In recent years, research on China’s image in the international mainstream media has received increasing attention from academics. This research has mainly focused on five aspects. The first aspect has researched the cognition of China’s image in the international mainstream media during a certain period. Zhang Ying (2011) analyzed the New York Times’ construction of China’s national image during the Obama administration, and concluded that the New York Times has been more objective in its coverage aiming at China since the Obama administration. In a study examining the coverage of China in the New York Times and Los Angeles Times between 1992 and 2001, Peng (2004) found that while the coverage of China increased significantly over time, the overall tone remained negative. The study also found that stories presented in political frames and ideological frames were more likely to be unfavorable and no significant differences were found between the two newspapers. He and Chen (2012) analyzed Newsweek’s reports on China from 2009 to 2010 and found a tendency of political generalization in its China image construction. The second area of focus is the study on the tendency of China image of mainstream media in a certain geographical area. Zhang and Cameron (2003) found that American mainstream newspapers tend to portray a negative image of China by analyzing the China-related reports of The New York Times, The Washington Post and The Los Angeles Times. As the United States is the largest dominant player in international public opinion, the tendency of China’s image in multiple domains (positive-neutral-negative) and China’s national image as a multi-ethnic country presented in various mainstream media reports have attracted scholars’ constant attention (Zhou and Zheng, 2010; Gao and Jia, 2016, among others). Lan and Luo (2013) analyzed the British mainstream media’s construction of China’s image, pointing out that their reports on economic development are more objective and neutral, while their reports on politics or emergencies, as well as “human rights issues” are ideologically biased and one-sided. Zhang and Wu (2017) utilized critical discourse analysis to examine the representation of China in English-language newspapers in China versus the UK. Their study revealed that China Daily emphasized the positive attributes of the One Belt, One Road Initiative, portraying China as a peace-loving nation and an emerging global economic power, while Financial Times presented mixed and conflicting images of China, depicting it as both having a significant impact on the global economy and as an authoritarian state and geopolitical threat. Huan (2023) examined the representation of China in Australian media discourses and found that the conventional binary perception of China–Australia relations as either an opportunity or a threat is oversimplified. According to the study by Huan and Deng (2021), the portrayal of China in South Africa's mainstream English-language newspapers transcended the traditional and often oversimplified dichotomy of partner or predator, recognizing the complexities, contradictions, and changing dynamics of Sino-SA relations. Li (2017) summarized the ways and characteristics of the Japanese mainstream media’s portrayal of China and its tendency to maliciously distort and negatively focus on China. Shen and Wu (2013) analyzed the reports on China by three regional mainstream media in Russia and found that the image of China reported by them was somewhat distant from the actual image of China and what was described in the joint statements/communiqués of the two countries. The third area is the study of China’s image shaping in international mainstream media in a certain field. Wang (2009) and scholars such as Wang and Han (2010) analyzed the image of China’s economy in overseas mainstream media by focusing specifically on economic reports related to China’s industries, China’s manufacturing, and the three rural issues. Xia (2012) interpreted the backward and negative image of China in the Western media by analyzing the New York Times’ coverage of the Forbidden City theft case. Pan and Dong (2017) used reports related to the 2016 China-Russia joint military exercises to scrutinize the negative Chinese military image constructed by the U.S. mainstream news media. Tang (2018) analyzed the portrayal of China in three US mainstream newspapers between 2008 and 2010, and revealed a predominantly negative representation. This representation is characterized by participant roles such as the Persuaded, the Criticized, the Labeled, the Contained, the Punished, the Helped, and the Praised, collectively contributing to a negative thematic emphasis, particularly in the domains of economy and trade. Wang (2018) conducted a corpus-assisted critical discourse analysis of news reports on air pollution published from 2008 through 2015 by China Daily. The study emphasized the submissive role of the official Chinese press under the CPC's strict censorship system and the role of international sporting events, such as the 2008 Beijing Olympic Games and the 2010 Guangzhou Asian Games, in raising awareness of environmental issues among both the public and the government. Teo and Xu (2021) conducted a critical discourse analysis comparing the portrayal of China’s Belt and Road Initiative (BRI) in China Daily (CD) and New York Times (NYT). They uncovered divergent discourses between the two newspapers, with CD depicting the BRI as a collaborative project aimed at unifying and benefiting member countries, while NYT constructed the BRI as a geopolitical threat to the waning global influence of the US. Fourth, in a study of the causes of China’s image shaping in international mainstream media, Zhang and Chen (2014) examined the political and cultural factors that interfered with the construction of China’s image by British and American broadsheets in their coverage of the geopolitical conflict—Diaoyu Islands incident. Xu and Wang (2016) focused on the New York Times’ ten-year China-related reports and analyzed the changes of China’s image in the discourse construction of the western media in recent years and the causes behind it. Fifth, research has also taken a linguistic perspective on the study of China’s image in international mainstream media. Zhong (2013) used critical discourse analysis theory to study the three-dimensional negative image of China constructed by the intertextuality of “Ye Shiwen’s discourse” in British mainstream newspapers. Chen and Chen (2017) used corpus linguistics to analyze the changes of China's image in the eyes of Western media with news discourse from two different time periods of The Guardian. Hu and Li (2022) conducted a corpus-based analysis comparing English translations of the Report on the Work of the Government with the State of the Union Address. Additionally, studies from a linguistic perspective have contributed to our understanding of China’s image in international mainstream media. Yang and Van Grop (2023) conducted an extensive analysis of diverging interpretations of the Belt and Road Initiative (BRI) through frames, shedding light on core claims constituting China’s discursive legitimation of the BRI, the differences between Chinese officials and foreign political-media elites, the continuity or change in foreign governments' positions on the BRI, and the increasing critical coverage by foreign elite media.

Tsirakis et al. (2017) proposed a platform for real-time opinion mining, named “PaloPro,” to address the challenges faced by companies in processing huge volumes of streamed data from social media and other sources. This platform utilized prior research and real application services to meet the needs for brand monitoring in near real-time and providing insights on data in motion. Hartmann et al. (2018) demonstrated that compared to traditional Support Vector Machines (SVM) and Linguistic Inquiry and Word Count (LIWC), Random Forest (RF) and Naive Bayes (NB) exhibited superior performance in automatically classifying social media text, particularly in three-class sentiment classification and with small sample sizes. In the context of evaluating China's national image across various international media platforms, the methodologies and insights provided by Nandwani and Verma (2021) on sentiment analysis and emotion detection offered a promising avenue for deepening our understanding of the underlying attitudes and sentiments conveyed in media texts. Their work highlighted the capability of sentiment analysis to discern the polarity of texts—identifying whether the sentiments are positive, negative, or neutral towards subjects such as nations, administrations, or individuals. This approach could potentially augment current research methodologies by providing quantitative data on the sentiment bias in media coverage of China, thus offering a nuanced perspective on its image construction. Furthermore, the application of advanced analytical methodologies, such as the automated public opinion monitoring mechanism described by Karamouzas, Mademlis, and Pitas (2022), which utilizes Natural Language Processing algorithms to quantify text polarity, offensiveness, bias, and figurativeness in social media content, could significantly enrich the current research landscape on China's image in international media. The study by Chaudhary et al. (2021) demonstrated the use of big data analytics to predict consumer behavior on social media platforms by analyzing diverse and high-volume data collected from platforms such as Facebook, Twitter, LinkedIn, YouTube, Instagram, and Pinterest. Ramya and Sivakumar (2021) proposed a comprehensive approach involving sentiment analysis and the identification of influential users in social media platforms. By utilizing techniques like weighted partition around medoids with artificial cooperative search (WPAM-ACS) and fuzzy deep neural network (FDNN), they addressed challenges in sentiment classification, especially in languages lacking adequate resources for NLP models. Additionally, the survey conducted by Adak et al. (2022) comprehensively reviewed various aspects related to the infosphere, including collaborative systems like Wikipedia, the role of the infosphere in facilitating scientific citations and interdisciplinary research, and challenges related to governance, such as addressing rising hateful behavior, abusive behavior, bias, and discrimination in online platforms and news reporting. Although research on China’s image in the international mainstream media have generated insights, there are still several shortcomings. First, the existing research covers a short time span, and there is a lack of dynamic monitoring and analysis of China’s image. The limited time span has broken the continuity of the diachronic evolution of the research on China’s image in the international mainstream media, making it impossible to grasp the dynamic changes in China’s image in real time. Second, the media and regional scope that have been studied are limited, and most studies have focused on the mainstream media in several developed countries in Europe and the United States. There is little research on important media in Germany, Russia, Japan and other countries. The voice of mainstream media in Asian and African countries is ignored. In particular, the mainstream media lack the recognition of China’s image in developing countries and are less representative due to the imbalance of the media’s regional choice. Third, the existing research focuses on topics in the political and economic fields with little attention paid to other topics. With the enhancement of China’s role as a responsible great power, it is difficult to present the complex composition and evolution of China’s image from a single dimension. With respect to China’s image, research must be extended to other dimensions, such as society, culture, military, diplomacy, climate and geography. Fourth, most of the existing studies focus on analyzing the causes at the macro level, with few targeted strategy suggestions. Most of the studies describe only the image of China that is presented by international mainstream media and analyze the macro causes, such as ideology, cultural background and international relations. These studies lack operational and targeted strategy suggestions for improving China’s international image. Fifth, the interdisciplinary and quantitative level of existing research is insufficient, and the scientificity of research methods must be improved. The disciplines involved in the research include politics, communication and linguistics. However, communication and linguistics are not sufficiently connected in terms of disciplinary theories and methods, thus greatly limiting the breadth and depth of the study. Wang (2016) proposed to develop a large diachronic corpus to compensate the deficiency of such research by providing constantly increasing objective descriptions and analyzable data.

In conclusion, the methods and tools used for researching China’s image in international mainstream media should be further enriched. The country and regional scope of media sources and the related disciplines should be expanded to include comprehensive linguistics, communication, sociology, political science, journalism, and statistics. A comprehensive analysis of qualitative and quantitative, diachronic and synchronic, language and framework methods with a long-time span should be carried out based on big data. However, the prerequisite for achieving this goal is the construction of a dynamic monitoring and analysis platform for the international mainstream media corpus on China’s image. With hot words as the search object and key time points as the horizon, this platform is technically characterized by dynamic monitoring and continuous expansion of the corpus. This platform can deepen the previous research on the causes of China’s image-building in the international mainstream media, the analysis of China’s image in the international media, and the combination of China-related public sentiment and opinion. Additionally, this platform can provide dynamic and quantitative information support for timely grasping the focuses and diachronic changes and comprehensively understanding the international mainstream media corpus on China’s image in the future. Since the construction of a dynamic monitoring platform for the international mainstream media corpus on China’s image is a highly technical work, the platform development principles, operating environment and how to realize media corpus crawling, automatic storage and corpus analysis during platform construction were introduced from the technical perspective in the following section.

According to the practical, rational, secure, extensible, maintainable and open principles for platform construction, the Python developing language was adopted under the Model-View-Template (MVT) architecture to build the operating framework of Django (Huang, 2019). For hardware operation, a 1.8 GHz Intel Core 2 or higher equivalent Advanced Micro Devices (AMD) processor was required, whose memory capacity was Windows 7:2 GB or Windows 10:4 GB, and hard disk capacity was above 15 GB. For software operation, PyCharm was used as the developing platform, MySQL was used to provide back-end service support, Redis and MongoDB as the nonrelational database, native HTML5 to design front-end pages and ECharts technology to render data.

The main function of the system is to automatically collect data on news websites through crawlers, generate corpus data in a standardized format and store them in the database. Many technologies were involved in generating the corpus, and the data from the news websites were presented mainly in the HyperText Markup Language (HTML) in the browser. Generally, the program simulated a browser to capture the HTML structure corpus of news websites while crawling news data. Upon obtaining the original website data, the original website should be analyzed manually. Each website adopts different technologies due to the complex original website data and the inconsistent format. For the simplest news websites, the data are displayed on the websites via HTML tags, while the data will be transmitted encrypted for some websites that adopt anticrawling measures. No matter how the data were crawled, as long as the information sources were different, the crawled data needed to be further processed. Conventional processing is to analyze the website format and extract the data content on the websites. However, data extracted by such a method usually contain special characters, hyperlinks and tags; consequently, the original corpus must be cleaned. In the automatic corpus cleaning stage, two main links were designed (Figure 9).

From the perspective of the topic model, the evolution of dimension of China’s image and the concentration of each dimension in the China-related coverage in 10 international mainstream media platforms, including The New York Times (USA), The Times (UK) and Asahi Shimbun (JP), in the past 10 years were analyzed. The topic mining effect of each topic in the past ten years and the changes in topic intensity over time were explored, and hot topics and topics with rising and falling intensity were analyzed and evaluated. Furthermore, the hot topics of coverage in the dimension were found out by measuring topic intensity, and the topics with decreasing intensity were analyzed.

First, during the analysis and processing of the text content, stop words should be removed correctly to ensure the accuracy of the analysis results. Then, the LDA topic model was realized through Python-based Sklearn to train the optimal number of topics, and the text content was analyzed by year according to the optimal number of topics, with 2000 iterations. The topic intensity recognized in each natural year and the word probability under each topic were recorded, and topics with an intentional tag were labeled based on the words. Topic intensity over 0.02 was set as a hot topic, and hot events were recognized. The change in the intensity of the same topic between different years was measured to obtain topics of increasing intensity, that is, hot topics, and the topic trend was visualized. The topics that declined in intensity were analyzed in the same manner.

Given that the number of topics directly influences the effect of recognized topics, it is necessary to preset the number of topics K before constructing the LDA model. A low value of K would lead to broadly recognized topics, which does not reflect the core essence of the document, while a high value leads to the recognized topics being too detailed and unsystematic. The generalization ability of the model should be evaluated by effective means for analyzing topic evolution, namely, the well-recognized “perplexity indicators” (calculated from corpus classification tags and high-frequency topic words). The lower value indicated the stronger generalization ability of the model. Generally, when the K value in which the number of topics keeps increasing while the perplexity value drops marginally was selected, some obvious inflection points result. To reflect the degree of difference between topics and ensure the effect of topic recognition, an average topic similarity indicator was added to measure the generalization ability of the model. The number of topics K was determined by integrating the perplexity and average topic similarity indicators to ensure a stronger generalization ability of the model and thus improve the semantic effect of recognized topics.

There was no similar direct calculation method for the time characteristics of topics. The time characteristic was measured by the release time of coverage, because the coverage released on the same day contributes the greatest to the topic heat, while the contribution of past coverage decreases over time. By analyzing the time characteristics of hot topics and using the topic heat calculation method with customized thresholds, the topic heat was comprehensively quantified. The most popular topic was selected from the topic collection to display to the users. Additionally, according to the dynamic changes in the topic content and time, the development trend of topics could be traced, and those topics that were no longer popular were eliminated, ensuring the high quality of hot topic collections.

In terms of the evolution path of topic concentration, by calculating the topic similarity in adjacent periods and determining the changes in topics before and after adjacent periods, the evolution path of topic content was constructed. If there was a strong correlation between the local topics and the global topics, starting from the strongly correlated local topics, the forward topic and the backward topic query were performed consecutively to create the evolution path of the global topics. Without a strong correlation between the local and global topics, a local topic that had a high correlation with the global topic in each period was selected to construct a topic evolution path based on the correlation between local topics in adjacent periods.

The system supervised training based on the provided corpus data. By continuing to iterate the number of cycles, the training effect was ensured to be better to improve the generalization effect of the program. In the experimental stage, combined with the deep model, various toolkits of NLP were used to allow the training effect to continue to regress. The ultimate goal was to control the convergence of the generalization results to a controllable range and to provide reliable results.