融合全局和局部注意力机制的自然语言框架识别方法

FRAME IDENTIFICATION METHOD INTEGRATING GLOBALAND LOCAL ATTENTION MECHANISM

框架识别是框架语义学研究中进行浅层语义分析的核心任务,要求根据句子中目标词的上下文语义场景从给定的框架库中选择最相符的框架。现有的识别方法只考虑了句子的全局特征,忽略了目标词周围的局部信息,基于此,提出一种融合全局和局部注意力机制的框架识别方法。通过BERT预训练模型生成输入文本的向量表示;利用全局注意力机制和局部注意力机制分别对上下文与目标词周边信息进行编码;融合全局和局部信息编码进行框架选择。实验结果表明,该方法在FrameNet和CFN数据集上分别取得了88.39%和74.90%的准确率,优于多个基线模型,且对中英文数据具有较好的适应性。

Frame identification is the core task of shallow semantic analysis in frame semantics research,which is to select the most consistent frame from a given frame base according to the contextual semantic scene of the target word in the sentence.The existing Identification methods usually only consider the global features of the sentence and fail to capture the local information features around the target word.Based on this,the paper proposes a frame identification method that combines global and local attention mechanisms.A vector representation of the input text was generated through the BERT pre-training model,and the global attention mechanism and the local attention mechanism were used to respectively encode the context and surrounding information of the target word.The global and local information encoding was combined for frame selection.The experimental results show that the proposed method achieves the accuracy of 88.39%and 74.90%on the FrameNet and CFN dataset,which is better than multiple baseline models,and has good adaptability to Chinese and English data.