【摘要】 为了综合考虑更多模态信息，对语境和语义特征进行了建模，并将它们融合在一起以提取情感原因对。针对语境模态，采用了子句嵌入方法来获取情绪和原因的表示，并通过双因素注意力机制得到全局语境矩阵。同时，通过构建子句间语义的图神经网络，得到了局部语义特征。最后，通过主模态和辅助模态的匹配，得到了融合特征，以进行多任务预测，包括情感句、原因句和情感-原因对的抽取。实验结果表明，在抽取经典中文情感原因对数据时，相较于最佳基线系统，所提模型的F测度提高了2.2%。更多还原

【Abstract】 To consider more modal information, contextual and semantic features are modeled in detail, and emotion cause pair extraction is carried out on the fusion of two modal features.For the contextual modality, a clause embedding method is employed to obtain representations of emotions and causes, and dual-factor attention mechanism is utilized to derive a global contextual matrix. In the mean time, by constructing a graph neural network for inter-clause semantics, local semantic features are obtained. Finally, the fusion features are obtained by the main and auxiliary mode matching method for multi-task prediction, including emotional sentence, cause sentence and emotion-cause pair extraction task. Experimental results indicate that when extracting classic Chinese emotion-cause pairs, compared to the best baseline system, the F-measure has improved by 2.2%.更多还原