<p>Emotion Recognition in Conversation is a critical component of next-generation AI, particularly for large-scale, real-time interaction systems. While previous studies highlight the importance of context modeling, existing methods often suffer from computational inefficiency due to information redundancy and fail to adequately capture emotional intensity. To address these challenges under strict latency constraints, this paper proposes a computationally efficient context-aware filtering method. By selectively screening relevant context, this module not only mitigates the impact of noise, but also optimizes resource allocation by reducing the processing load. Furthermore, we design a parallelizable dual-branch architecture: a Directed Acyclic Graph Neural Network captures structural dependencies in a causal manner suitable for online streaming, while a separate branch processes semantic information. Finally, a FiLM layer efficiently modulates these heterogeneous features via multiplicative interaction. Experiments on four benchmark datasets demonstrate that our model achieves state-of-the-art performance. Crucially, the inherent parallelism and streamlined design of our architecture ensure high inference throughput, validating its suitability for deployment in real-time, high-performance computing environments.</p>