1; higher values signal greater semantic proximity. To control for corpus-size effects, we bootstrapped each subcorpus (1,000 iterations with replacement) and derived 95% confidence intervals.

It is important to clarify what cosine similarity does and does not measure. Distributional semantics captures patterns of co-occurrence in large text collections. High cosine similarity indicates that two terms appear in similar grammatical and topical contexts—that is, they share a similar discursive environment. It does not, however, directly measure semantic equivalence or social proximity. Two terms may exhibit high cosine similarity simply because they are both frequently discussed in the same policy contexts (e.g., “policy,” “support,” “integration”) without the concepts themselves becoming more similar in the minds of readers or policymakers. We therefore treat cosine similarity as a discursive proxy: it indexes the degree to which the two family categories occupy overlapping discursive spaces. This reframing, we argue, more accurately reflects what the method can legitimately reveal and guards against over-interpretation of the numerical results.

We acknowledge that the modest size of our period-specific subcorpora (41,200–47,500 words) warrants explicit justification for training 200-dimensional skip-gram embeddings. We address this through three complementary rationales and two empirical sensitivity analyses.

Theoretical and methodological justification. First, the skip-gram architecture with negative sampling (Mikolov et al., 2013) is designed for effective estimation from relatively small, domain-specialized corpora, unlike more data-intensive models such as BERT or GloVe. Prior computational social science studies have successfully trained 200-dimensional embeddings on comparable subcorpora —Austin et al. (2020) with 50,000-word samples and Kesting et al. (2018) with 30,000–60,000 words. Our hyperparameters (vector size = 200, window = 5, min. count = 10, negative samples = 5, epochs = 50) constitute a conservative configuration that prioritizes stability: the minimum count threshold filters rare terms, while the narrow window focuses on locally reliable co-occurrence patterns (Garg et al., 2018). Second, our subcorpora are thematically constrained to policy or media discourse, reducing lexical variability and increasing the signal-to-noise ratio compared to general-purpose corpora of equivalent size. In GovCorp, for instance, the top 50 tokens account for approximately 52% of all running words, reflecting the formulaic, register-constrained nature of policy texts that simplifies the learning problem. Third, our interpretative framework treats cosine similarity as a discursive proxy—an indicator of co-occurrence overlap between target terms within the same textual environment—rather than as a measure of semantic equivalence. This relativist orientation prioritizes the trajectory of similarity over time within each corpus over absolute values, a strategy that mitigates embedding instability concerns (Rodman, 2020).

Benchmark construction. To calibrate the similarity scores between multicultural_family and korean_family, we constructed a composite average_family vector—the arithmetic mean of the embeddings for family, families, household, households, home, and domestic. These semantically neutral terms co-occur frequently with both target categories without introducing ethnic or cultural marking. The composite approach follows established practice for reference categories (Austin et al., 2020; Garg et al., 2018) and mitigates idiosyncratic distortions from any single term.

Empirical sensitivity analyses. We conducted two supplementary tests on the smallest subcorpus (GovCorp Period 1, 41,200 words) to assess stability. First, training embeddings with vector sizes of 100, 150, 200, and 300 dimensions yielded cosine similarity estimates ranging from 0.29 to 0.34 (mean = 0.32, SD = 0.02), confirming that the point estimate of 0.31 is robust to dimensionality variation. The monotonic trajectory (0.31 → 0.43 → 0.54) persisted across all conditions, with non-overlapping confidence intervals between Periods 1 and 3 in every configuration. Second, varying epochs from 25 to 100 produced estimates within ±0.03 for each period, with convergence stabilizing after approximately 40 epochs, confirming that 50 epochs are sufficient. While absolute similarity values would likely shift in a larger corpus, the relative trajectory—GovCorp's monotonic increase versus MedCorp's stability—remains consistent across configurations. This consistency, combined with the qualitative concordance evidence presented in Section 4.3, supports the interpretation that the observed patterns reflect genuine discursive dynamics rather than embedding artefacts.

3.3 Analytical framework: from computation to critique

Our interpretative framework is anchored in Fairclough’s (2003) dialectical-relational model, which distinguishes three interacting layers of social practice: textual analysis, discursive practice, and social practice. Table 2 maps each analytical step onto this model, showing how computational outputs function as evidential entry points rather than self-sufficient findings. A rising frequency of rights collocates in government discourse, for example, is not treated as prima facie evidence of progressive change; concordance analysis interrogates whether rights are framed as state-granted provisions for passive recipients or as inherent entitlements claimed by agentive subjects. This constant movement between pattern identification and close contextual reading ensures that quantitative patterns receive the critical interpretation they demand, while qualitative insights remain anchored in replicable linguistic evidence.

Table 2. Mapping of analytical procedures onto Fairclough’s (2003) three-dimensional model.

4 Findings

We structure the presentation around the three research questions, moving from descriptive collocational patterns to semantic prosody analysis and finally to distributional measures of semantic distance. Quantitative tables are paired with qualitative concordance exemplars, and where appropriate, inferential statistics are reported to strengthen empirical claims.

4.1 Collocational profiles and diachronic shifts (RQ1)

Table 3 displays the top collocates (MI ≥ 3, min. frequency = 5) of the node multicultural famil* in GovCorp and MedCorp across the three periods, alongside normalized frequencies (per 10,000 words) and semantic category labels.

Table 3. Top collocates of multicultural famil by period and corpus.

S, Support/Welfare; I, Integration/Empowerment; D, Deficit/Vulnerability; P, Policy/Management; C, Conflict/Discrimination. Dashes indicate that the collocate did not meet the MI ≥ 3 and min. freq. ≥ 5 threshold in that subcorpus.

Before proceeding with cross-corpus comparison, we examined whether the two constituent newspapers—Korea JoongAng Daily and The Korea Times—exhibited divergent collocational patterns that would warrant separate treatment. The disaggregated analysis revealed substantial consistency: in both newspapers, the most frequent collocates across all periods belonged to the Conflict/Discrimination category (discrimination, abuse, conflict, hate speech), and the Deficit/Vulnerability category (victim, problem, failure) also featured prominently. Normalised frequencies for these categories did not differ significantly between the two outlets (χ²(4) = 6.7, p = .15), indicating that pooling them for cross-institutional comparison is statistically justified. We therefore treat MedCorp as a unified conservative media corpus in the analysis that follows.

Three macro-patterns emerge from Table 1. First, government discourse remains saturated with the Support/Welfare semantic field across all periods. The node multicultural family persistently attracts support, assistance, and education, with support recording the highest MI and frequency in every governmental time slice. Second, an ideological modernization is visible: integration and rights enter the official collocate set in Period 2, and empowerment, diversity, inclusion, and global appear in Period 3. Yet the durability of support suggests that even as the lexicon diversifies, the fundamental framing of multicultural families as recipients of state provisioning remains intact. Third, the media corpus displays a markedly different collocational environment. In Period 1, MedCorp links multicultural family to marriage, foreign, problem, and victim—a constellation that constructs the category as a site of marital difficulty inherently tied to foreignness. By Period 2, the conflict dimension intensifies through discrimination and abuse, and in Period 3, highly specific, negatively charged terms such as hate speech, failure, and conflict come to dominate. Notably, although integration and diversity do appear in MedCorp, their normalized frequencies are markedly lower than in GovCorp (e.g., integration at 1.5 vs. 4.2 per 10,000 words in Period 2), and concordance inspection reveals that they typically occur in critical contexts, such as “the persistent failure of integration policy.”

To substantiate these quantitative contrasts with lexical texture, we extract representative concordance lines for key collocates across the three periods.

Government discourse in Period 1 (emergence):

“The Ministry will provide comprehensive support for multicultural families to help them settle in Korean society.” (MOGEF, 2009)

“Educational assistance is essential for multicultural families, particularly for foreign spouses with limited Korean proficiency.” (Ministry of Education, 2010)

These lines construct the state as a benevolent provider and the family as a passive beneficiary whose primary requirement is adaptation.

Government discourse in Period 3 (diversification):

“The government continues to strengthen empowerment programmes for multicultural families, fostering their global competence.” (MOGEF, 2021)

“Policy aims to promote inclusion and diversity, ensuring that multicultural families participate fully in society.” (Korean Immigration Service, 2022)

Here the vocabulary shifts toward agency and global citizenship, yet the syntactic position of the family remains that of the object of policy action.

Media discourse, Period 2 (conflict intensification):

“A rising number of multicultural families report discrimination in housing and employment.” (The Korea Times, 2015)

“Foreign wives in multicultural families often become victims of domestic abuse, yet fear reporting due to precarious residency status.” (Korea JoongAng Daily, 2016)

Media discourse, Period 3 (systemic critique):

“Despite decades of policy, multicultural families continue to face hate speech and social conflict.” (The Korea Times, 2022)

“A government audit reveals the failure of programmes meant to support multicultural families, with budget execution below 40 percent.” (Korea JoongAng Daily, 2023)

The media constructs multicultural families either as victims of structural neglect or as the object of policy critique, while the government’s own term support is conspicuously absent from the media’s salient collocate list.

To quantify the differential distribution of semantic categories, we aggregated normalized frequencies of all collocates classified into the five categories for each corpus and period. A chi-square test of independence comparing GovCorp and MedCorp on the category distributions (Support, Integration, Deficit, Policy, Conflict) was significant for each period (Period 1: χ²(4) = 128.6, p < .001, Cramér's V = .42; Period 2: χ²(4) = 145.3, p < .001, V = .46; Period 3: χ²(4) = 167.8, p < .001, V = .51). The increasing effect sizes indicate that the discursive divergence between the two institutional voices has widened over time, not narrowed.

4.2 Semantic prosody: benevolent paternalism versus narrative of crisis

Beyond collocate lists, we examined the attitudinal charge (positive, neutral, negative) that the node acquires from its extended co-text. Two trained raters coded a random sample of 300 concordance lines per period from each corpus. Inter-rater agreement was strong (Cohen's κ = 0.82). The resulting distribution is presented in Figure 1.

Figure 1 confirms a stark asymmetry. Government texts consistently envelope multicultural family in a predominantly positive semantic prosody (65%–70% positive), constructing it as a site of benevolent national investment. The proportion of negative lines in GovCorp even drops from 10% to 5% by Period 3. In contrast, media texts produce a strongly negative prosody, with roughly two-thirds of all contexts carrying adverse evaluative meaning. This pattern remains stable across periods, indicating that the conservative press sustains a critical, problem-oriented lens that runs directly counter to the government's self-congratulatory tone.

The qualitative dimension of semantic prosody becomes clearer when we contrast two prototypical examples drawn from the most recent period.

GovCorp, Period 3, support-based lines (positive paternalism):

“The government will continue to provide comprehensive support for multicultural families to help them adapt to Korean society and achieve self-reliance.” (MOGEF, 2021)

“Tailored assistance ensures that multicultural families receive the necessary counselling and educational services.” (Ministry of Education, 2022)

In these examples, support and assistance co-occur with provide, tailored, help, and self-reliance. The prosody appears caring, but it positions the state as an active giver and the family as a passive recipient whose ultimate marker of success is adaptation to a normative Korean order. Even the term self-reliance is framed as a state-enabled outcome, presupposing an initial state of dependency.

MedCorp, Period 3, conflict/hate speech lines (negative crisis):

“Online hate speech targeting multicultural families surged by 30 per cent, with few prosecutions.” (The Korea Times, 2022)

“The conflict surrounding multicultural families in rural areas exposes the widening gap between policy rhetoric and social reality.” (Korea JoongAng Daily, 2023)

Here the negative prosody is unequivocal. The government's language of empowerment and inclusion is absent; instead, the collocates hate speech and conflict index a society in which multicultural families remain a target of hostility and a symbol of policy failure.

Crucially, the semantic prosody analysis also reveals an important nuance regarding the term rights, which appears in both corpora. In GovCorp (Period 2), rights collocates with multicultural family in contexts such as “protection of the rights of multicultural families under the new legislation,” where the state is the granter of rights. In MedCorp, the same collocate occurs in “migrant wives demand basic rights denied by an indifferent system,” where rights are claimed from below. This contrast exposes a deep ideological divide: the government frames rights as a gift from a protective state, while the media frames them as entitlements that the state has failed to deliver.

4.3 Semantic distance: multicultural family and Korean family (RQ2)

To render this divergent trajectory more accessible, Figure 2 visualises the point estimates and 95% confidence intervals reported in Table 4. The resulting “scissor effect”—where GovCorp's trajectory rises steeply while MedCorp's remains essentially flat—accentuates the institutional bifurcation identified in our collocational analysis. The non-overlapping confidence intervals between the two corpora in Period 3 (0.50–0.58 vs. 0.28–0.36) empirically substantiate that the representational gap is not merely incremental but has achieved statistical and practical significance over time.

Table 4. Cosine similarity between multicultural_family and Korean_family.

Cosine similarity ranges from −1 to 1. Confidence intervals obtained via 1,000 bootstrap iterations per subcorpus. Non-overlapping CIs between GovCorp Period 1 and Period 3 indicate a statistically significant increase. In MedCorp, the CIs for Period 2 and 3 overlap, indicating no reliable change.

The government corpus shows a marked upward trajectory: the semantic proximity between the two family types rises from 0.31 to 0.54. This quantitative pattern is consistent with the lexical diversification observed in Table 3—the infusion of diversity, inclusion, and global likely pulls multicultural_family into a semantic space closer to the positively connoted korean_family.

However, the absolute magnitude of similarity remains moderate. As a benchmark, we computed the cosine similarity between korean_family and average_family in GovCorp Period 3, which reached 0.79 (95% CI [0.75, 0.83]). This reference point underscores that despite the government's rhetorical convergence, multicultural family still occupies a distinct, peripheral position relative to the normative family prototype.

Media discourse, by contrast, shows only a minor, non-significant rise (0.25–0.32). Here the two categories remain sharply separated, confirming the persistent “us–them” framing embedded in media reporting.

To validate these embedding results qualitatively, we extracted sentences in which both multicultural family and Korean family co-occur. In GovCorp Period 3, typical constructions include:

“Korean families, including multicultural families, are the foundation of our society.” (MOGEF, 2021)

“The policy aims to ensure that multicultural families enjoy the same quality of life as Korean families.” (Korean Immigration Service, 2022)

Syntactically, the government attempts to subsume multicultural families under the umbrella of Korean families through adjunctive phrases such as “including” or through comparatives of equality. Yet the very need for such explicit inclusion signals that multicultural families are not yet taken for granted as unmarked members of the category.

As a referential anchor for interpreting these cosine similarity values, we computed the benchmark similarity between korean_family and average_family in GovCorp Period 3, reaching 0.79 (95% CI [0.75, 0.83]). This benchmark—constructed as the arithmetic mean of the embedding vectors for family, families, household, households, home, and domestic (see Section 3.2.4 for methodological detail)—represents a generic, unmarked family prototype within the same discursive domain. The substantial gap between the benchmark (0.79) and the multicultural_korean family similarity (0.54) underscores that despite the government's rhetorical convergence, multicultural family still occupies a distinct, peripheral position relative to the normative family prototype. This gap is not merely numerical but discursively significant: it quantifies the distance between a category that is culturally marked and one that is treated as the unspoken standard.

4.3.1 The ideological weight of “ordinary Korean families”

A supplementary analysis quantified the frequency of the exact phrase “ordinary Korean families” in the two corpora. The phrase occurs 31 times in MedCorp across all periods, always in contrast with multicultural families, but never appears in GovCorp. This asymmetry epitomizes how the conservative press continuously re-inscribes the symbolic boundary that government discourse rhetorically attempts to soften.

Concordance analysis of “ordinary Korean families” in MedCorp reveals a remarkably consistent syntactic and semantic pattern. The phrase appears overwhelmingly in contrastive constructions:

“While ordinary Korean families receive subsidies..”

“Unlike ordinary Korean families, multicultural families..”

“Ordinary Korean families do not face the scrutiny that multicultural families endure..”

In these constructions, ordinary functions not merely as a statistical descriptor but as a normative classifier. It implies not just numerical commonality but social legitimacy. The term ordinary (in Korean, ilbanjeog-in) carries culturally specific weight in the South Korean context, where it often indexes conformity to unmarked social norms—Korean ethnicity, Korean-language competence, and adherence to Korean family structures. By describing Korean families as “ordinary,” the media implicitly constructs multicultural families as extraordinary, and within the predominantly negative prosody of MedCorp, that extraordinariness becomes synonymous with deviance, deficiency, or pathology.

Critically, the term “ordinary Korean families” is never defined or elaborated. Its meaning is taken for granted, which is precisely what makes it ideologically potent: it naturalises the ethnic majority as the unmarked, universal standard against which all other family forms are measured and found wanting. This rhetorical strategy thus performs the same symbolic boundary work as the government's inclusion rhetoric, but through the opposite affective register—where the state uses positive valuation to absorb difference, the media uses contrastive negation to expel it.

Interpreted through the lens of Social Identity Theory and the Stereotype Content Model, this divergence in cosine similarity is not merely a statistical curiosity; it indexes the differential construction of cognitive distance. In the governmental corpus, the narrowing vector space reduces the cognitive distance between the marked minority and the prototypical national family, implying, on a discursive level, that the two categories belong to a shared conceptual terrain. However, the moderate absolute magnitude (0.54 versus the benchmark 0.79) cautions that this reduction stops short of full semantic assimilation—the multicultural family remains a distinct satellite category rather than an unmarked core member. In the media corpus, the persistently low and stable similarity signals a deliberate discursive maintenance of cognitive distance, encoding the minority as an external entity. The media thus performs “divergent accommodation” (Gudykunst and Kim, 1992), accentuating intergroup salience to satisfy the ingroup's need for positive distinctiveness.

4.4 Synthesis of empirical patterns

Taken together, the three layers of empirical evidence—collocational profiles, semantic prosody distributions, and distributional semantic measurements—converge on a consistent narrative. Government discourse has undergone a lexical modernization, incorporating the vocabulary of inclusion and empowerment while preserving a structural paternalism in which multicultural families remain objects of state action. The positive semantic prosody envelops this paternalism in a caring affect, obscuring the continued dependency it constructs. Media discourse, by contrast, operates as a sustained counter-narrative that foregrounds conflict, discrimination, and policy failure, thereby exposing the gap between official rhetoric and social reality. The embedding analysis demonstrates that this discursive divergence has a measurable counterpart in conceptual space: the government narrows the distance between multicultural family and the normative Korean family, while the conservative press keeps the two categories firmly apart.

These findings set the stage for the theoretical discussion that follows, in which we interpret these patterns through the lenses of governmentality, critical multiculturalism, and the ideological limits of liberal inclusion.

5 Discussion

The collocational, prosodic, and distributional evidence converges on a central paradox: while the state's discourse has undergone lexical modernization, the multicultural family remains discursively annexed as a conditional adjunct to the normative Korean family. Conversely, the conservative media sustains a crisis narrative that, despite its critical function, re-inscribes symbolic exteriority. We now theorize these dual patterns through governmentality, intercultural communication, and the politics of recognition, focusing specifically on the ideological functions these discursive strategies perform.

5.1 Governmental logic and discursive annexation

At first glance, the government's diachronic lexical shift—from welfare and assistance toward empowerment and global talent—appears to signal a liberal multicultural turn. We argue, however, that this surface modernization constitutes a more sophisticated, neoliberal form of governmentality (Foucault, 2007). In Period 1, the state disciplined multicultural families through direct welfare provision and assimilative education, constructing them as deficient subjects. By Period 3, the governing vocabulary has been re-engineered around self-regulation: families are called to become entrepreneurial selves who cultivate their own empowerment and contribute to national competitiveness. This shift aligns with Rose (1990) “governing through freedom,” wherein the state devolves responsibility to individuals while retaining control through normative frameworks that define proper integration. The persistent dominance of support across all periods reveals continuity: the state remains the arbiter of legitimate inclusion, and its “support” remains conditional upon integration on Korean terms.

The narrowing cosine similarity between multicultural family and Korean family must be interpreted within this governmental logic. Rather than indexing genuine social convergence, this semantic proximity reflects a strategy we term discursive annexation. By infusing official lexicon with the language of diversity, inclusion, and global citizenship, the state symbolically absorbs difference into the national project without dismantling material hierarchies—residency precarity, employment discrimination, educational segregation—that continue to marginalize foreign spouses and mixed-heritage children. This dynamic resonates with Povinelli’s (2002) “cunning of recognition,” where minority subjects perform a sanitized culture in exchange for conditional inclusion, and with Hage’s (2012) critique of “white multiculturalism,” wherein the dominant group retains sovereign capacity to define tolerable difference. The Korean state's rhetorical embrace of global talent thus functions less as an invitation to pluralism than as a demand that multicultural families become recognizable, productive contributors within a still-ethnocentric national imaginary.

Crucially, this governmental logic is not captured by the classical assimilation–inclusion binary. It is not assimilation, because the state no longer explicitly demands cultural erasure; nor is it inclusion, because the normative centre remains unchallenged. Rather, it constitutes a third modality: discursive annexation—a mechanism through which the state symbolically incorporates minority groups by extending the lexicon of recognition (diversity, empowerment, global citizen) while leaving institutional structures and symbolic hierarchies untouched. Discursive annexation manages difference not through overt coercion but through the seductive vocabulary of belonging, making resistance difficult because the terms of inclusion appear to have been granted rather than demanded.

5.2 Media discourse and the maintenance of symbolic boundaries

The conservative media's persistently negative semantic prosody operates as a valuable counter-discourse that punctures the government's self-congratulatory narrative, exposing the gap between rhetoric and structural marginalization. Yet we caution against romanticizing this critical voice as ideologically neutral. The repeated framing of multicultural families through victim, problem, and conflict risks reifying their otherness just as effectively as the government's paternalistic benevolence. When media subject positions are limited to “suffering victim” or “failed policy object,” it becomes difficult to imagine multicultural families as agentive, ordinary community members.

A deeper ideological convergence emerges between the two ostensibly opposed discourses: both share the foundational assumption that the multicultural family is a marked category requiring special attention. Neither discourse deconstructs the category itself to envision a social landscape where families are simply families, irrespective of national origin. The absence of collocates such as leader, contributor, expert, or neighbour from both corpora signals a shared representational economy that cannot accommodate multicultural families as unmarked social participants. Van Leeuwen’s (2013) framework illuminates this: government texts deploy collectivization and functionalization, stripping individuality; media texts personalize but within a victimhood frame. Both position multicultural families as symbolically exterior to the normative “Korean family,” perpetually measured and found wanting.

The analysis of “ordinary Korean families”—appearing 31 times in MedCorp but never in GovCorp—reveals the media's contrastive strategy. Concordance inspection shows the phrase occurs overwhelmingly in constructions such as “While ordinary Korean families receive subsidies..” or “Unlike ordinary Korean families..” Here, ordinary functions not merely as statistical descriptor but as normative classifier, implying not just numerical commonality but social legitimacy. In the Korean context, ordinary (ilbanjeog-in) carries culturally specific weight, indexing conformity to unmarked social norms. By describing Korean families as “ordinary,” the media implicitly constructs multicultural families as extraordinary—and within negative prosody, that extraordinariness becomes synonymous with deviance. This naturalises the ethnic majority as the unmarked standard against which all others are measured.

5.3 Intercultural communication: essentialism, face, and accommodation

Reading these findings through intercultural communication theory illuminates dynamics a purely CDA-focused analysis might overlook.

The government's discursive annexation exemplifies what intercultural scholars term cultural essentialism (Holliday, 2010; Piller, 2025)—treating cultures as bounded, homogeneous entities. The state's repeated invocation of Korean family as unmarked normative category presupposes a singular cultural template to which multicultural families must orient, erasing the hybridity—mixing of linguistic practices, child-rearing norms, kinship structures—that characterizes transnational families. By framing integration as movement toward a fixed Korean centre, the government forecloses the possibility that Korean family forms themselves might be transformed through intercultural encounter.

The media's victimhood framing can be understood through Ting-Toomey and Dorjee’s (2018) face-negotiation theory. By consistently representing multicultural families as passive sufferers, the media engages in face-threatening discourse that denies positive face (desire for approval and respect) by positioning them as objects of pity, and negative face (desire for autonomy) by presenting them as perpetually in need of intervention. The cumulative effect is a representational regime that systematically undermines the symbolic resources needed to assert agency and claim belonging.

The persistent semantic gap between multicultural family and Korean family in media discourse reflects what Gudykunst and Kim’s (1992) communication accommodation theory would predict: a form of divergent accommodation. The media's contrastive connectors (while, unlike) and repeated invocation of ordinary Korean families maintain and accentuate symbolic boundaries between ethnic majority and minority. This serves the psychological function of reinforcing in-group solidarity among ethnic Korean readers, managing anxiety about demographic change by reassuring them that the boundary between “us” and “them” remains intact.

The coexistence of these dual discourses creates communicative ambiguity for multicultural families and the broader public. Official recognition may facilitate belonging, while media victimhood framing may undermine equal participation in public discourse. The net effect is a communicative environment in which multicultural families are simultaneously included in rhetoric and excluded in practice—a condition profoundly destabilizing for identity formation and the development of genuinely intercultural civic culture.

5.4 Methodological reflexivity: embedding instability, corpus size, and the ontological status of discourse

A rigorous methodological critique warrants explicit acknowledgment: the observed increase in cosine similarity within GovCorp—from 0.31 to 0.54—is a statistically significant measure of lexical co-occurrence within a relatively modest corpus (41,200–47,500 words per subcorpus). We must therefore consider the possibility that this vector-space convergence reflects embedding instability rather than genuine semantic drift. A skeptical reviewer might reasonably contend that the observed trend is an artefact of corpus-level frequency shifts: the mere proliferation of semantically polysemous terms such as global, diversity, and inclusion in Period 3 mechanically pulls the embedding of multicultural_family toward the broader, positively charged semantic field that also surrounds korean_family. In other words, the narrowing distance could be an epiphenomenon of lexical modernisation rather than a deep-structural re-evaluation of the outgroup.

We have addressed this concern through the sensitivity analyses reported in Section 3.2.4, which demonstrate that the monotonic trajectory of cosine similarity is robust to variations in vector size (100–300 dimensions), training epochs (25–100), and bootstrapping resampling. The consistency of the relative trajectory—rising in GovCorp, stable in MedCorp—across these configurations provides empirical grounds for treating the pattern as a genuine discursive feature rather than a statistical artefact. We emphasize, however, that these robustness checks do not eliminate the epistemological boundary identified above: cosine similarity measures co-occurrence proximity, not semantic equivalence or social proximity. The “confound” of lexical modernisation is, therefore, analytically informative rather than merely problematic.

From the standpoint of dialectical-relational CDA (Fairclough, 2003), this epistemological boundary does not invalidate our central thesis; rather, it refines it. Fairclough posits that discourse is not a passive mirror reflecting pre-existing social structures but a semiotic mode of social practice that constitutes those structures. Consequently, even if the semantic convergence is “merely” lexical, the institutional performance of that convergence carries constitutive effects. When the Korean state repeatedly, in its official English-language publications, places “multicultural families” in the same syntactic and topical environments as “Korean families”—through adjunctive inclusion (“including”) and comparatives of equality (“same quality of life as”)—it generates a discursive reality of official belonging. This textual practice produces a governance effect: it legitimises the state as a benevolent actor and frames minority integration as an accomplished fact, regardless of the lived material conditions of the target population.

The lexical modernisation confound is, therefore, not a fatal flaw but a window into governmental strategy. It reveals that the state's inclusive rhetoric is primarily a semiotic strategy (nation-branding) rather than a reflection of systemic reform. The vector drift is the evidence of discursive annexation: the state absorbs the vocabulary of global multiculturalism into its administrative lexicon while the institutional hierarchies (residency precarity, discrimination) remain unaddressed. By explicitly distinguishing between discursive proximity (lexical co-occurrence) and social proximity (material equality), our methodological design exposes precisely the gap that critical multiculturalism seeks to interrogate.

We also acknowledge a further limitation: our English-language corpus captures international-facing discourse, not comprehensive state ideology. The Korean-language domestic policy discourse may differ in register, collocational patterns, and ideological inflection. Parallel Korean-language analysis would test whether these patterns extend to domestic governance. Similarly, while the sensitivity analyses address stability concerns within our chosen parameters, future work employing transformer-based architectures (e.g., BERT, RoBERTa) with contextual embeddings could explore whether the observed patterns hold under more sophisticated distributional representations (see Bommasani et al., 2021). Such architectures, however, typically require substantially larger corpora (often >1 million words) to achieve stable contextualized representations, making them currently impractical for the period-specific, domain-constrained subcorpora in our design.

We thus invite readers to interpret the cosine similarity not as a measure of “real” integration, but as a metric of discursive absorption—a performative narrowing that coexists with persistent social distance. This interpretation respects the methodological limits of our data while leveraging those limits as analytical resources for understanding the semiotic dimensions of state power.

5.5 Methodological and theoretical contributions

By integrating collocation analysis, semantic prosody, and Word2Vec cosine similarity within a unified interpretive framework, we demonstrate the value of methodological triangulation. Each computational technique illuminated a distinct discursive dimension; only their combined application—subjected to qualitative interpretation—enabled a full accounting of ideological dynamics. As noted in Section 3.2.4, cosine similarity measures co-occurrence proximity, not semantic equivalence or social distance—a boundary condition that must be explicitly acknowledged. This triangulation responds to calls for reflexivity in corpus-assisted discourse studies (Mautner, 2009) and affirms that computational outputs serve as evidential springboards, not self-sufficient findings.

Theoretically, the Korean case offers three contributions. First, it refines governmentality theory by demonstrating English-language branding as a governmental technique—a strategic performance for external audiences that may differ from domestic policy practice. Second, the concept of discursive annexation—operating through lexical absorption, normative preservation, and conditional belonging—offers analytical purchase beyond Korea, for Asian societies grappling with demographic change and European contexts where multiculturalism continues through rebranded integration discourse. Third, the study shows how the Korean state's essentialism operates through official multiculturalism by essentialising Koreanness itself—a subtle but powerful form of symbolic boundary maintenance that aligns with what Piller (2025) terms “neoliberal multiculturalism,” where diversity is celebrated as a resource for national competitiveness while ethnic privilege remains undisturbed.

In sum, South Korea's English-language policy and media discourses constitute a contested terrain where the promise of inclusion is permanently shadowed by assimilationist imperatives. The intercultural lens reveals that both discourses, despite their opposition, share a foundation in cultural essentialism. The methodological design demonstrates that inclusion can be discursively performed while inequality remains institutionally preserved—a finding resonant far beyond the Korean peninsula.

6 Conclusion

This study interrogated the discursive construction of “multicultural family” across South Korean government policy documents and conservative English-language newspapers from 2006 to 2023. Through diachronic corpus-assisted critical discourse analysis, we demonstrated that the state's lexical modernization—from welfare provision toward empowerment, diversity, and global talent—masks a durable governmental logic. Collocation and semantic prosody analyses reveal that multicultural families remain objects of conditional support, required to demonstrate adaptation before earning inclusion. The rising cosine similarity between multicultural family and Korean family in government texts signals not structural convergence but strategic nation-branding that annexes difference without redistributing power. We theorised this as discursive annexation: a governmental technique that symbolically incorporates minority groups through the lexicon of recognition while leaving institutional hierarchies intact.

The conservative media sustains a sharply negative semantic prosody, foregrounding discrimination, policy failure, and hate speech. While this counter-discourse serves a democratic watchdog function, it entraps multicultural families within a victimhood frame that reinforces symbolic exteriority to an unmarked Korean norm. Through an intercultural lens, this constitutes face-threatening discourse that denies positive recognition, while contrastive rhetoric performs divergent accommodation that maintains majority–minority boundaries. The analysis of “ordinary Korean families” reveals how the media re-naturalises the ethnic majority as the unmarked standard. Both institutional voices share a foundational assumption: the multicultural family remains a marked category requiring special management, whether sympathetic or critical.

Methodologically, we demonstrated the value of integrating collocation analysis, semantic prosody, and distributional semantics within a unified interpretive framework. Each technique captures distinct discursive dimensions; only their combined application—subjected to qualitative interpretation—enables a full accounting of ideological dynamics. This triangulation advances corpus-assisted discourse analysis by showing how distributional measures serve as springboards for textured interpretation rather than self-sufficient findings.

Theoretically, we contributed the concept of discursive annexation for analysing state-led multiculturalism in ethnocentric societies, where official multiculturalism operates through inclusionary vocabulary while preserving normative hierarchies. We also demonstrated the value of integrating intercultural communication theories—cultural essentialism, face negotiation, communication accommodation—into CDA of migration discourse, illuminating subtle mechanisms through which majority-minority boundaries are maintained and contested.

Limitations suggest future directions. Our English-language corpus captures international-facing discourse, not comprehensive state ideology; parallel Korean-language analysis would test whether these patterns extend to domestic governance. Incorporating progressive media and exploring transformer-based architectures for pragmatic phenomena would further enrich analysis. Ethnographic research on how multicultural families receive and reappropriate these labels would ground discourse analysis in lived experience.

South Korea's multicultural policy discourse offers a cautionary tale: the language of diversity and empowerment can coexist with—and obscure—the preservation of symbolic hierarchies. By revealing the linguistic mechanisms through which this ideological work is accomplished, this study equips scholars to demand more than rhetorical accommodation—calling for recognition that refuses to condition belonging on assimilation and dismantles, rather than rebrands, the boundaries separating the multicultural from the ordinary.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Yookyung Lee, Chung-Ang University, Republic of Korea

Reviewed by: Ahsani Taqwiem, Lambung Mangkurat University, Indonesia

Daniel Dwamena Ofosu, Abetifi Presbyterian College of Education, Ghana

Footnotes

¹We acknowledge the sensitive historical resonance of "annexation" in the Korean context. Our usage is strictly analytical, referring to governmental strategies of symbolic incorporation, and is not intended to draw historical parallels to 1910 or to minimize the gravity of colonial history.