Extracting Causal Relationships Between Events and Sentiments in Chinese-English Business Communications Using Hierarchical Shared Encoder

This paper introduces HSE-TTLS, a model that extracts causal relationships between events and sentiments in Chinese-English business communications texts using a hierarchical shared encoder. Drawing from linguistic theories of contextual sentiment analysis and causal inference patterns, the model improves cross-lingual business text processing through two innovative approaches: a business -specific layer selection strategy and a sentiment dictionary filtering mechanism. For layer selection, the model precisely selects and combines features from different BERT layers based on business language characteristics, enhancing its understanding of technical terms and complex sentence structures. The sentiment dictionary filtering implements a multilevel business sentiment dictionary system with a soft filtering mechanism to improve sentiment recognition accuracy. Experimental results show that the model achieves an F1 score of 0.834 on standard datasets, outperforming the best existing models by 3.4%. The model shows strong performance in Chinese business text processing with a 10.4% improvement in F1 score. Ablation studies further validate the effectiveness of each innovative component, offering a new solution for cross-lingual text analysis in business.