基于多特征多注意力模型的中文事件关系识别

Chinese event relation recognition based on MFMA model

当前的关系识别任务多采用基于词或字粒度单一特征进行,忽略了全局信息对于关系识别的重要性,而且现代汉语具有词类和句法成分关系复杂的特点,这使得特征选择成为中文文本处理中的重点与难点。文中所用多特征多注意力模型除事件自身特征外,充分地考虑到位置、事件要素和上下文三类额外特征,利用全局信息以解决模型特征矩阵语义表征不足的问题。该模型结合双向注意力机制、点积注意力机制和双向门控循环神经网络进行事件关系识别,结合注意力机制的神经网络模型来较好地提取文本中的深层语义信息。其中双向注意力从特征矩阵两个方向提取事件自身有效信息,点积注意力提取事件之间的对应关系,双向门控循环神经网络提取矩阵中的上下文特征。在CEC2.0中文突发事件语料库上的实验结果表明,文中方法以及所用模型均有较好的识别效果。

The current relationship recognition tasks are mostly based on a single feature of word granularity,ignoring the importance of global information for relationship recognition. Besides,modern Chinese has characterized of complex relationship between parts of speech and syntactic components,which makes feature selection become the focus and difficulty in Chinese text processing. In addition to the characteristics of the event itself,the multi feature and multi attention(MFMA)model used in this paper takes full account of the three additional characteristics of location,event elements and context,and makes full use of the global information to solve the problem of insufficient semantic representation of the feature matrix of the model. In combination with bidirectional attention mechanism,dot product attention mechanism and Bi-GRU(bi-gated recurrent unit),the model is used to identify the relationship between events, and combine with the neural network model of attention mechanism to well extract the deep semantic information in the text. The bidirectional attention is used to extract the effective information of the event itself from two directions of the feature matrix,dot product attention is used to extract the corresponding relationship between events,and Bi-GRU is used to extract the context features in the matrix. The experiments on the CEC2.0Chinese emergency corpus show that the method and the model presented in this paper have good recognition effect.