Seeing China from different perspectives: Analyzing Media reporting on the sustainability impacts of Chinese overseas investments.

Over the last two decades, China has quickly expanded its overseas engagement through large-scale investments across the Global South, especially its 2013 Belt and Road Initiative. However, narratives around the impacts of these investments diverge greatly, with potentially great variation between perspectives from the West and the Global South. Existing research has yet to systematically examine various narratives around China’s overseas investments. To fill this knowledge gap, our study explores the changing dynamics of different media narratives regarding the impacts of Chinese overseas investments and potential causal relationships within these narratives. To do so, we focus on Southeast Asia – a region having a pivotal role in China’s Belt Road Initiative as recipient of large amounts of China’s development finance and foreign investments. Our computational analysis investigates the reports on Chinese overseas investments from mainstream media outlets in four key Southeast Asian countries, influential international media led by the West and also key state media in China between 2010 and 2024 To systematically analyse these narratives, this paper explores cutting-edge development in the field of natural language processing (NLP), focusing on the application of deep learning techniques in political science. We first use qualitative content analysis to classify a small group of media reports on the impacts of Chinese investments into three thematic types: positive, negative, and neutral narratives. Then, we leverage these manually labelled samples to train a supervised machine learning model, aiming to develop an automatic text classification system and reduce the high research intensity required by manual coding. Building upon the RoBERTa model, we introduce a classification layer—essentially a smaller neural network—on top of the pre-trained model. This fine-tuning process allows the pre-trained model to be adapted for more specialised tasks. To enhance its capacity to perform classifications with greater precision, we further train the model on the previous smaller, manually coding data. This approach enables a high accuracy of classifications even with a relatively limited training set. Following this approach, we seek to investigate the evolving trends over time by analysing the classification results aggregated across the three narrative types. The time series analysis also vividly highlights notable or special cases. To better understand the mechanisms within these competing narratives, we apply causal mapping techniques to explore potential causal relationships between Chinese investments and their sustainability (economic, social and environmental) impacts underlying these narratives. Causal mapping is a method developed by Causal Map Ltd (https://www.causalmap.app/) for analysing large volumes of text to identify and visualise causal relationships described within it. This approach uses a new generative Large Language Model (LLMs) as a supportive assistant to systematically and transparently identify a wide range of individual causal claims within texts. As the first paper applying these text-as-data techniques to a topic of this kind, we believe our study can make novel contributions to the political science literature on media narratives and China’s global influence, providing valuable insights into practices and standards for the use and validation of recent NLP tools.