Financial news exerts a profound impact on market volatility, particularly during abrupt negative events. However, traditional risk management approaches relying on historical data face significant challenges in real-time analysis, due to the high velocity and unstructured nature of incoming news, as well as the difficulty in promptly identifying relevant information. In addition, they lack the ability to explicitly reveal which parts of the news contribute to risk predictions, limiting their interpretability and usefulness in decision-making. To address these issues, we propose INDRE, a novel framework that integrates real-time financial news with market data for dynamic risk assessment. Leveraging Rhetorical Structure Theory (RST), the model decomposes news texts into Elementary Discourse Units (EDUs) and constructs semantic graphs where EDUs serve as nodes and rhetorical relationships as edges. A Graph Attention Network (GAT) captures fine-grained semantic dependencies among events. Concurrently, a BiLSTM network extracts temporal market features, while a cross-modal attention mechanism fuses textual and market information to predict trends and assess risks. Experimental results confirm the superior performance of INDRE. Furthermore, INDRE enhances interpretability by visualizing the influence of specific EDUs and their semantic roles, making the reasoning behind risk assessments transparent. This study offers methodological advancements for real-time financial risk analysis and demonstrates the practical value of structured text representation in event-driven quantitative investment.