2.1 The concept of framing

Scholars have explored how PB travels and becomes integrated into local political systems using terms such as “idea journey”, “transfer”, and “diffusion” (Bartocci et al., 2023; Ganuza & Baiocchi, 2012; Röcke, 2014; Sintomer et al., 2008, 2016). Sintomer et al. (2008) summarise the three principles of the Porto Alegre model: grassroots democracy (where citizens’ assemblies set priorities and elect representatives), social justice (which ensures more funds go to areas with deficient infrastructure), and citizen control (where elected boards oversee the budget process to ensure local priorities are included). Therefore, the model is designed to render governance more inclusive and equitable, addressing inequality and promoting the well-being of residents, especially in disadvantaged areas.

However, Sintomer et al. (2008) found the diverging patterns of PB adoption in Europe, noting that while the Porto Alegre model served as an initial inspiration, no direct replication of the model has been implemented. In Asia, countries like the Philippines, Indonesia, and Korea began developing “PB-like” programmes in the 2000s, aiming to promote citizen participation in municipal-level policymaking within a broader context of national democratic reforms (Touchton et al., 2023b). In contrast to the grassroots activism seen in the Porto-Alegre model (Sintomer et al., 2008), national governments in these countries have mandated PB after evaluating local pilot programmes (Touchton et al., 2023b).

While some scholars have raised concerns about the lack of conceptual clarity caused by this diversity of PB (Miller et al., 2017), others, such as Röcke (2014, p. 9), have adopted a different approach by the reasons that people have “framed” PB differently in different national contexts. Röcke (2014) argues that frames matter when new ideas like PB are imported and transformed into policy practices. Frames are “relatively coherent but flexible idea packages that transport a certain meaning and change over time” (Röcke, 2014, p. 10). Framing might shape different discourses of “equality” (participation for everyone) or “equity” (participation for disadvantaged groups) regarding PB, creating distinctive guiding principles, values and practices (Escobar, 2021; Pape & Lerner, 2016). For instance, PB in Seoul, the capital city of Korea, has prioritised equality over equity with a vision of broad and open participation without an explicit goal of redistribution (No & Hsueh, 2022). Therefore, framing highlights specific aspects of an issue rather than simply its presence or absence (Hänggli, 2012).

Frames also facilitate political communication by providing common ground for discussing complex issues (Calvert & Warren, 2014; Van Gorp, 2010). In policy discourse, frames define social issues in a particular way and set the boundaries of cognitive representations, influencing how specific problems are defined, interpreted, and become policy agendas (Pan & Kosicki, 1993; Rein & Schön, 1993). Therefore, previous studies have examined how actors strategically create frames to influence public deliberation (Brewer & Gross, 2005; Pan & Kosicki, 2001).

If framing is a crucial practice, it is important to know whose frames matter most. Frames constructed by powerful actors, such as politicians and high-profile officials, could be more influential than those created by ordinary individuals. Therefore, Röcke (2014) studied framing in public discourses at the micro (e.g., politicians), meso (e.g., policy frames) and macro levels (e.g., nation-state framework) to provide empirical evidence on the power relationship between actors who diffuse PB frames. Röcke (2014, p. 5) found that the nation-state framework was crucial for developing country-specific PB features, as shown in the cases of France, Germany, and the UK.

France, known for having one of the most centralised state systems in Europe, framed PB as part of broader decentralisation reforms aimed at bridging the gap between municipal governments and the public. French politician Jean-Pierre Raffarin played a key role in promoting the concept of proximity democracy, which influenced how PB was framed in the country. In Germany, PB was framed with a focus on modernisation and transparency in public administration. Under the concept of the Citizens’ Town, citizens were encouraged to consult and offer perspectives on budgetary decisions, though their role remained largely advisory. In the United Kingdom, PB was integrated into a national strategy for community empowerment, particularly under the New Labour government. The focus was on empowering marginalised groups and addressing social exclusion through participatory processes.

Röcke (2014) observed that concepts like proximity democracy, the citizens’ town, and community empowerment became intertwined with PB in these countries, creating distinct approaches to the practice. These unique frames illustrate how PB is shaped by local contexts and political cultures. Overall, framing analysis helps to understand how PB is interpreted, adapted, and ultimately accepted by local institutions, allowing it to become indigenised (Escobar, 2021; Yoon & Lim, 2019).

2.2 Media framing

While the news media have received limited attention in the PB literature, media studies have long investigated media frames, recognising that people rely on the mass media for political information (Entman, 1993, 2007; Vliegenthart, 2012; Vliegenthart & Van Zoonen, 2011). At the micro-level, journalists and news outlets bring multiple frames to an issue, resulting in macro-level patterns of framing (Entman, 2007; Hendrickson & Tankard Jr, 1996). Therefore, news content is viewed “as a system of organised signifying elements that both indicate the advocacy of certain ideas and provide devices to encourage certain kinds of audience processing of the texts” (Pan & Kosicki, 1993, p. 55). Communication texts are essential for understanding the persistent selection and emphasis of metaphors, catchwords, slogans and symbols (Calvert & Warren, 2014; Entman, 2007; Pan & Kosicki, 1993; Vliegenthart & Van Zoonen, 2011).

Originating from Goffman (1974), frame(ing) analysis has become an established method in the social sciences for studying how frames shape perceived reality (van Dijk, 2023; Vliegenthart & Van Zoonen, 2011). This article adopts Entman’s (1993, p. 52) classical definition of framing:

To frame is to select some aspects of a perceived reality and make them more salient in a communicating text, in such a way as to promote a particular problem definition, causal interpretation, moral evaluation, and/or treatment recommendation for the item described (emphasis in original).

In this definition, selection and salience are crucial elements (Entman, 1993). Like cropping an image, selection involves setting boundaries within which a phenomenon is described, labelled, and classified, manifested by “the presence or absence of certain keywords, stock phrases, stereotyped images, sources of information, and sentences” (Entman, 1993, p. 52). Salience is about highlighting, referring to “making a piece of information more noticeable, meaningful, or memorable to audiences” (Entman, 1993, p. 53). If this pattern persists across time and outlets, it signifies a systematic media bias (Entman, 2007).

Previous studies have identified several factors influencing media framing, including political power, readership, partisanship, ownership and market competition (Hänggli, 2012; Puglisi & Snyder Jr, 2015). This article focuses on two factors: government intervention (legislative amendments) and newspaper types (local/national). First, government intervention, specifically through legislative amendments, has significantly transformed how PB has been practised in Korea, as municipalities are obliged to engage in it (No, 2017). Cho et al. (2020, p. 196) found that the government intervention “standardised the academic discussion”, with most studies after 2011 adhering to implementation and national PB guidelines. This article investigates how media framing has changed post-reform.

Second, newspaper type constitutes another crucial factor. Holt and Major (2010) argue that national and local newspapers differ in framing due to the scope of coverage and readership. Local newspapers often reflect local perspectives, covering detailed specific issues and promoting citizen participation because of their area-based readership. For instance, during the coverage of Korea’s nuclear safety inspection forgery scandal, Choi et al. (2016) found that national newspapers relied more on public officials as sources, while local newspapers interviewed residents and non-governmental organisations to highlight local concerns about safety and health. By analysing the national-local media framing of PB, this study provides insight into how the news media reflect varying perspectives from both national and local contexts.

3.1 Data

This article employed a structural topic model to analyse the media framing of PB in Korea. All news articles mentioning PB between 1991 and 2022 was collected using BigKinds, a Korean newspaper digital archive hosted by the Korea Press Foundation (https://bigkinds.or.kr). BigKinds provides access to daily news content and metadata from 54 Korean media organisations dating back to 1990.

The following Boolean search string was used to capture relevant articles: (resident OR citizen OR people) AND (“participatory budget”).¹ The search retrieved 31,259 records, including titles, content, keywords, publishing dates and newspaper names. While content refers to the unstructured text data of original news articles (e.g., “participatory budgeting is a good programme”), keywords are word tokens (e.g., “participatory budgeting,” “good,” “programme”), which are units in natural language processing. Common words, such as “is” and “a,” known as stopwords, were excluded as they offer little informational value. BigKinds provides all keywords from each news article and written content up to 200 words. Keywords were used for topic modelling after removing duplicate articles (with the same title and texts), empty strings, numbers, and website links. The number of articles processed for the analysis was 28,480.

3.2 Structural topic model

In recent years, topic models have received increasing attention as an automated content analysis method for investigating recurring topics in extensive text corpora (Heidenreich et al., 2019; Shin & Boonjubun, 2021; Walter & Ophir, 2021; Ylä-Anttila et al., 2022). These models operate based on the principle that a news article contains multiple topics in varying proportions, and each topic consists of words that tend to co-occur within a context (Blei, 2012). When sets of words frequently co-appear across documents, topic models identify these word clusters as distinctive topics. In this statistical approach, a word derives meaning from its relationship with other words (DiMaggio et al., 2013). For instance, the word “Donald Trump” might signify a president when associated with words like “election” and “Republican Party”, but it can also refer to an entrepreneur when combined with “real estate” and “golf clubs”. Researchers label these word clusters (e.g. Topic 1: President) by interpreting their meanings based on domain knowledge.

Ylä-Anttila et al. (2022) argue that topics derived from topic models can approximate frames based on two conditions. First, topic model outputs must enable researchers to discern how the media define a general problem, subject, causal interpretation and judgment, and recommendation – aligning with Entman (1993)’s definition of framing. The keywords in each topic must be interpretable, allowing researchers to identify distinct frames (internal validity). Second, external validity is another crucial checkpoint: the identified topics should align with the findings of previous studies (Ylä-Anttila et al., 2022). The present study assessed the external validity of the model by matching the model outputs with previous Korean PB studies (for a review, see Cho et al., 2020).

There exists a wide range of topic models, such as Latent Semantic Analysis (LSA), Latent Dirichlet Allocation (LDA) (Blei, 2012), the Structural Topic Model (Roberts et al., 2016), and, more recently, Top2Vec (Angelov, 2020), BERTopic (Grootendorst, 2022), and the Contextualized Topic Model (CTM) (Bianchi et al., 2021), which use pre-trained models to account for semantic similarity between words or sentences. Despite these advances, reviews comparing models have found that no single model performs best across all applications and settings (see Churchill & Singh, 2022; Egger & Yu, 2022). Therefore, it is crucial to clarify the rationale behind selecting a specific model for a given purpose.

This article employed the Structural Topic Model (STM), developed by Roberts and colleagues (2014, 2016, 2019), to infer hidden topic structures within a large corpus of news articles. The STM was particularly developed for social sciences, as it allows researchers to examine how document-level covariates (e.g., author, date) influence the content of the text (Roberts et al., 2016). This function is relevant to this study, which examines a collection of news articles contributed by various outlets over different periods. Previous studies found that the STM outperforms the popular LDA in capturing the relationships between covariates and the content of the text (Schmiedel et al., 2019; Wesslen, 2018). Therefore, STM has been widely used in studies analysing customer reviews (Hu et al., 2019; Park et al., 2020), social media (Dehler-Holland et al., 2021; Jo et al., 2022; Shin & Boonjubun, 2021; Stelmach & Boudet, 2022; Zhang et al., 2021) and academic literature (Bai et al., 2021). The present study used STM to examine the effects of government intervention (before/after becoming mandatory) and newspaper types (local/national) on the prevalence of topics (the proportion of a document devoted to a topic) using a generalised framework (Bai et al., 2021; He et al., 2020):

$\mathit{\text{Prevalence}}=g(\mathit{\text{Intervention}}, \mathit{\text{Type}}, \mathit{\text{Intervention}}\times \mathit{\text{Type}})$

4.1 Media coverage of participatory budgeting

Figure 1 shows the number of news articles mentioning PB over the past three decades (1991–2022). In Korea, PB was first introduced in 2001 through social campaigns led by a minor left-wing party and civic society organisations (Yoon & Lim, 2019). In 2002, PB entered the public discourse for the first time during local elections, when the party pledged to engage in PB (Kwack, 2007; No, 2017). Media discourse on PB began the same year, as shown in Figure 1.

Figure 1: The number of news articles that mentioned PB.

In 2003, Bukgu, a district in the City of Gwangju, piloted PB for the first time, formalising the process with an ordinance the following year as part of the pledge. Other early adopters, such as Ulsan and Daejon, also implemented the Porto Alegre model, influencing subsequent initiatives (Kwack, 2007; Yoon & Lim, 2019). These local efforts attracted attention from national policy-makers and soon aligned with the participatory agenda of the Roh Moo-Hyun government (2003–2007), which identified itself as the “Participatory Government” (Sintomer et al., 2012). The national government’s first recommendation for public participation in local budgeting came in 2003, encouraging municipalities to use online polls, public hearings and meetings (Ministry of the Interior and Safety, 2003).

After these initial pilots, Figure 1 shows three significant peaks in media coverage: 2006, 2011 and 2019. The first peak in 2006 followed the revision of the Local Finance Act in June 2005, which introduced Article 39, allowing local governments to implement PB at their discretion. The Article reads as follows: “The head of a local government may prepare and implement procedures for residents to participate in local budgeting processes, as prescribed by Presidential Decree [emphasis by the author]”. As a result, the idea of PB was first enshrined in national law, albeit on a non-mandatory basis. Although the national government provided the first national guidelines in 2006, municipalities were slow to adopt it due to inadequate training for public officials and negative attitudes from local councils, possibly due to perceptions that it would infringe their right to set budgets (Kwack, 2007). By 2007, only 62 municipalities (25.2%) had adopted PB (Kwack, 2007), leading to decreasing media attention between 2005 and 2010.

A second peak occurred in 2011, driven by the national government’s decision to mandate PB after a public scandal in Seongnam, which declared an unprecedented moratorium on its debts in 2010, and revelations of inefficient spending and corruption in a few municipalities (K. W. Lee, 2018). The national government introduced the second version of its guidelines with three exemplary models to enhance the transparency and capacity of local finance (No, 2017). In 2011, the revision of the Local Finance Act required all municipalities to adopt PB. However, the national government focused on standardising PB ordinances across municipalities while neglecting their local contexts (Yoon & Lim, 2019), thus limited the media’s interest in the topic.

The third peak began in 2017, following the Moon Jae-In government’s inclusion of PB as a national task. A 2018 revision of the Local Finance Act expanded PB to cover all budgeting processes, including national budgeting. By 2022, 235 municipalities (98%) were operating PB, a 67% increase from 2019, with 32,787 PB cases with a total budget of €5.6 billion (2.3% of municipal expenditure) (Ministry of the Interior and Safety, 2023).

Overall, Figure 1 reveals an increasing trend in media coverage, with 3,633 articles mentioning PB before it became mandatory and 24,847 articles afterwards. Local newspapers accounted for most of the coverage (68.7%, 19,559 articles), followed by business/finance newspapers (4,941), national newspapers (3,634) and broadcasting companies (346).² This distribution highlights the significant role of local media in framing PB.

4.2 Major topics and interpretation

A prerequisite of the STM involves specifying a hyperparameter, K, which controls the number of topics to be generated by the model. The present study used exclusivity and semantic coherence metrics to find the optimal range of topic numbers, and 17 topics were selected in this analysis (see Appendix 1).

Table 1 presents the results of the 17-topic model, showing the most representative keywords for each topic. The second column lists five keywords, with “P” indicating the highest probability words based on occurrence frequency. These probability-based keywords help identify words that frequently appear in each topic; however, common words, such as “budget” and “citizen”, appear across multiple topics due to the nature of the keyword search and thus may not provide semantic insight (Roberts et al., 2014, p. 5). Therefore, the second column also includes keywords ranked by the FREX metric, denoted as “F”. FREX is the weighted harmonic mean of a word’s rank based on frequency (how often it occurs in each topic) and exclusivity (how distinct it is to that topic compared to others), which provides more semantically distinct word sets (Lucas et al., 2015).

Topic correlations, which quantify the likelihood of two topics (e.g., Topic A and B), were then calculated. Figure 2 illustrates the topic correlation network, focusing on the positive correlations (Lucas et al., 2015). Previous STM analyses have used the Louvain community detection algorithm (Blondel et al., 2008) to identify prominent clusters in the topic correlation network (Jo et al., 2020; Ophir, 2018). The algorithm assigns densely connected nodes to communities that maximised a modularity score Q, a measure of the quality of network partition (Blondel et al., 2008). In this analysis, the algorithm identified four clusters, with a modularity score of Q = 0.5. In the final step, representative articles associated with each topic were manually inspected. The topics are labelled in the third column of Table 1. Topic labelling by reading highly associated articles is recommended in the STM manual (Roberts et al., 2019, p. 14).

Figure 2: A network of correlated topics. Note: node size is proportional to the estimated proportion of the topic (the fourth column in Table 1), and edge width reflects the strength of the correlation between topics. Different colours are assigned by the Louvain community detection algorithm based on topic correlation, indicating that topics with the same colour are more likely to appear in the same news article.

4.3 The effect of the intervention and newspaper types

Government intervention

Previous studies have shown that policy practices and academic discourse underwent significant change after PB was mandated in 2011 (Cho et al., 2020; No, 2017). In line with the STM manual (Roberts et al., 2019, p. 17), this study estimated the differences in the proportions of each topic between the pre-mandatory period (2001–2011) and the post-mandatory period (2001–2022) to investigate whether media discourse experienced a similar shift. Here, covariate analysis was used to evaluate how topic distributions varied across these periods, while confidence intervals were applied to assess the statistical significance of the changes (Roberts et al., 2014). Figure 3 presents a graphical display of these estimated topic proportions, and a detailed table can be found in Appendix 2.

Figure 3: Difference in topic proportions (pre-mandatory period vs. post-mandatory period).

In Figure 3, the dots represent the mean differences in topic proportions between two periods, while the bars display the 95% confidence interval of these differences. Topics with dots and bars located either to the right or left of the central dotted line demonstrate a statistically significant difference in prevalence between the two periods. For instance, Topic 2 (Open call) appeared approximately 9% more frequently in news articles after 2011 with a confidence interval of 0.08–0.1%, indicating a statistically significant difference. By contrast, Topic 14 (Award) showed a slight decrease of 0.5%, but its confidence interval of -0.01–0.002% suggests that this difference was not statistically significant. Overall, most topics showed significant shifts in media coverage after 2011, except for Topic 14.

A distinct pattern emerges from these findings as the media shifted their focus from political to implementation topics. As Ganuza and Baiocchi (2012, p. 1) noted, the global diffusion of PB has transformed it into a “politically malleable device” that requires translation and adaptation to local contexts. During its early phases, PB functioned as “a centrepiece of a political strategy” (emphasis in original), involving various actors, values, and conflicts. Figure 3 supports this observation, as all topics in the political cluster (e.g., governance, decentralisation, elections) were more prominent in media coverage before PB became mandatory. However, PB became decoupled from its Brazilian origins, evolving into a more “politically neutral device” (Ganuza & Baiocchi, 2012). The substantial increase in Topic 2 (Open call) reflects this transformation, as municipalities increasingly adopted PB in a more routine, operational context.

Newspaper types

Figure 4 illustrates the differences in how local and national newspapers framed PB, using the same method as in the earlier analysis. Local newspapers focused more on delivering practical information to residents about how to participate in PB processes, which is reflected in the higher prevalence of topics related to information dissemination, such as Topic 1 (Announcement) and Topic 2 (Open call). Additionally, local newspapers framed PB as a tool for regeneration (Topic 3), emphasising its role in revitalising communities and improving local governance at the county level (Topic 5).

Figure 4: Difference in topic proportions (local vs. national newspapers).

By contrast, national newspapers emphasised the political aspects of PB, highlighting national government initiatives and political campaigns that promoted PB as a mechanism for decentralisation. Topics such as budgeting (Topic 7), decentralisation (Topic 8), and elections (Topic 9) appeared more frequently in national coverage. National newspapers also tended to showcase actual cases illustrating what PB looked like in practice, reinforcing PB’s role as a political device through various implementation examples, including community (Topic 13), welfare (Topic 15), and facility improvements (Topic 16).

Thes findings indicate that national newspapers framed PB as part of the national government’s broader push for decentralisation, focusing on integrating PB into formal political structures and validating it through various local cases. By contrast, local newspapers framed PB as a tool for regenerating local communities, focusing more on local economic improvements and grassroots participation. Their neutral information dissemination was also key in helping residents stay informed about local affairs and events.