3.2.1 The representation of global media in reporting cross-Strait relations.

To examine how global news media represented the evolution of cross-Strait relations between 2014 and 2023, this study operationalizes media representations as aggregated patterns of attention, interactional balance, evaluative impact, and tone embedded in large-scale news reporting, rather than as direct reflections of underlying political reality. This operationalization is consistent with prior research that treats international relations as being mediated and reconstructed through news production processes rather than merely reported.

Drawing on structured event information from the GDELT, six dimensions are employed to capture different aspects of global media representations of cross-Strait relations.

(a) Attention Index. Let A(t) denote the attention index, NCT(t) is the number of events in month t for events initiated by Mainland China toward Taiwan Region and NTC(t) is the number of events in month t for events initiated by Mainland China toward Taiwan Region.

(1)

The Attention Index captures the overall salience of cross-Strait relations in global news coverage, reflecting the degree to which cross-Strait interactions enter the global media agenda in a given period. Logarithmic transformation is applied to reduce the influence of extreme values during periods of heightened media activity.

(b) Balance Index. Let B(t) denote the Balance Index, NCT(t) and NTC(t) is the same meaning in the formula (1). If B(t)=0.5, it means there is a perfectly balanced coverage between the two sides cross the Strait.

(2)

Deviations from this value indicate asymmetric attribution of initiative or agency in media representations, highlighting which side is more frequently portrayed as the actor in cross-Strait events.

Unlike the impact and tone indices, which rely on continuous event attributes and weighted aggregation, the Balance Index captures the structural directionality of media coverage by measuring the proportion of Mainland China-to-Taiwan Region events relative to total cross-Strait interactions. As this indicator is intended to describe the observed configuration of media representations rather than to estimate an underlying latent distribution, the main analysis reports the original monthly proportions. Bootstrap resampling is therefore not required for substantive interpretation and is used only as a robustness check in supplementary analyses.

(c) Impact Index. Let I(t) denote the Impact Index of cross-Strait relations, NumMentionsi the NumMentions (weight) for event i, and GoldsteinScalei the GoldsteinScale for event i. In t, the weighted average is given by:

(3)

Because GoldsteinScalei∈[-10,10], affine mapping I(t) translates the original [−10, 10] range to [0,1] enabling direct comparability with other normalized metrics.

(4)

To address potential noise arising from automated event coding, media duplication, and weighting effects in GDELT data, uncertainty estimation is conducted using nonparametric bootstrap resampling. Specifically, events within each month are resampled with replacement 2,000 times. For each resampled event set, the Impact Index are recalculated based on weighted aggregation of event attributes. The resulting bootstrap distributions are then used to derive monthly point estimates (means) and 95% confidence intervals. This procedure allows temporal peaks and troughs to be interpreted in terms of their statistical stability rather than as isolated point estimates.

(d) Tone Index. Let T(t) denote the tone of global reporting of cross-Strait relations, NumMentionsi the NumMentions (weight) for event i, and AvgTonei the average tone for event i. In t, the weighted average is given by:

(5)

Because AvgTonei∈[-10,10], affine mapping I(t) translates the original [−10, 10] range to [0,1] enabling direct comparability with other normalized metrics.

(6)

The same event-level bootstrap resampling strategy is applied to the Tone Index to ensure methodological consistency across indices. Using identical monthly resamples as Impact Index, the Tone Index is recalculated for each bootstrap iteration based on weighted average tone scores, and 95% confidence intervals are constructed from the resulting distributions.

(e) The Evolution of QuadClass and EventRootCode. GDELT assigns each event to one of four QuadClasses: Verbal Cooperation, Material Cooperation, Verbal Conflict, Material Conflict. For each month t, count frequencies Q(t)=[Q1, Q2, Q3, Q4]. CAMEO’s 20 category EventRootCode taxonomy covers a wide spectrum of political and social behaviors. Monthly frequency vector E(t)∈N20.

These categorical distributions are used to examine structural changes in the types of interactions emphasized by global media, rather than to estimate continuous trends; therefore, uncertainty estimation is not applied at this stage.

3.2.2 Identifying high-impact media outlets and conceptualizing the deviation index.

To determine which news organizations wield high influence in cross-Strait reporting, this study leverages GDELT’s SourceURL field. Each news event in GDELT is tagged with a URL. Parsing these URLs yields the hostname, which in turn serves as a proxy for the media outlet and its country/region of origin. This organizational-level identification is consistent with prior large-scale studies of international news flows and enables systematic comparison among media actors embedded in the global news ecosystem.

A cutoff of 1,000 total cross-Strait event mentions over the 2014–2023 period was chosen to distinguish outlets with sustained engagement from those with sporadic coverage. The initial domain list comprised 51 that exceeded the threshold. Two URLs (yahoo.com and news.yahoo.com) represent the same corporate entity and their counts were merged, yielding the final set of 50 high-impact outlets.

The analytical focus of this study extends beyond the concentration of global visibility to the internal differentiation among these already highly visible outlets. As discussed in Section 2.2, when global news attention is disproportionately concentrated among a limited set of media organizations, the extent to which these outlets converge or diverge in their reporting structures becomes theoretically consequential for understanding the production and fragmentation of global discourse. The Deviation Index is designed to capture this form of structural differentiation in reporting profiles, rather than to measure influence, ideological bias, or normative stance.

Conceptually, the Deviation Index operationalizes deviation as a multidimensional distance between an outlet’s longitudinal coverage pattern and the aggregate distribution produced by the global news system. The global baseline thus serves as a relational reference point, representing an emergent equilibrium of international reporting rather than a normative standard. Deviation is therefore analytically neutral: a higher index value indicates consistent divergence from the global baseline, not greater bias or stronger influence.

The measurement framework applied to fifty high‑impact outlets comprises six complementary dimensions: four continuous indices (EventCount, BalanceIndex, ImpactIndex, ToneIndex) and two categorical distributions (QuadClass and EventRootCode). Each dimension captures a different operational layer at which editorial differentiation may manifest, including reporting intensity, directional balance, event salience, emotional orientation, and the distribution of interaction types.

For continuous dimensions, each outlet’s monthly time series is compared to the global media baseline using the root-mean-square error (RMSE). To ensure scale invariance and comparability across dimensions, RMSE values are normalized by the baseline’s sample standard deviation σ, yielding σ-normalized RMSE (σ-NRMSE). These values are then standardized across outlets (Z-scores) and transformed using a logistic sigmoid function to produce bounded deviation scores in the interval (0, 1). This transformation ensures that no single continuous dimension dominates the composite index due to scale, variance, or outlier effects.

Categorical dimensions are modeled as monthly probability distributions over the four QuadClass categories and twenty EventRootCodes, respectively. Each outlet’s histogram is compared to the baseline distribution using the Jensen-Shannon distance, which symmetrizes and smooths the Kullback-Leibler divergence and yields values in the range [0, 1]. This approach captures divergence in the structural composition of reported interactions, rather than differences in volume alone.

The six dimension-specific deviation scores are aggregated using an unweighted arithmetic mean to form the overall Deviation Index. This aggregation is intended as a descriptive summary of overall editorial differentiation, rather than as an estimate of a latent variable. Equal weighting reflects a deliberate theoretical stance: no a priori assumption is made regarding the relative primacy of any single dimension in defining structural differentiation among high-impact media. Assigning equal weights thus avoids imposing normative or theoretically underdetermined hierarchies among reporting features and provides a transparent and replicable basis for cross-outlet comparison.

To address the concern that equal weighting constitutes a substantive analytical choice rather than a neutral default, this study explicitly evaluates alternative aggregation strategies and tests the sensitivity of the composite index to its weighting and metric assumptions. First, an empirically derived weighting scheme based on principal component analysis (PCA) was applied to the six deviation dimensions. The resulting weights are broadly balanced across dimensions, with no single metric exhibiting clear dominance, and the PCA-weighted composite index displays an almost perfect rank-order correspondence with the equal-weighted index (Spearman’s ρ = 0.991, p < 0.001). Second, a leave-one-dimension-out sensitivity analysis was conducted by iteratively recalculating the composite index after removing each dimension in turn. In all cases, the resulting indices remain highly correlated with the full specification (Spearman’s ρ ranging from 0.937 to 0.999), indicating that the overall structure of outlet differentiation is not driven by any single component. Third, robustness checks were performed by substituting alternative distance and error metrics in the construction of the deviation measures, including replacing RMSE with MAE for continuous dimensions and Jensen–Shannon distance with the Hellinger distance for categorical distributions. The composite index derived from these alternative specifications exhibits strong rank-order stability relative to the baseline index (Spearman’s ρ = 0.934), limited average rank displacement, and a high degree of quartile consistency across outlets. These analyses suggest that while equal weighting is a theoretically explicit baseline choice, the substantive conclusions drawn from the Deviation Index are robust to alternative weighting schemes, dimensional exclusions, and measurement specifications, and therefore reflect a stable underlying pattern of editorial differentiation among high-impact media outlets.

Accordingly, the Deviation Index should be interpreted as an integrative measure of how distinct a given high-impact outlet’s reporting profile is relative to the global news system as a whole. By focusing on relative deviation rather than absolute dominance, the index enables systematic comparison among media organizations originating from diverse institutional, political, and regional contexts, thereby empirically linking global visibility concentration to internal differentiation among media elites.

3.2.3 Textual analysis of significant news events with machine learning.

To operationalize news framing under conditions of mediatized politics, this study conducts a micro-level textual analysis of significant cross-Strait news events by integrating topic modeling and sentiment analysis. Following the computational framing approach outlined in Section 2.3, framing is conceptualized as a combination of issue emphasis and evaluative orientation, which can be approximated through the joint analysis of topic structures and headline sentiment.

Machine learning techniques are applied to conduct a micro-level textual analysis, which comprises three key steps.

(a) Selection of Textual Samples. News events are ranked based on the NumMentions metric, and the top 100 cross-Strait news reports per year are selected as the sample of significant events, totaling 1,000 reports over the ten-year period. 50 reports are selected from the “Mainland China-Taiwan Region” (CHN-TWN) group and 50 from the “Taiwan Region-Mainland China” (TWN-CHN) group annually. The full texts of these reports are retrieved via the SourceURL field. To ensure sample integrity, duplicate reports are systematically removed and replaced to maintain a consistent sample size of 100 reports per year. This selection strategy prioritizes events that are most salient within the global media agenda, thereby aligning the analysis with the mediating role of news coverage emphasized in mediatized politics. The rationale for focusing on high-salience reports is grounded in mediatized politics, where global political perception is primarily structured by highly visible news events rather than the total volume of sporadic reporting.

(b) Topic Model Analysis. Prior to topic modeling, all news texts undergo a standardized preprocessing pipeline designed to enhance semantic coherence while preserving issue-relevant expressions in cross-Strait reporting.

First, texts are tokenized at the word level using a rule-based regular expression tokenizer. All tokens are converted to lowercase, and punctuation, numerical strings, and non-alphanumeric characters are implicitly removed through tokenization rules. Standard English stop words are excluded using an externally defined stop-word list, which additionally incorporates corpus-specific functional terms (e.g., “a/an,” “according,” “anything”) that do not contribute to topical differentiation.

Second, to preserve semantically meaningful expressions frequently used in cross-Strait discourse, a predefined list of domain-specific multiword expressions is incorporated prior to tokenization. These include commonly occurring phrases such as “Hong Kong,” “Taiwan Independence,” “One China policy,” and “South China Sea,” and treated as single lexical units during modeling. This step mitigates semantic fragmentation and improves topic interpretability.

Third, lemmatization is applied using WordNet-based lemmatization to reduce inflectional variation while preserving part-of-speech consistency. After preprocessing, separate Latent Dirichlet Allocation (LDA) models are estimated for each yearly corpus, using the Gensim LdaMulticore implementation with 500 passes (iterations) to ensure model convergence. To improve computational efficiency, parallelized inference is employed with eight worker processes, and a fixed random seed (random_state = 42) is specified to ensure reproducibility.

Rather than selecting a single optimal topic solution a priori, LDA models are estimated across a predefined range of topic numbers (K = 2–10). Model performance is evaluated using both perplexity and topic coherence. Perplexity scores are computed based on held-in likelihood, while coherence scores are calculated using the original tokenized texts and dictionary representation. All topic solutions within the specified range are retained, and the top 30 keywords for each topic under each topic number are extracted and reported, allowing for systematic comparison of thematic structures across alternative model specifications.

Instead of treating topics as predefined frames, the resulting topic distributions are interpreted as the issue dimension of news framing, reflecting how cross-Strait relations are structured and emphasized in global media discourse over time.

(c) Sentiment Analysis of the Headlines. To capture the evaluative dimension of news framing, this study analyzes the sentiment orientation of news headlines. Headlines are chosen because they function as highly condensed, interpretive cues that frequently encode evaluative judgments and emotional tone.

Instead of relying on lexicon-based sentiment tools such as VADER, which may be limited in handling political and geopolitical contexts, this study employs a comparative large language model (LLM)-based sentiment classification approach. Two independent LLMs, GPT-4.1-mini (OpenAI) and Grok-4 (xAI), were accessed via their respective web interfaces to classify each headline as positive, negative, or neutral with respect to the core event subject. To ensure reproducibility and account for potential model updates, all sentiment classification queries were conducted on December 9, 2025.

To further discern the sentiment orientation toward different news subjects, the core event subject, distinct from the grammatical subject, is manually identified and annotated with its associated country or region. For example, in the headline “Experts Say Arms Sale to Taiwan Answer Defense Needs, But Spur New Questions”, although “Experts” serves as the grammatical subject, the core event subject is “Arms Sale to Taiwan”, which is thus tagged with the Taiwan region.

Given the possibility of model-specific bias, a multi-stage validation procedure is implemented. First, sentiment classifications produced by the two LLMs are compared. For headlines where the two models yield inconsistent sentiment labels, the cases are subjected to manual review. Two researchers independently examine these headlines and jointly determine the final sentiment classification through discussion and consensus, thereby resolving ambiguity in evaluative interpretation.

To assess and enhance the reliability of LLM-based sentiment classification, a multi-stage validation procedure is implemented.

First, sentiment labels produced by the two LLMs are compared. Headlines are divided into two subsets: cases where the two LLMs produce identical sentiment labels (LLM consensus cases), and cases where the two models yield different labels (LLM disagreement cases).

For instances where divergent classifications emerged between ChatGPT and Grok, human annotation was introduced as an external gold standard to evaluate the reliability of model-based judgments. Within this framework, a sample was operationalized as a valid identification in the context of model-assisted sentiment recognition if either model’s output aligned with the human benchmark. This dual-model validation strategy was employed to mitigate systematic bias inherent in relying on a single algorithmic architecture. Empirical results indicate that for samples with inter-model disagreement, the proportion of cases where at least one model corroborated the human annotation reached 94.14%. Furthermore, by synthesizing a unified predictive label and comparing it against human coding, we obtained a Cohen’s kappa of 0.9021, demonstrating that model-assisted sentiment classification maintains robust alignment with human judgment even in scenarios characterized by high uncertainty.

Regarding the subsample where Grok and GPT reached a consensus, the study eschewed the assumption of unconditional reliability, instead implementing a stratified random sampling protocol for manual verification. Specifically, 150 news headlines were selected from the consensus pool which is balanced across sentiment categories, for rigorous human auditing to ascertain the empirical accuracy of the model agreement. The verification revealed an inter-rater agreement rate of 0.96 and a Cohen’s kappa of 0.94, suggesting that the observed model consensus is grounded in empirical consistency rather than stochastic alignment.

Building upon these validation outcomes, the final sentiment variable was constructed through a multi-stage integration process. For all manually annotated or audited samples, human judgment was prioritized as the definitive result. For the remaining data, model outputs were adopted as the final sentiment labels only if consensus was achieved between ChatGPT and Grok. A subsequent data audit confirmed the absence of any residual cases involving model disagreement without human intervention, thereby ensuring that unverified coding results were excluded from the analytical phase. Finally, by integrating the certainty of human-coded samples with the empirical accuracy of the consensus-based samples, a weighted accuracy estimation for the aggregate sentiment variable was calculated at 0.973. Through this systematic triangulation, the study achieves rigorous validation and robustness control of automated sentiment analysis while circumventing dependency on any single computational tool.

4.1.1 Evolution of global attention and coverage balance.

The number of monthly reported events and attention index, which serves as an indicator of changes in global media attention (Fig 1). On an annual basis, the years 2016, 2022, and 2023 witnessed higher reporting volumes, each exceeding 40,000 events, with 2022 and 2023 surpassing 50,000 events, while other years recorded roughly 30,000 events. This pattern suggests that factors such as U.S.-China strategic competition and the Russo-Ukrainian war are reflected in the elevation of the Taiwan Strait as a hotspot of international focus, with the intensity of media attention still rising. At the monthly level, notable peaks occurred in November 2015, December 2016, August 2022, and April 2023. These peaks correspond respectively to significant events including the first Xi-Ma Meeting, the telephone call between President Trump and Tsai Ing-wen, Speaker Pelosi’s controversial visit to Taiwan, and Ma Ying-jeou’s visiting mainland China for ancestral worship, which demonstrates the global media’s sensitivity and timeliness in covering major cross-Strait events.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Monthly events and attention index of global reporting on cross-Strait relations.

https://doi.org/10.1371/journal.pone.0342755.g001

Over the sixty-month span from 2014 to 2023, the balance index exceeded 0.5 in 94 of the 120 months under review (Fig 2), indicating that, for the majority of this period, global media devoted greater attention to events initiated by the mainland China vis-à-vis Taiwan region rather than vice versa. Within the cross-Strait relations framing employed by international outlets, the mainland China consistently occupies the more more prominent role and is depicted as taking greater initiative in cross-Strait developments within global news narratives. This predominance is underpinned by mainland China’s expansive national capacity and its resolute pursuit of national reunification. The white paper “The Taiwan Question and China’s Reunification in the New Era” released in August 2022 declares that “resolving the Taiwan question and realizing China’s complete reunification is a shared aspiration of all the sons and daughters of the Chinese nation. It is indispensable for the realization of China’s rejuvenation” [48]. The mainland China has enshrined the “One China” principle as the non-negotiable precondition for establishing diplomatic relations. Amid intensifying great-power competition, mainland China has repeatedly cautioned the United States that the Taiwan issue represents a red line in Sino-U.S. relations, thereby underscoring its firm resolve.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Balance index of global reporting on cross-Strait relations.

https://doi.org/10.1371/journal.pone.0342755.g002

Nevertheless, this does not imply that Taiwan merely adopts a reactive posture or that it is deprived of global media attention. Indeed, the balance index has ranged between a maximum of 0.608 in January 2021 and a minimum of 0.446 in July 2014, suggesting international reportage of cross-Strait affairs has remained relatively balanced, without manifest clear bias.

4.1.2 Evolution of the impact index and tone index.

The Impact Index, derived from the GDELT’s GoldsteinScale, exhibits marked fluctuations in response to major cross-Strait events (Fig 3). In the aftermath of the 2014 Sunflower Movement, the index fell sharply before recovering during the Xi‑Ma meeting, and then declined again after the Democratic Progressive Party’s (DPP) election victory in early 2016.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Evolution of the impact index of global media.

https://doi.org/10.1371/journal.pone.0342755.g003

Based on bootstrap resampling, February 2014 exhibits the highest estimated monthly Impact Index (mean = 0.664, 95% CI: [0.652–0.677]), indicating an unusually strong concentration of high-impact cross-Strait reporting during this period. This peak coincides with a series of high-level political interactions, coinciding with State Council Taiwan Affairs Office Director Zhang Zhijun’s inaugural ministerial meeting with Taiwan Mainland Affairs Council (MAC) Chair Wang Yu‑chi in Nanjing [49]. One week later, Xi Jinping received former Kuomintang (KMT) chairman Lien Chan at the Great Hall of the People in Beijing, where he advocated for peaceful development as part of China’s national rejuvenation agenda [50]. Conversely, lowest estimated monthly Impact Index (mean = 0.427, 95% CI: [0.409–0.445]) was recorded in September 2021, when Mainland Chinese planes flew into Taiwan Region’s air defense identification zone (ADIZ) after Han Kuang military exercises [51]. Moreover, Taiwan applied to join the Comprehensive and Progressive Agreement for Trans-Pacific Partnership (CPTPP) during that month, which elicited a public rebuke from Beijing [52]. These concurrent military and political pressures correspond with the pronounced trough in the Impact Index and highlight the complex challenges facing cross-Strait relations in 2021.

The global media’s Tone Index on cross-Strait relations displays notable fluctuation (Fig 4). Based on bootstrap estimates, beginning in March 2015, the overall sentiment of coverage progressively shifted toward a more negative orientation and remained at a relatively low level for an extended period, with the monthly Tone Index means consistently falling below the neutral midpoint of 0.5 thereafter. And November 2014 represents a local maximum in the Tone Index time series (mean = 0.650, 95% CI: [0.642–0.657]), whereas April 2016 corresponds to one of the lowest estimated values (mean = 0.207, 95% CI: [0.193–0.221]) over the 120-month period analyzed.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Evolution of the tone index of global media.

https://doi.org/10.1371/journal.pone.0342755.g004

Both extremes are closely linked to significant political developments. The peak in November 2014 coincided with the 22nd APEC Economic Leaders’ Meeting held in Beijing, where Vincent Siew represented Taiwan under the designation “Chinese Taipei”. On November 9, Chinese President Xi Jinping met with Siew at the Great Hall of the People, reaffirmed the “1992 Consensus” and opposed “Taiwan independence”, emphasizing mutual trust and respect for divergent development paths as a foundation for long-term stability in cross-Strait relations [53]. This high-level engagement attracted substantial international media attention, contributing to the elevated Tone Index recorded that month.

In contrast, the trough in April 2016 corresponded to China’s extradition of 77 telecom fraud suspects from Kenya, including 45 Taiwanese for prosecution in the mainland [54]. This incident took place shortly before the inauguration of Tsai Ing-wen and was highly politicized by pro-independence media outlets in Taiwan. It was associated with a media storm and was widely interpreted by international outlets as a signal of rising cross-Strait tensions, contributing to the sharp decline in the Tone Index.

4.1.3 Evolution of event categories.

The evolution of the QuadClass of global media on cross-Strait shows that from the perspective of monthly fluctuations in cooperative and conflict events, November 2015, December 2016, August 2022, and April 2023 emerge as peak periods for cross-Strait cooperation, whereas December 2016, August 2022, and April 2023 similarly register heightened episodes of conflict (Fig 5). These convergence points in cooperation and confrontation generally coincide with significant political developments affecting both sides of the Taiwan Strait.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Evolution of the QuadClass on cross-Strait relations.

https://doi.org/10.1371/journal.pone.0342755.g005

It is notable that cooperation and conflict events frequently co-occur within the same month. This seemingly paradoxical pattern likely results from the classification mechanism of GDELT’s CAMEO-based coding algorithm. Due to the limitations of natural language processing, the system relies principally on keyword matching to infer event typologies, leading at times to the incorrect classification of ostensibly cooperative actions reported in a negative context. For example, the December 2016 phone call between President Trump and Tsai Ing-wen was widely regarded as a negative incident, provoking strong protests from Beijing. Yet in the GDELT, media coverage of this event appears as a series of cooperative behaviors. The most frequent recorded event codes include “Consult”, “Make an appeal or request”, “Engage in negotiation”, “Make a visit”, “Make a statement”, and “Host a visit”. GDELT may interpret Beijing’s political appeals and public statements as evidence of cross-Strait cooperation, thereby producing a misleading cooperation narrative in the GDELT.

The evolution of 20 different event categories on cross-Strait relations reported by global media illustrates that, during the decade, consultative and declaratory events have consistently dominated cross-Strait interactions (Fig 6). “Consult” (EventRootCode 04) ranks first in frequency, with pronounced peaks in November 2015, August 2022, and April 2023, reflecting moments when institutional dialogue was most vigorous. “Make public statement” (EventRootCode 01) and “Appeals” (EventRootCode 02) rank second and sixth in frequency. Their surges in December 2016, February 2018 (the trough) and August 2022 closely match the peaks in consultative events, suggesting that official pronouncements and public entreaties often precede or coincide with high‑level talks. “Engage in diplomatic cooperation” (EventRootCode 05) and “Express intent to cooperate” (EventRootCode 03) rank fourth and fifth. Their peak months almost perfectly overlap those of consultations, suggesting that once dialogue begins, both sides promptly move to formalize collaborative frameworks.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Evolution of the EventRootCode on cross-Strait relations.

https://doi.org/10.1371/journal.pone.0342755.g006

Meanwhile, “Engage in material cooperation” (EventRootCode 06) and “Provide aid” (EventRootCode 07) remain relatively infrequent yet exhibit a notable uptick in August 2022, implying that even amid heightened tension humanitarian and technical exchanges endure as limited safety-net channels. Other mid-frequency modalities, including “Yield” (EventRootCode 08) and “Demand” (EventRootCode 10), tend to track alongside material aid activities, serving as routine instruments within broader negotiation sequences.

In contrast, the conflict and deterrence cluster consists of “Disapprove” (EventRootCode 11), “Reject” (EventRootCode 12), “Threaten” (EventRootCode 13) and “Exhibit military posture” (EventRootCode 15). These events surged together in August 2022 and again in April 2023, coinciding with rises in “Reduce relations” (EventRootCode 16) and “Fight” (EventRootCode 19). This pattern illustrates how reports of provocations by pro-independence actors in Taiwan or external interventions are frequently accompanied by media-coded both diplomatic rebukes and demonstrations of force.

Nevertheless, extreme forms of violence such as “Engage in unconventional mass violence” (EventRootCode 20), “Assault” (EventRootCode 18) and “Protest” (EventRootCode 14) remain exceptionally rare. The global media representation indicates that despite occasional confrontations, the two sides have generally maintained a restrained, peace-oriented posture grounded in the “1992 Consensus” and a preference for resolving disputes through dialogue rather than coercion.

Overall, the media-recorded dynamics over the decade can be described as alternating phases of consultation, declaration and cooperation punctuated by periods of opposition and deterrence, yet consistently framed within a broader narrative of peaceful engagement.

4.2.1 Structural visibility and event production among high-impact media.

Fifty high-impact media outlets were identified based on sustained event production in cross-Strait reporting (Table 1). These media collectively contributed 166,796 reports, accounting for 42.18% of all cross-Strait events in the dataset. Rather than representing a simple aggregation of prolific news producers, this subset constitutes a structurally salient stratum within the global news ecosystem, occupying a disproportionate share of informational visibility.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. High-impact media in cross-Strait reporting.

https://doi.org/10.1371/journal.pone.0342755.t001

A pronounced pattern of visibility concentration characterizes the global reporting structure. Although more than 8,000 unique media domains appear in the GDELT records of cross-Strait events, fewer than 1% of outlets account for over 40% of total coverage. This skewed distribution indicates that global narratives on cross-Strait relations are disproportionately represented by a relatively small number of media actors with sustained capacity for event reporting and international dissemination. Under such conditions, the internal dynamics among these high-impact outlets become analytically consequential.

The structural salience of high-impact media does not map neatly onto national power or institutional prominence. Mainland Chinese outlets, including China News Service, Global Times, China Daily, People’s Daily, and Xinhua News Agency, constitute only a limited share of total cross-Strait coverage, despite their organizational scale and state affiliation. This pattern reflects broader asymmetries in global communication, in which international visibility is unevenly distributed and shaped by transnational media infrastructures rather than solely by domestic media capacity.

Beyond the dominant Anglo-American and Taiwanese outlets, regional clusters of high-impact media emerge in Southeast Asia and the Middle East. Media organizations based in Singapore and Malaysia display sustained engagement with cross-Strait developments, reflecting regional security concerns and economic interdependence. Similarly, outlets from the United Arab Emirates, Jordan, and Qatar occupy visible positions within the dataset, suggesting efforts by non-Western actors to establish globally resonant news platforms within an otherwise uneven communication order.

Digital-native platforms and news aggregators further amplify the concentration of visibility. Outlets such as MSN, Yahoo News, Global Security Network, Breitbart News, and World News account for a substantial share of reporting volume, underscoring the growing role of platform-mediated dissemination in the circulation of global news flows. Importantly, the prominence of these outlets reinforces the centralization of attention around a limited number of nodes, rather than expanding discursive diversity.

These patterns indicate that global reporting on cross-Strait relations is characterized by a highly concentrated visibility structure. As a result, the question is no longer whether high-impact outlets matter, but how they differentiate their reporting practices within this structurally unequal environment. This structural condition motivates the subsequent analysis of internal differentiation among high-impact media, operationalized through the Deviation Index.

4.2.2 Mapping structural differentiation among high-impact media via the deviation index.

After computing Deviation Index for high-impact media outlets, substantial heterogeneity emerges within this structurally salient group (Table 2). These differences unfold within a subset of outlets that already command disproportionate visibility in global reporting on cross-Strait relations, underscoring that concentrated attention does not entail narrative or structural homogeneity.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. The deviation index of high-impact media.

https://doi.org/10.1371/journal.pone.0342755.t002

The outlets with the lowest Deviation Index values, including Want China Times, Taipei Times, Focus on Taiwan, The China Post, and Eye on Taiwan, exhibit reporting profiles that closely approximate the aggregate global baseline across all six dimensions. This proximity does not imply neutrality or lack of editorial agency. Rather, it reflects a high degree of structural embeddedness within the prevailing equilibrium of global news production. As frequent contributors with sustained event volumes, these outlets participate directly in defining the statistical baseline, thereby reproducing dominant patterns of event distribution, tonal orientation, and issue framing within the global news ecosystem.

In contrast, the highest deviation index values occur for World News Network, The Economist, People’s Daily, Fox News and The Washington Post, indicating pronounced editorial divergence from the aggregate baseline. These deviations are not simply the result of limited coverage or occasional attention; rather, they indicate consistent differences in how highly visible media outlets select, frame, and emphasize cross-Strait events.

People’s Daily provides a particularly illustrative case of such internal differentiation. Over the ten‑year period, it recorded cross-Strait coverage in only 97 months, with fewer than ten events in 55 of those months, yielding a low EventCount score. Its mean balance index of 0.590 exceeds the global average of 0.539, reflecting a systematic emphasis on mainland‑initiated actions. The outlet’s QuadClass distribution further illustrates this orientation. Compared with the global baseline, People’s Daily allocates a substantially higher share of coverage to verbal cooperation (71.4% vs. 59.6%), while systematically underrepresenting material cooperation (5.9% vs. 9.0%), verbal conflict (16.6% vs. 19.9%), and material conflict (6.2% vs. 11.3%). On the continuous indices, its Impact Index is markedly higher than the global mean (1.149 vs. 0.352), whereas its Tone Index remains less negative (–1.223 vs. –1.767), indicating a relatively restrained affective framing. At the level of event typology, People’s Daily disproportionately emphasizes low-intensity, symbolic, and policy-oriented interactions. EventRootCode analysis shows that “Consult” (34.0%), “Make Public Statement” (14.2%), and “Express Intent to Cooperate” (7.0%) occur more frequently than in global coverage (25.5%, 11.2%, and 6.9%, respectively), while overtly confrontational actions, such as “Threaten” (2.4%), “Exhibit Military Posture” (0.5%), and “Fight”(1.6%), are consistently marginalized, each accounting for roughly half or less of their global proportions. These deviations reflect not episodic editorial selectivity but a stable and institutionally grounded reporting logic. As the official newspaper of the Chinese Communist Party, People’s Daily functions as a central policy communication instrument. Its coverage of cross-Strait relations prioritizes the signaling of Mainland China’s Taiwan policy, including emphasizing peace, stability, consultation, and national unity, while deliberately downplaying conflictual or militarized framings. This role helps explain why the outlet systematically exhibits a more positive and cooperative profile relative to the global reporting equilibrium: its primary function is not to mirror the diversity of international news narratives, but to articulate and disseminate the Mainland’s authoritative stance on cross-Strait relations within the global information environment.

Crucially, the Deviation Index captures differentiation independently of media influence as conventionally measured. Univariate and multivariate regression analyses reveal no statistically significant association between average NumMentions per event and any individual deviation dimension or the composite index. This finding confirms that editorial differentiation among high-impact media is not a function of prominence or visibility per se. Instead, it reflects how structurally salient actors exhibit selective patterns of attention, framing, and tone within a highly unequal communication environment.

The Deviation Index is statistically independent of media influence as measured by the average NumMentions per event. Uni-variate and multivariate regression analyses demonstrate no significant association between influence and any of the six deviation dimensions or their composite, indicating that editorial distinctiveness does not inherently correlate with an outlet’s prominence in cross-Strait reporting. This finding confirms that editorial differentiation among high-impact media is not a function of prominence or visibility per se. Instead, it reflects how structurally salient actors selectively organize attention, framing, and tone within a highly unequal communication environment.

In this sense, the Deviation Index does not rank outlets by influence or ideological extremity, but identifies the spectrum of editorial divergence among those already positioned at the center of global visibility. The coexistence of low-deviation stabilizers and high-deviation differentiators within the high-impact media cluster highlights the dual structure of global communication imbalance: while visibility is externally concentrated, narrative and editorial practices remain internally differentiated.

4.3.1 Issue-based framing of cross-Strait relations via thematic modeling.

By balancing perplexity and coherence, the optimal number of topics for each year was determined through manual review of the original texts and the summarization of thematic keywords (Table 3). The detailed keywords of the themes are presented in S1 Table of supporting information.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Optimal number of topics, perplexity, coherence and themees for annual coverage.

https://doi.org/10.1371/journal.pone.0342755.t003

From 2014 to 2023, global media coverage of cross-Strait relations distilled into five dominant frames, each illustrated by representative topics: (1) event-centered conflict and risk framing, emphasizing crises like South China Sea disputes (2014T1), aviation disasters (2014T2; 2015T1), strategic squeezes amid U.S.-China rivalry (2018T3), and Pelosi’s visit framed as the catalyst for the Fourth Taiwan Strait Crisis (2022T4), illustrating the media’s tendency to prioritize high-impact events over routine interactions; (2) electoral politics and internal contestation framing, interpreting dynamics through Taiwan’s domestic lens via the Sunflower Student Movement (2014T4), election cycles (2015T4; 2018T1; 2019T2), and Tsai Ing-wen’s inauguration (2016T1), underscoring sovereignty debates’ evolution through democratic processes; (3) great-power rivalry and strategic instrumentalization framing, positioning Taiwan as a U.S.-China pivot in the Trump-Tsai call (2016T1), Indo-Pacific alignments (2018T3; 2019T4), high-level visits (2020T2), and military tensions (2021T4), de-emphasizing bilateral history for geopolitical signaling; (4) ideological governance and regime contrast framing, juxtaposing mainland control with Taiwan’s norms in human rights cases (2017T2; 2021T3) and Hong Kong events like protests (2014T5; 2019T5) and the National Security Law (2020T4), with declining prominence post-2020; and (5) functional interaction, disruption, and conditional reopening framing, addressing exchanges such as solar trade disputes (2014T3), Foxconn investments (2017T6), pandemic responses (2020T1; 2021T2), and post-pandemic diplomacy (2023T1; 2023T3), coexisting with security narratives. Collectively, these frames highlight a securitization shift influenced by Western biases, interplaying domestic, regional, and global factors in cross-Strait discourse.

First, major event reporting functions as a primary organizing logic through which cross-Strait relations are rendered visible to global audiences. Early coverage situates Taiwan-related issues within broader regional contexts, such as maritime disputes and regional security tensions, as in 2014T1, or frames Taiwan as a site of transnational risk through aviation disasters and commercial incidents, exemplified by 2014T2 and 2015T1. Nevertheless, Western media’s disproportionate share of coverage has led to a prevalence of conflict incidents in headlines, marginalizing themes such as Mainland China’s reunification policies, popular affinity across the Taiwan Strait, and international adherence to the “One China” principle.

Second, thematic emphasis shifts from non-security-centered issues in the earlier years to predominantly security- and sovereignty-focused narratives after the late 2010s. Before 2018, global reporting frequently approached cross-Strait relations through economic, industrial, and social lenses that did not place Taiwan at the center of military confrontation. Topics such as trade disputes in the solar industry in 2014T3 and Foxconn’s investment in the United States in 2017T6 portray Taiwan primarily as an economic actor embedded in global production and policy networks rather than as a focal point of strategic risk. From 2018 onward, however, military security, diplomatic confrontation, and great-power rivalry increasingly dominate global headlines. Taiwan is progressively framed as facing mounting external pressure amid intensified U.S.-China strategic competition, as illustrated by 2018T3 and subsequent coverage of high-level visits and security alignment in 2020T2. In this later period, economic and social themes do not disappear but are increasingly subordinated to security-oriented narratives. A limited reappearance of interaction-focused reporting occurs only in 2023, following post-pandemic reopening and symbolic cross-Strait engagements such as Ma Ying-jeou’s mainland visit in 2023T1.

Third, global media recurrently frame cross-Strait relations through Taiwan’s internal political contestation and its entanglement with broader ideological narratives. The 2014 Sunflower Movement, highlighted in 2014T4, marks a critical turning point by foregrounding societal apprehension toward economic integration with the mainland. Subsequent reporting increasingly depicts cross-Strait relations as filtered through electoral politics, partisan competition, and debates over democratic legitimacy, as seen in 2018T1 and 2019T2. Pro-independence forces are often portrayed as invoking Hong Kong as a rhetorical reference point, generating narratives that conflate Taiwan’s sovereignty debate with other regional autonomy movements, particularly in 2014T5 and 2019T5. Since 2020, however, the implementation of Hong Kong’s National Security Law and the decline of large-scale protest movements have reduced the prominence of this framing.

These frames demonstrate that global media coverage of cross-Strait relations from 2014 to 2023 has evolved from event-driven and low-politics reporting toward a predominantly high-politics narrative structured by electoral contestation, ideological contrast, and great-power rivalry. While moments of functional interaction and social exchange periodically re-enter the media agenda, they remain discursively secondary to conflict-oriented and strategic framings, producing a complex yet asymmetrically securitized portrayal of cross-Strait relations.

4.3.2 Actor-sentiment framing in cross-Strait news headlines via LLM tools.

The principal actors and predominant sentiment categories appearing in the most consequential cross-Strait news headlines during the 2014–2023 period are also summarized (Table 4). Rather than merely reflecting episodic developments, these headline-level patterns reveal how global media selectively reconstruct the discursive representation of political reality through recurrent framing devices under conditions of intensified geopolitical mediatization. Three salient patterns emerge from this global media sample.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Sentiment classification of headlines by nation/region.

https://doi.org/10.1371/journal.pone.0342755.t004

Negative sentiment dominates the most consequential news headlines related to cross-Strait relations, whether the focus is Mainland China, Taiwan Region, or the United States, a large majority of affective terms are negative. From a framing perspective, this dominance should not be interpreted solely as a passive mirror of worsening relations, but as the interplay between real-world tensions and a media logic that privileges conflict, crisis, and threat as cognitively salient news frames. This prevalence reflects two convergent factors. First, the objective escalation of cross-Strait tensions, including Taipei’s policy reversals, intensified great-power involvement, and a higher frequency of confrontational incidents, has provided fertile discursive material for conflict-centered framing. Second, the inclusion of disaster reporting, most prominently the two TransAsia Airways crashes in July 2014 and February 2015, which together claimed over one hundred lives, including twenty‑eight mainland passengers, introduces non-political tragedy into the dataset; however, these events are nevertheless integrated into broader narratives of risk and instability surrounding cross-Strait interaction.

Actor focus in headline narratives has shifted markedly over the decade. In 2014–2016, references to Mainland China (inclusive of Hong Kong) and Taiwan appeared with comparable frequency, while US‑centric headlines were relatively rare. From 2017 onward, however, U.S.-focused headlines increased steadily and by 2021 surpassed those centered on Taiwan. This transformation signals a reconfiguration of dominant frames, in which cross-Strait relations are progressively subsumed under the broader interpretive schema of China-U.S. strategic rivalry.Rather than being framed primarily as a bilateral or intra-Chinese issue, the Taiwan Strait increasingly functions as a symbolic and strategic node within global power competition, reflecting the growing influence of international media logics that prioritize great-power confrontation over localized political dynamics.

The composition of other international actors has also evolved. Although constituting a smaller share of coverage, countries such as Japan, Canada, Australia, and India remain consistently present, indicating sustained interest in Taiwan‑related affairs. Their recurring visibility suggests that global media framing of the Taiwan issue extends beyond immediate stakeholders, embedding it within a wider network of alliance politics and normative positioning. Periodic attention to Taiwan’s diplomatic partners corresponds with episodes of diplomatic realignment, illustrating how moments of diplomatic rupture are selectively elevated as newsworthy signals of cross-Strait instability. After 2016, Southeast Asian states recede from headline prominence, likely reflecting a stabilization of Sino-ASEAN relations, particularly in trade, investment, and South China Sea cooperation, as well as a declining compatibility between routine regional cooperation and the conflict-oriented news frames favored by international media.

These patterns underscore how changes in geopolitical structure and media logic jointly condition sentiment distribution and actor visibility, structuring global news framing of the Taiwan issue in a manner consistent with the dynamics of mediated politics which refracts and interprets diplomatic reality.