Advanced Natural Language Processing Analysis on Cross-Border Media Sentiment from China and South Korea

Introduction

Nationalism, a powerful and often polarizing force, has shaped the course of history and continues to influence global politics and social dynamics. The emotional and instinctual ties that bind individuals to their nations are not only born out of shared geography or language but also out of a collective sentiment that resonates with the group’s history and social experiences. As Handman (1921) insightfully observed, these sentiments are central to the development of nationalism, where beliefs and shared memories serve as the bedrock for collective identity.

In the arena of international relations, the interactions between neighboring countries provide a rich tapestry of cooperation and conflict (Schultz, 2015). These relationships are underpinned by a myriad of factors, including but not limited to historical legacies, economic dependencies, cultural affinities, and political alliances. The proximity of these nations often leads to intertwined destinies, where actions taken by one can have immediate and profound effects on the other. This interconnectedness necessitates a delicate balance, as periods of peace and collaboration can swiftly give way to tensions and disputes driven by changes in regional security landscapes, shifts in trade dynamics, or cultural misunderstandings.

Furthermore, the role of the media in shaping perceptions and sentiments across national borders cannot be overstated (Kokeyo, 2023). In an era where information is more accessible than ever, the media possesses an unparalleled ability to influence public opinion and national sentiment (Entman, 2004; Sobkowicz et al., 2012). The narratives propagated through news outlets, social media platforms, and cultural productions play a crucial role in constructing the image of the “other,” often serving as a lens through which domestic populations view their neighbors. As noted by Fuentes and Peterson (2021) and further explored by Alawade and Obun-Andy (2024), the media’s impact on international relations is profound, with the power to both bridge divides and deepen fissures between nations.

Media outlets often can reflect and amplify national attitudes towards foreign countries, influencing international relations (Berry et al., 2016; Nisbet & Myers, 2011). For instance, how Chinese and Korean media portray each other can affect public sentiment, potentially leading to changes in policy, consumer behavior, and even direct diplomatic relations. The portrayal of critical events, political decisions, and cultural exchanges in the media can either bridge gaps between the two nations or exacerbate misunderstandings and conflicts (e.g., Disputes over kimchi; Gries & Masui, 2022).

Given the media’s significant role, it is vital to study cross-border media sentiment to understand how it can reflect and influence the complex dynamics between China and Korea. Such an analysis can offer insights into the undercurrents of bilateral relations, providing a more nuanced understanding of how each country views the other and the implications of these perceptions for regional stability and cooperation.

Even though the nature and importance of relation measurement between two countries is scarce, this area of research deals with a comprehensive understanding of media sentiment at the country level. Media sentiment can significantly fluctuate in response to specific events, such as diplomatic meetings, military conflicts, cultural festivals, or trade agreements. Analyzing these fluctuations provides insights into the responsiveness of media narratives to international developments and their potential to influence public opinion and bilateral relations (e.g., Nagashima, 1970). Fisher et al. (2022) offers a unique perspective on analyzing media sentiment, particularly in the context of foreign policy and diplomatic relations. He collects and analyzes sentiment from hundreds of thousands of articles from state media and foreign affairs ministries of countries like North Korea, China, Russia, and Iran. This analysis provides insights into how governments attempt to influence international perceptions and the impact of media narratives on bilateral relations and public opinion. However, most of research are released in financial aspect (Kearney & Liu, 2014; Kim, 2023; Kim et al., 2023; Soon & Kim, 2023; Tetlock, 2007).

This study analyzes how the media can portray China and Korea, employing English text data and advanced Natural Language Processing (NLP) techniques. It aims to achieve the following four objectives: 1) measure sentiment, 2) identify common themes, 3) understand the relationship between the two countries, and 4) compare how historical events influence media coverage. The concepts-section provides a comprehensive overview of the research methodology and process, and the data section explains the collection of documents methods and general exploration; the measurement section presents the sentiment analysis results and discusses those results. The findings and ways ahead section highlights major contributions of this study and suggests directions for further study. Finally, the study concludes with a summary of its investigations and implications.

Concepts

The overviews and processes of study are divided into three key phases as follows:

• 1. Collections and Initial Analyses of Text Data: The initial phase involves extensive data collection through scraping news articles, reports, and other media documents related to China and Korea. Following the collection, the study applies NLP techniques to tokenize the titles of these documents and extracts preliminary information. This step sets the foundation for a comprehensive analysis by organizing and preparing the data for deeper examination.
• 2. Sentiment Analysis: The core of this study lies the sentiment analysis that is conducted by using advanced deep learning models, specifically transformer-based models like FinBERT-FOMC, which have been adapted for economic texts. This phase allows for an automated evaluation of the sentiments expressed in the media concerning the economic and geopolitical dynamics between China and Korea, offering insights into prevailing attitudes and perceptions.
• 3. Verification with Actual Events: The final phase involves comparing the sentiment analysis results with actual geopolitical and economic events between China and Korea. This comparison aims to validate the reliability of sentiment analysis techniques in accurately reflecting complex international relations and economic sentiments as reported in the media. This step ensures the findings are grounded in reality and enhances the study’s credibility by linking the analysis with the real world.

Data

The dataset for this study was obtained through a comprehensive scraping of the ProQuest database, incorporating a broad array of variables such as publication date, title, abstract, document type, publishing company, publisher, and place of publication. And, we focus on a subset of the following variables: ‘Publication Date’, ‘Title’, ‘Document Type’, ‘Publisher’, and ‘Company’, to conduct a textual mood analysis across two countries. <Figure 1> illustrates the distribution of document types within the dataset. Most of the data was classified as newspaper articles, followed by wire feeds, magazines, and reports. It is important to note that in this study, the weight attributed to each document type was not considered. This decision was made despite the potential for reports or scholarly journals to exert a relatively higher impact on the market.

Figure 1.

Document Types of Raw DataNote: Data has been downloaded in ProQuest Central Database (https://www.proquest.com/)

<Figure 2> displays the top 20 publishers within the dataset, with ‘Asia News Monitor’ emerging as the foremost distributor, succeeded by ‘The Korea Times’ and ‘China Daily’. Moreover, <Figure 3> delineates the top 20 organizations or institutions classified as companies in the dataset, with NATO and ASEAN occupying the leading positions.

Figure 2.

Top20 Publisher by FrequencyNote: Data has been downloaded in ProQuest Central Database (https://www.proquest.com/).

Figure 3.

Top20 Publishing Companies by FrequencyNote: Data has been downloaded in ProQuest Central Database (https://www.proquest.com/).

<Figure 4> illustrates the daily frequency of documents, presenting an overview of the data retrieved from the ProQuest database using the keywords ‘China’, ‘Chinese’, ‘Korea’, and ‘Korean’. A total of 11,598 documents were identified. However, it is noted that some documents may only mention these countries tangentially, potentially introducing noise into the analysis. The second and third plots focus specifically on documents with titles containing the keywords “China” and “Korea” respectively, aiming to refine the dataset to more relevant content concerning the two countries. Additionally, the fourth plot highlights the intersection of occurrences where both China and Korea are mentioned, with April and October emerging as months with notably high activity in the dataset.

Figure 4.

Frequency of Document Entries in 2023Note: The dataset consists of 11,598 documents. Out of these, 3,469 articles mention “Korea”, while 4,664 articles reference “China” in their titles. Additionally, 1,535 articles contain references to both “China” and “Korea” in their titles.

In our analysis, we opt to focus on documents whose titles specifically mentioned ‘China’ or ‘Korea’. <Figure 5> presents a graph of 6,598 documents deemed suitable for analysis. The decision is based upon the premise that rely solely on documents featuring keywords for only one country could result in skewed information, while exclusively considering documents with cooccurring keywords for both countries might be overly restrictive. Consequently, we aggregate the data to encompass both criteria.

Figure 5.

Frequency of Document Entries for China or KoreaNote: In the dataset, there are 6,598 articles that mention either “China” or “Korea” in their titles.

Measurement

Initially, we employed natural language processing (NLP) techniques to extract fundamental information from the documents. Tokenization at the word level, while straightforward, proved to be a valuable method for applying economic analysis. Subsequently, we utilize deep learning techniques, specifically transformer-based models (Vaswani et al., 2017), to conduct sentiment analysis. This approach allow us to automatically gauge the economic sentiment between the two countries, offering insights into the underlying perceptions and attitudes present in the textual data.

Natural Language Processing

The document titles have been tokenized to separate them into individual words. It means that a title such as “China slams IAEA’s nod for Japan’s plan” has been broken down into “China,” “slams,” “IAEA,” “nod,” “Japan,” and “plan.” And then, each title has been split into keywords, and the monthly frequency of each keyword has been calculated. This provides useful information about the frequency of certain topics. In Table 1, the top 10 keywords for each month in 2023 are listed with their corresponding frequency values. For example, in January ‘23’, “Russia” appeared 9 times, “trade” appeared 7 times, and “Foreign” appeared 7 times. The note at the bottom of the table explains that common terms such as “mr.,” “said,” “would,” “one,” “could,” “new,” “China,” “Korea,” “says,” and other similar generic keywords have been excluded manually. It is important because it ensures that the frequency count reflects more meaningful data by removing less informative words. By analyzing the table, we can observe that some keywords are consistently prominent across several months (e.g., “Russia,” “trade,” “Biden”), which may indicate ongoing themes or issues relevant during that time. Other keywords may reflect events or discussions specific to a particular month (e.g., “Antidumping” in May ‘23’).

Table 1.

Top10 reference of Keywords by Monthly

The presence and frequency of certain keywords can provide insights into the geopolitical or economic discourse during the year 2023, and this table is likely a part of a larger study aiming to analyze trends, sentiments, or topics of interest in relation to the keywords and the context in which they were used.

Sentiment Analysis with Deep Learning

Our approach to deepening the textual analysis for discerning the mood surrounding bilateral issues is indeed methodologically sound. Economic texts often present a challenge for sentiment analysis mainly due to their complexity and the specialized language used, which typical pretrained models may not accurately interpret. As highlighted by Kim et al. (2023), the intricacies involved in economic texts necessitate more than a general understanding that pre-trained models provide.

Liu et al. (2021) and Gössi et al. (2023) have addressed this challenge by updating and refining the BERT language model specifically for economic texts, resulting in what you refer to as the FinBERT-FOMC model. BERT, or Bidirectional Encoder Representations from Transformers, is a transformer-based machine learning technique for natural language processing pre-training. The original BERT model is trained on a large corpus of text and then fine-tuned for specific tasks. However, economic texts require a model that is fine-tuned on a relevant corpus that captures the nuances of economic language and sentiment.

The FinBERT-FOMC model the previous studies applied has been specifically fine-tuned for complex economic texts, which likely includes Federal Open Market Committee (FOMC) statements or similar materials. The FOMC statements are known for their impact on financial markets and contain rich economic language that can be essential for a model that aims to analyze economic sentiment.

Using the FinBERT-FOMC model, the sentiment analysis can more accurately capture the subtleties of economic discourse, allowing for a nuanced understanding of the sentiment in economic texts regarding China and Korea. By applying this model, we can determine whether the sentiment is positive, negative, or neutral, which can be particularly useful for gauging market sentiment or the overall mood of bilateral economic relations.

<Figure 6> depicts the weekly average sentiment scores for news data in the year 2023, comparing two sets of results obtained from sentiment analysis using the FinBERT model. Dashed line This line represents the sentiment scores obtained from the original application of the FinBERT model (Gössi et al., 2023) on the dataset. Solid Line (Adjusted FinBERT Sentiment) shows the sentiment scores after adjusting the original FinBERT results by stressing the influence of certain negative words, specifically ‘war’, ‘nuclear’, ‘military’, and ‘Taiwan’.

Figure 6.

Weekly Sentiment Trends in 2023Note: The adjusted FinBERT sentiment scores are calculated by deducting the impact of negative words by ‘war’, ‘nuclear’, ‘military’, and ‘Taiwan’. Consequently, these adjusted scores are lower than the original FinBERT sentiment values. The shaded red areas seem to highlight periods of volatility or significant changes in sentiment regarding political issues.

The graph shows fluctuations in sentiment throughout the year with both lines following a somewhat similar pattern—this indicates that overall trends in sentiment are preserved even after the adjustment for negative words. However, the adjusted sentiment line is consistently lower than the original sentiment line, which means that once the identified negative terms are discounted, the overall sentiment becomes more negative on average. The observation by Kim (2023) that a negativity-focused sentiment score may yield better predictive power of economic conditions is noteworthy. This perspective is grounded in the theory that negative news can have a disproportionate impact on market sentiment and investor behavior, often resulting in more significant and rapid market reactions compared to positive news. This phenomenon is sometimes referred to as the “negativity bias” in the field of behavioral economics, where negative events or emotions have a greater effect on an individual’s psychological state and processes than neutral or positive things (Kim, 2023).

When applying sentiment analysis, especially in the context of economic forecasting or market analysis, incorporating a weighting scheme that emphasizes negative sentiment could therefore improve the model’s ability to forecast economic downturns or market volatility. This approach would be particularly relevant if historical data supports the premise that markets respond more vigorously to negative news.

However, it is also important to consider that an overemphasis on negative events can lead to a skewed perspective if not balanced by an understanding of positive developments and their potential impact. A comprehensive sentiment analysis should consider the full spectrum of news sentiment to provide a balanced view of market dynamics.

Inference

When discussing the main issues reflected by a sentiment score, especially in the context of economic relations between two countries, it is important to consider various dimensions that can be captured by news sentiment. These dimensions often include changed in political events, economic policies, trade relations, and other significant developments. Here’s how these issues can be reflected in a sentiment score (See Table 2):

Table 2.

Main issue between China and Korea in 2023

Findings and ways ahead

Conclusion

This study highlights the importance of Natural Language Processing (NLP) and sentiment analysis in interpreting media sentiment, especially regarding China and Korea. The study reveals complex sentiment patterns that reflect the geopolitical and economic dynamics between the two nations by examining datasets. Using transformer-based models for sentiment analysis represents a significant advancement, allowing for more nuanced sentiment detection.

The correlation between sentiment analysis results and actual geopolitical events proves the effectiveness of NLP methodologies and underscores their potential for predictive analytics in international relations. This research suggests that media sentiment could serve as a leading indicator for shifts in geopolitical stances and economic policies. This opens new avenues to understanding the impact of media narratives on public perception and foreign policy.

The study recommends an integrated approach to media analysis, combining quantitative NLP techniques with qualitative assessments to accurately capture national sentiment and its implications. As geopolitical landscapes change, analyzing and interpreting media sentiment becomes crucial for policymakers, analysts, and scholars to anticipate and respond to international developments.

In conclusion, this research provides valuable insights into media sentiment dynamics towards China and Korea, showcasing the transformative potential of NLP and sentiment analysis in international relations research. Applying these methodologies promises to deepen our understanding of global media narratives, offering a potent tool for navigating the complexities of international diplomacy and economic relations.

AI Acknowledgment

Generative AI or AI-assisted technologies were not used in any way to prepare, write, or complete essential authoring tasks in this manuscript.

References

• Alawade, S. O., & Obun-Andy, M. K. (2024). The role of media in shaping public perception of climate change. Journal of Research and Innovation in Social Science, 8(1), 2442–2448. [https://doi.org/10.47772/IJRISS.2024.801179]
• Berry, M., Garcia-Blanco, I., & Moore, K. (2016). Press coverage of the refugee and migrant crisis in the EU: A content analysis of five European countries. United Nations High Commissioner for Refugees.
• Entman, R. M. (2004). Projections of power: Framing news, public opinion, and US foreign policy. University of Chicago Press.
• Fisher, S., Klein, G. R., & Codjo, J. (2022). Focusdata: Foreign policy through language and sentiment. Foreign Policy Analysis, 18(2), orac002.
• Fuentes, A., & Peterson, J. V. (2021). Social media and public perception as core aspect of public health: The cautionary case of @realdonaldtrump and COVID-19. PLoS ONE, 16(5), e0251179. [https://doi.org/10.1371/journal.pone.0251179]
• Gössi, S., Chen, Z., Kim, W., Bermeitinger, B., & Handschuh, S. (2023). FinBERT-FOMC: Fine-Tuned FinBERT Model with sentiment focus method for enhancing sentiment analysis of FOMC minutes. Proceedings of the Fourth ACM International Conference on AI in Finance, 357–364.
• Gries, P., & Masui, Y. (2022). How history wars shape foreign policy: an ancient kingdom and the future of China–South Korea relations. Journal of East Asian Studies, 22(1), 1–21.
• Handman, M. S. (1921). The sentiment of nationalism. Political Science Quarterly, 36(1), 104–121.
• Kearney, C., & Liu, S. (2014). Textual sentiment in finance: A survey of methods and models. International Review of Financial Analysis, 33, 171–185.
• Kim, W. (2023). Words that matter: The impact of negative words on news sentiment and stock market index. [https://doi.org/10.48550/arXiv.2304.00468]
• Kim, W., Spörer, J. F., & Handschuh, S. (2023). Analyzing FOMC minutes: Accuracy and constraints of language models. [https://doi.org/10.48550/arXiv.2304.10164]
• Kim, W., Spörer, J. F., & Handschuh, S. (2023). Words that Matter: The Impact of Negative Words on News Sentiment and Stock Market Index. [https://doi.org/10.48550/arXiv.2304.00468]
• Kokeyo, A. (2023). Exploring the dynamics of social media in shaping narratives and perceptions in the Israeli- Palestinian conflict: preliminary reflections. African Journal of Emerging Issues, 5(17), 181–194.
• Liu, Z., Huang, D., Huang, K., Li, Z., & Zhao, J. (2021). Finbert: A pre-trained financial language representation model for financial text mining. Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, 4513–4519.
• Nagashima, A. (1970). A comparison of Japanese and US attitudes toward foreign products. Journal of Marketing, 34(1), 68–74.
• Nisbet, E. C., & Myers, T. A. (2011). Anti-American sentiment as a media effect? Arab media, political identity, and public opinion in the Middle East. Communication Research, 38(5), 684–709.
• Schultz, K. A. (2015). Borders, conflict, and trade. Annual Review of Political Science, 18, 125–145.
• Sobkowicz, P., Kaschesky, M., & Bouchard, G. (2012). Opinion mining in social media: Modeling, simulating, and forecasting political opinions in the web. Government information quarterly, 29(4), 470–479.
• Soon, B. M., & Kim, W. (2023). The impact of African swine fever news sentiment on the Korean meat market. Plos ONE, 18(6), e0286520.
• Tetlock, P. C. (2007). Giving content to investor sentiment: The role of media in the stock market. The Journal of finance, 62(3), 1139–1168.
• Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30.