보수 언론과 진보 언론의 북한 전문가 활용 방식의 차이 탐색 : 인용문에 대한 KPF-BERT 기반 딥러닝 분석을 중심으로
초록
본 연구는 언론이 정치성향에 따라 북한 전문가 인용 보도에서 편향성을 나타내는 실태를 정파적 지표화(partisan indexing) 현상과 딥러닝 모형의 적용을 통해 살펴보았다. 언론의 지표화 가설(press indexing hypothesis)에 따르면, 언론은 공적 정보원의 영향을 받아 보도 범위를 설정한다. 반면, 보수와 진보 언론이 자신의 정치적 성향에 맞춰 정보원과 인용 내용을 선택적으로 활용하는 현상은 정파적 지표화라고 할 수 있다. 본 연구진은 북한 보도에서 이런 현상을 탐색하고자 북한 관련 메시지를 보수와 진보 성향으로 분류하는 딥러닝 모형을 KPF-BERT를 기반으로 개발했다. 이어서 모형을 활용해 빅카인즈에서 제공하는 북한 전문가 80인의 인용문 42,375건(2011년~2021년)에 대해 정치성향 지수를 측정했다. 분석 결과, 북한 전문가들은 보수 성향에서 진보 성향에 이르기까지 다양하게 분포했으며, 극단적 정치성향에서는 진보 전문가보다 보수 전문가가 많은 것으로 나타났다. 언론사 분포에서는 보수 언론사 3곳(조선일보, 중앙일보, 동아일보)이 분석 대상 52곳 가운데 보수 성향 순서로 2, 4, 5위를 기록했다. 한편, 진보 언론사 한겨레신문과 경향신문은 중도에 가까운 위치를 나타냈고, 내일신문만 뚜렷하게 진보 성향을 보였다. 핵심 분석 결과는 보수와 진보 언론이 자신의 성향과 유사한 전문가를 더 자주 인용하고, 동일 전문가에 대해 자신의 성향과 유사한 내용을 더 많이 선택해 보도했다는 점이다. 북한 보도에서 ‘정파적 지표화(partisan indexing)’ 현상이 유의미하게 검증된 것이다.
Abstract
This study examined whether conservative and progressive Korean media exhibits bias in the news coverage quoting North Korean experts. According to the press indexing hypothesis, the news media are influenced by public sources to set coverage limits. On the other hand, the partisan indexing phenomenon refers to the case in which conservative and progressive media selectively use sources of information and quotes according to their political orientation. To explore this phenomenon in the news coverage of North Korea, this study developed a deep-learning model that classifies news editorials about North Korea into conservative and progressive messages based on the KPF-BERT model. This model was constructed by learning news articles accumulated for 20 years by the Korea Press Foundation’s Big-Kinds archive system, and showed excellent performance in natural language processes compared to other Korean BERT models. The KPF-BERT-based deep-learning model developed by this study finally recorded an accuracy of 93.77%. Next, this model, measured the political orientation index for 42,375 quotes (from 2011 to 2021) of 80 North Korean experts provided by BIG-KINDS. As a result, North Korean experts were widely dispersed from conservative to progressive positions, and some conservative experts were found to be in more extreme positions than progressive ones. This implies that extreme conservative opinions are allowed in the Korean media, but extreme progressive opinions are hardly reflected. In terms of the media distribution, three conservative media (Chosun Ilbo, JoongAng Ilbo, and Dong-A Ilbo) ranked 2nd, 4th, and 5th in the order of conservative political orientation among the 52 media. On the other hand, two progressive media such as Hankyoreh and Kyunghyang Shinmun were placed at a near-central position, and only the Naeil Shinmun clearly showed progressive orientation. This implies that the intensity of political orientation should be considered beyond the distinction between conservative and progressive media in the media analysis. A key point of this analysis is that conservative and progressive media tend to selectively quote experts similar to their political orientation and selectively report content similar to their orientation in the same expert. The phenomenon of partisan indexing was significantly observed in the news on North Korea. It is worrying that the political bias in Korean media is reflected in experts’ quotes beyond the news staff's writings. Methodologically, the application of the KPF-BERT-based deep-learning model is a suitable method to analyze tons of news data accumulated every day. Hopefully, this method can be widely and actively used in future journalism research.
Keywords:
North Korean news, experts, partisan indexing, KPF-BERT, deep-learning키워드:
북한 보도, 전문가 인용, 정파적 지표화, 딥러닝Acknowledgments
This work was supported by Press Promotion Fund of the Korea Press Foundation in 2022. 이 연구는 2022년도 한국언론진흥재단 언론진흥기금의 지원으로 이루어진 것임.
References
- Bang, E., & Kim, S. (2012). A study of SNS sources in Korean major newspapers. Journal of Cybercommunication Academic Society, 29(4), 145-189.
- Bennett, W. L. (1990). Toward a theory of press-state relations in the United States. Journal of Communication, 40(2), 103-125. [https://doi.org/10.1111/j.1460-2466.1990.tb02265.x]
- Choi, H. (2010). A study on the diversity of Korean newspapers: Analyzing the tendencies of covering three major issues. Korean Journal of Journalism & Communication Studies, 54(3), 314-348.
- Choi, J. (2012). Political empowerment of media after democratization in Korea: Focusing on the analysis of the major newspapers' articles during the 13-17th presidential election campaigns. Doctoral dissertation at Hanyang University, Seoul, Korea.
- Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2018). Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv:1810.04805, .
- Ha, S., & Lee, M. (2012). A news frame analysis by the South Korean press on the livelihoods of North Koreans. Korean Journal of Communication & Information, 58, 222-241.
- Hall, S., Critcher, C., Jefferson, T., Clarke, J., & Roberts, B. (1978). Policing the crisis: Mugging, the state and law and order. London, UK: Macmillan. [https://doi.org/10.1007/978-1-349-15881-2]
- Han, D., & Lim, J. (2001). A study on the use of news sources in the media and the struggle for hegemony: Focusing on the analysis of news sources in North Korean reports. Paper presented at Korean Society for Journalism & Communication Studies.
- Kim, K. (2019). Kim Ki-Hyun's natural language processing and deep learning camp (PyTorch edition): From basic to advanced deep learning-based natural language processing. Seoul: Hanbitmedia.
- Kim, K., & Noh, G. (2011). A comparative study of news reporting about North Korea on newspapers in South Korea. Korean Journal of Journalism & Communication Studies, 55(1), 362-498.
- Kim, Y. (2007). Source competition and dependency on issue attributes: Issue competition between the government and the activists on the issue of screen quota. Korean Journal of Communication & Information, 39, 140-177.
- Kiousis, S., Mitrook, M., Wu, X., & Seltzer, T. (2006). First-and second-level agenda-building and agenda-setting effects: Exploring the linkages among candidate news releases, media coverage, and public opinion during the 2002 Florida gubernatorial election. Journal of Public Relations Research, 18(3), 265-285. [https://doi.org/10.1207/s1532754xjprr1803_4]
- Lee, J. (2003). A political economy study on North Korea-related news coverage: An editorial analysis of the North Korean nuclear issue (1993-2003) by Chosun-Ilbo and Hankyoreh. Paper presented at Korean Society for Journalism & Communication Studies.
- Lee, K. (2021). Do it! Learning natural language processing with BERT and GPT: Core principles of transformers and how to use the huggingface package. Seoul: Easyspublishing.
- Lee, R., Park, E., & Lee, J. (2020). Comparison of variant BERT models on Korean. Proceedings of Korean Institute of Commucations and Information Sciences, 901-902.
- Lee, W., & On, B. (2021, June). Emotion word masking method for emotional classification of BERT model. Proceedings of Korean Institute of Information Technology, 413-416.
- Lee, W., & Son, Y. (2011). The media’s agenda setting on the nuclear test of North Korea. Korean Journal of Communication & Information, 56, 175-193.
- Özçift, A., Akarsu, K., Yumuk, F., & Söylemez, C. (2021). Advancing natural language processing (NLP) applications of morphologically rich languages with bidirectional encoder representations from transformers (BERT): An empirical case study for Turkish. Automatika, 62(2), 226-238. [https://doi.org/10.1080/00051144.2021.1922150]
- Pappagari, R., Zelasko, P., Villalba, J., Carmiel, Y., & Dehak, N. (2019, December). Hierarchical transformers for long document classification. Proceedings of 2019 IEEE Automatic Speech Recognition and Understanding Workshop(ASRU), 838-844. [https://doi.org/10.1109/ASRU46091.2019.9003958]
- Park, D (2014). Biases by bursts of quoting sources in Journalism. Communication Theories, 10(1), 295-324.
- Seol, J., & Nam, G. (2007). A study on the relationship between television news and diplomacy : A content analysis of Six Party Talks news. Korean Journal of Broadcasting and Telecommunication Studies, 21(4), 69-100.
- Shoemaker, P. J., & Reese, S. D. (1996). Mediating the message: Theories of influences on mass medai content. New York, NY: Longman.
- Sim, J. (2021). A study on automatic classification of orofanity sentences of elementary school students using BERT. Journal of Creative Information Culture, 7(2), 91-98.
- Song, Y. (2005). A study on the rhetorical devices to enhance the factuality in the news of the society pages in Korean national newspapers. Korean Journal of Journalism & Communication Studies, 49(3), 80-104.
- Usuga-Cadavid, J. P., Lamouri, S., Grabot, B., & Fortin, A. (2022). Using deep learning to value free-form text data for predictive maintenance. International Journal of Production Research, 60(14), 4548-4575. [https://doi.org/10.1080/00207543.2021.1951868]
- van Dijk, T. A. (1990). News as discourse. Hillsdale, NJ: Lawrence Erlbaum.
- Yoo, W., & An, S. (2022). Introduction to natural language processing using deep learning. WikiDocs. Retrieved from https://wikidocs.net/109251, .
- Yoon, Y. (1997). Changes in inter-korean relations and media: Comparing reporting times of North and South Korean newspapers. In Y. C. Yoon, Y. S. Kim, & Y. S. Choi (Eds.), International Communication and Korean Society (pp. 121-152). Seoul: Nanam.