Abstract: This study examines how the People’s Republic of China and Taiwanese media frame United States-Taiwan arms sales using unsupervised text mining methods. It analyses English-language news coverage from 2019 to 2026, a period marked by delayed U.S. arms transfers and political debate over Taiwan’s defence spending. Drawing on articles from Taiwanese and PRC-affiliated media, the study compares term frequencies, word associations, and topic structures. The results show that Taiwanese media primarily frame arms sales in terms of defence readiness, security, and policy, while PRC-affiliated media emphasise sovereignty, territorial integrity, and external interference. These differences suggest that the two media systems interpret the same arms sales through distinct political and security narratives.

Problem statement: How do Taiwanese and PRC media differ in their presentation of United States–Taiwan arms sales and related political debates?

So what?: Analysts and policymakers should treat the various narratives as political signalling about deterrence, sovereignty, and international positioning. Understanding these differences is necessary to interpreting how arms sales are communicated and contested across media systems.

Introduction

A record backlog of delayed United States arms transfers and a stalled special defence budget have made military procurement a contested political issue in Taiwan. By early 2026, Taiwan awaited roughly $32 billion in approved but undelivered U.S. weapons, including a December 2025 package featuring HIMARS artillery systems, loitering munitions, and anti-tank missile systems.[1] At the same time, the Lai Ching-te administration proposed a $39.5 billion special defence budget for 2026–2033 focused on air defence and drones, but the plan has been repeatedly blocked in the Legislative Yuan, where opposition parties support a much smaller NT$400 billion (approximately USD 12.7 billion) alternative that excludes key systems. An additional U.S. arms package valued at around $14 billion also remains uncertain following high-level U.S.-People’s Republic of China diplomatic engagement.[2] Although some transfers are proceeding, delays between approval and delivery continue to shape debate in Taiwan over defence readiness and the reliability of U.S. security commitments.[3]

Media coverage shapes how these developments are understood. Even when different outlets report the same facts, they can present them differently through framing and word choice.[4] As a result, the meaning of United States-Taiwan arms sales is not fixed but shaped by how they are communicated.

PRC and Taiwanese media often present similar events differently, reflecting differences in political context and media systems.[5] In conflict, media narratives can reinforce distrust and ideological division rather than provide neutral descriptions.[6] Digital media further strengthen these patterns by increasing exposure to similar, reinforcing viewpoints, contributing to echo chambers and polarisation.[7] Over time, repeated exposure makes alternative perspectives less visible.[8]

These dynamics are especially important in Taiwan, where media narratives are closely tied to national identity and sovereignty. Prior research shows that media play a key role in shaping how cross-Strait relations are understood,[9] and identity is influenced by continued exposure to political information within specific media environments.[10]

United States-Taiwan arms sales provide a clear case to examine these dynamics. Although they are formally defence transactions, they also serve as political signals regarding deterrence, alliance credibility, and international positioning.[11] As a result, how these arms sales are described reflects broader narratives about security, sovereignty, and political responsibility.

Methodology and Research Question

Through language patterns, term associations, and topic structures, the variance in how each country frames these issues becomes evident. The central research question is how these media differ in their presentation of arms sales and related political debates.

The analysis applies unsupervised text mining methods, including term frequency analysis, term association, and topic modelling, to compare language use across media groups. The objective is to identify systematic differences in framing through variation in word distributions and thematic structure.

• H1: PRC state-affiliated media and Taiwanese media differ in how they frame United States–Taiwan arms sales.
• H01: There is no systematic difference in framing.
  • Sub hypothesis 1a: PRC media use more terms related to risk, provocation, and external interference.
  • Sub hypothesis 1b: Taiwanese media use more terms related to defence, deterrence, and security.
  • Sub hypothesis 1c: Topic distributions differ across groups, reflecting distinct thematic structures.

The unit of analysis is the individual news article. The dataset consists of English-language news articles collected from Nexis Uni and ProQuest Global Newsstream using the search terms “United States” AND “arms sales” from January 1, 2019, to March 29, 2026.

The Taiwanese dataset comprises 90 articles, totalling about 393,580-word characters, with 75 from Taiwan News (2019–2026) and 15 from The China Post (2019–2020). The PRC dataset includes China Daily, People’s Daily, and South China Morning Post, with 580 articles and a total of about 3,566,102 word characters.

Both datasets were downloaded as aggregated files rather than structured article-level data. To reconstruct individual observations, the raw text was split into segments using paragraph breaks (“\n\n”) as separators. Empty segments were removed, and short fragments were filtered using minimum length thresholds. This step is necessary because methods such as Latent Dirichlet Allocation require multiple documents as input. Without splitting, the dataset would be treated as a single document, making topic modelling and document-level analysis impossible. As a result, each document serves as a meaningful unit for analysis rather than a fragmented text.

The final dataset is organised as a corpus, with each document representing a single article and each labelled by media group. The PRC corpus is substantially larger than the Taiwanese corpus. Rather than balancing the data, each corpus is treated as representative of its media environment, not directly comparable in size. To account for this difference, normalised measures such as topic proportions are used, and results are interpreted with attention to corpus size.

The Taiwanese dataset is processed as a standard corpus, while the PRC dataset is constructed as a VCorpus due to differences in data structure after extraction, particularly the need for additional cleaning after text reconstruction. The PRC data requires more iterative cleaning following reconstruction from aggregated text, whereas the Taiwanese data is more structured and requires fewer steps. Preprocessing includes lowercasing, removal of punctuation, numbers, and stopwords, as well as filtering out high-frequency but low-information terms such as “Taiwan,” “China,” “United States,” and “said.” Word stemming reduces terms to their root forms.

Additional dataset-specific cleaning is applied. Non-content artefacts introduced during the database export, such as “proquest”, “document”, and “url”, are removed. Non-alphabetic characters are filtered, and unusually long tokens are excluded to eliminate noise such as URLs or encoding errors. These steps ensure the analysis focuses on substantive language. Document-length filtering is used as a quality-control step. Documents below a minimum character threshold are removed to exclude incomplete or fragmented entries generated during text splitting.

The cleaned corpus is converted into a document term matrix, where rows represent documents and columns represent term frequencies. Terms are filtered using frequency thresholds to reduce sparsity. The threshold is lower for the Taiwanese dataset and higher for the PRC dataset due to differences in corpus size. This helps reduce noise in the larger corpus while retaining sufficient term coverage in the smaller one. Documents with zero term counts are removed before modelling, since topic models cannot process empty documents.

Topic modelling is conducted using Latent Dirichlet Allocation. The number of topics is evaluated using CaoJuan2009 and Deveaud2014 metrics, and a three-topic model is selected to balance interpretability and comparability across groups. Topic labels are based on the highest probability terms and reflect observed word groupings rather than predefined categories.

The analysis proceeds in three steps. First, term frequency distributions are used to identify the most common words and compare usage across groups. Second, term associations are calculated to examine how key terms such as defence, security, provocation, and interference are connected to other words. Third, topic modelling is used to identify broader thematic patterns.

A supplementary sentiment analysis is conducted using the Bing lexicon. Sentiment scores are calculated for each document and compared across groups using summary statistics and distribution plots. This is treated as a descriptive measure because dictionary-based methods do not capture context well in political and security-related texts.

Results and Analysis

The results suggest that Taiwanese and PRC media frame United States-Taiwan arms sales differently, even though they share a common baseline vocabulary. Rather than simply reporting events, these patterns suggest different interpretive frameworks.

From the Taiwanese side, the document-term matrix contains 209 documents and 3,353 terms after filtering, with high sparsity (97%). The most frequent terms include “defence” (353), “relations” (343), “arms” (322), “military” (305), “government” (259), and “security” (216). These remain stable after applying a minimum frequency threshold, indicating a consistent core vocabulary. The language is centred on defence capacity, military capability, and policy.

The topic model identifies three topics with proportions of 0.388, 0.304, and 0.308. The first topic includes terms such as military, defence, government, security, and weapons, which point to defence capacity. The second includes arms, Taiwan, support, region, and strait, which reflects external support and regional context. The third includes the president, sale, Washington, policy, and Tsai, focusing on political decision-making and U.S.-Taiwan relations. The topics are relatively balanced and closely related, suggesting a consistent framing around defence, support, and policy.

Association results follow a similar pattern. “Defence” is linked with terms such as missile, air, fighter, capability, and spending, which are concrete military concepts. “Security” is connected to cooperation and deterrence-related terms. In contrast, “provocation” and “interference” appear less central and show weaker, more scattered associations. Overall, the Taiwanese corpus tends to present arms sales in the context of defence preparation and security management.

From the PRC’s side, the dataset includes 1,907 documents and 1,290 terms after filtering. The most frequent terms are “arms” (1711), “military” (1645), “relations” (1526), “nation” (1255), “country” (1252), “island” (1221), and “strait” (1197). Compared to the Taiwanese corpus, there is greater emphasis on geopolitical and territorial language.

The topic model also identifies three topics with proportions of 0.319, 0.347, and 0.333. The first topic includes military, nation, force, development, security, and people, which reflects a state-centred perspective. The second includes arms, sales, island, strait, independence, and foreign affairs, linking arms sales to sovereignty and territorial issues. The third includes relations, Washington, administration, Biden, president, and policy, focusing on U.S. political actors and external involvement. These topics place arms sales within a broader geopolitical context.

Association patterns are more tightly structured. “Defence” and “security” are strongly connected with terms such as combat, operations, training, command, and modernisation, with high correlation values. “Provocation” is associated with sovereignty, separatist, independence, and territory, while “interference” is linked with external, reunification, sovereignty, and cross-strait relations. These connections indicate a closer linkage between arms sales, political conflict, and sovereignty claims.

Comparing the two, both media systems use similar core terms such as arms, military, and security, but the way these terms are organised differs. In the Taiwanese corpus, they appear within a framework of defence capability and policy. In the PRC corpus, they are more closely tied to sovereignty, territorial integrity, and external involvement.

These patterns are consistent with the expectations in Sub-hypotheses 1a and 1b. PRC media more often connect arms sales to provocation, interference, and sovereignty-related language, while Taiwanese media emphasise defence and security. The difference appears not only in terms of frequency but also in how words are associated with each other.

The topic modelling results are also consistent with Sub-hypothesis 1c. Taiwanese topics focus on defence readiness, regional support, and policy processes, while PRC topics focus on national development, territorial issues, and U.S. involvement. The relatively balanced topic proportions in both corpora suggest that these patterns are not driven by a single dominant theme.

Sentiment analysis shows a different pattern. Across the full dataset, sentiment scores are close to neutral, with a mean of -0.72 and a median of -1. Most values fall within a narrow range. PRC articles are more frequently negative, with about 62.7% below zero, while Taiwanese articles show a more balanced distribution, with about 38.5% positive and 40.7% neutral. However, the magnitude of these values remains small, and most scores fall between 0 and -20.

The distributions reflect this. The PRC corpus shows a wider range and more dispersion, while the Taiwanese corpus is more concentrated around neutral values. This indicates some difference in distribution, but the overall tone remains relatively moderate in both groups. Taken together, the results suggest that differences between the two media systems are more visible in framing and topic structure than in sentiment.

Limitations and Suggestions

These results are limited by the method itself. The Bing lexicon is dictionary-based and does not account for context, which is a major issue for political and security-related text. Words associated with conflict, military activity, or risk are often coded as negative, even when used in neutral or descriptive contexts. As a result, the method may overrepresent negative classifications in content that simply discusses defence or geopolitical tensions. This is likely one reason why the PRC corpus appears more negative, since it contains more frequent references to conflict, sovereignty, and external threat.

The preprocessing steps also affect sentiment measurement. The use of stopword removal, stemming, and custom filtering may remove or alter words that carry sentiment, especially modifiers and context-dependent terms. This reduces the sentiment model’s ability to capture nuance and may compress the distribution toward neutrality.

More importantly, the dataset itself introduces structural limitations. All articles are English-language and drawn from internationally oriented outlets. This creates a translation effect. English-language reporting, especially from PRC state-affiliated media, is produced for a global audience and is typically more controlled, formal, and moderated in tone. Strong or emotionally charged language is less likely to appear because it reduces credibility in an international context. This directly constrains sentiment variation and explains why even the “negative” scores remain relatively mild.

Audience targeting also matters. The dataset consists exclusively of English-language reporting from outlets whose primary domestic audiences consume news in Mandarin Chinese. As a result, these texts are not primarily written for domestic audiences. They are designed for international audiences, enabling them to communicate positions to external readers, including policymakers, analysts, and international observers. This encourages a more neutral and strategic tone. In the PRC’s case, state-affiliated outlets may function less as independent journalism and more as official signalling, which further reduces expressive or emotional language.

Finally, the imbalance in dataset size reinforces these issues. The PRC corpus is much larger, increasing the likelihood of capturing more variation, including outliers, whereas the smaller Taiwanese corpus remains more compressed. This makes direct comparison of distributions less reliable.

Taken together, the sentiment analysis does not provide strong evidence for meaningful differences in tone. Instead, it shows that both media systems operate within a relatively narrow and controlled linguistic range in English-language reporting. The more useful insight is methodological: sentiment analysis using a dictionary-based approach is not well suited to capturing framing differences in this type of political text.

Future work should address these limitations by using original-language sources, especially Mandarin-language corpora, to reduce translation effects. Expanding the dataset to include more domestic outlets would improve representation and increase tone variation. More advanced methods, such as context-sensitive models or supervised classification, would also provide a more accurate measure of sentiment and framing in political discourse.

[1] Joseph O’Connor, “Taiwan Arms Sale Backlog, March 2026 Update – Abrams and Altius Delivered, but Further Delays Emerge,” Taiwan Security Monitor, April 14, 2026, https://tsm.schar.gmu.edu/taiwan-arms-sale-backlog-march-2026-update-abrams-and-altius-delivered-but-further-delays-emerge/.

[2] Michael Martina, Trevor Hunnicutt, Yimou Lee, and Ben Blanchard, “Exclusive: New US Weapons for Taiwan Could Be Approved after Trump’s China Trip, Sources Say,” Reuters, March 13, 2026, https://www.reuters.com/world/china/new-us-weapons-taiwan-could-be-approved-after-trumps-china-trip-sources-say-2026-03-13/.

[3] Reuters, “Taiwan Says Sale of Second Package of Arms from US Is Proceeding on Schedule,” Reuters, March 17, 2026, https://www.reuters.com/world/china/taiwan-says-it-has-received-no-information-us-delay-second-arms-sale-2026-03-17/.

[4] Haewoon Kwak, Jisun An, and Yong-Yeol Ahn, “A Systematic Media Frame Analysis of 1.5 Million New York Times Articles from 2000 to 2017,” in Proceedings of the 12th ACM Conference on Web Science (New York: Association for Computing Machinery, 2020), https://doi.org/10.1145/3394231.3397921.

[5] Weixin Zeng, “Reframing News by Different Agencies,” Babel 66, nos. 4–5 (2020): 847–66, https://doi.org/10.1075/babel.00172.zen.

[6] Steven W. Hook and Xiaoyu Pu, “Framing Sino-American Relations under Stress: A Reexamination of News Coverage of the 2001 Spy Plane Crisis,” Asian Affairs: An American Review 33, no. 3 (2006): 167–83, https://doi.org/10.3200/AAFS.33.3.167-183.

[7] Pablo Barberá, “Social Media, Echo Chambers, and Political Polarization,” in Social Media and Democracy: The State of the Field, Prospects for Reform, ed. Nathaniel Persily and Joshua A. Tucker (Cambridge: Cambridge University Press, 2020), 34–55, https://doi.org/10.1017/9781108890960.004.

[8] Ludovic Terren and Rosa Borge, “Echo Chambers on Social Media: A Systematic Review of the Literature,” Review of Communication Research 9 (2021): 99–118, https://doi.org/10.12840/ISSN.2255-4165.028.

[9] Ming-yeh T. Rawnsley, review of The Construction of National Identity in Taiwan’s Media, 1896–2012, by Chien-Jung Hsu, International Journal of Taiwan Studies 3, no. 2 (2020): 373–75, https://doi.org/10.1163/24688800-00302013.

[10] Chu Yi-wei, “Cross-Strait Identity Formation among Taiwanese Youth: Media Exposure, Political Socialization, and Civic Engagement in a Democratic Society,” International Journal of Science and Society 7, no. 2 (2025): 125–38, https://doi.org/10.54783/ijsoc.v7i2.1416.

[11] Charles W. Mahoney, “Deterrence, Dollars, or Diplomacy? Why the United States Sells Arms to Taiwan,” International Relations (published online August 2025), https://doi.org/10.1177/00471178251356841.