基于融合模型的名词隐喻识别-《计算机技术与发展》

文章信息/Info

作者:: 苏魁麟; 张凯; 吕学强; 张乐; 1. 北京信息科技大学网络文化与数字传播重点实验室,北京 100101;
2. 首都师范大学文学院中国语言智能研究中心,北京 100048

Author(s):: SU Kui-lin1 ; ZHANG Kai2 ; LYU Xue-qiang1 ; ZHANG Le1; 1. Beijing Key Laboratory of Internet Culture & Digital Dissemination Research,Beijing Information Science & Technology University,Beijing 100101,China;
2. Research Center for Language Intelligence of China,School of Literature Capital Normal University,Beijing 100048,China

Keywords:: metaphor recognition; noun metaphor; feature fusion; semantic information; CNN; BERT

摘要:: 隐喻识别是自然语言处理各前沿领域中所面临的难题。为了解决名词性隐喻中忽视的潜在特征和语义的信息利用不足从而导致隐喻识别效果不高的问题,利用深度学习的优势,该文提出一种特征融合神经网络模型冥CB,针对名词性隐喻进行识别。使用卷积神经网络模型冥CNN 挖掘名词性隐喻句中的潜在特征,预训练表征模型冥BERT 对词语之间的关系和词的位置信息进行向量化表征,以此有效地学习名词性隐喻句中的语义信息。在隐藏层特征维度上融合两者提取到的信息,最后通过线性分类器进行识别。由于模型本身具有局限性,名词性隐喻句中还蕴含少量抽象的特性,因此无法只依靠模型挖掘所有的特征信息,但针对大部分非抽象名词性隐喻句能够在不耗费人力资源的条件下有较好的识别效果。经过实验对比发现 CB 模型在准确率上达到 0. 904 7、召回率 0. 936 2、F1 值 0. 926 2,其综合指标均高于现有的最优深度学习模型。

Abstract:: Metaphor recognition is a difficult problem in various fields of natural language processing. To solve the problem of poormetaphor recognition caused by hidden features overlooked in metaphors and insufficient use of semantic information,taking advantage ofdeep learning,we introduce the feature fusion neural network model ( CB) to recognize nominal metaphors. First,convolutional neuralnetwork ( CNN) is used to excavate the potential features in metaphorical sentences. Then, BERT, a pre - training model, is used torepresent words and positions to excavate the semantic information in metaphorical sentences. It can effectively learn the semantic information in nominal metaphorical. Finally,the information extracted from CNN and BERT is recognized by the linear classifier. Dueto the limitations of the model itself,nominal metaphorical sentences contain a small amount of abstract features,so it is impossible tomine all the feature information only by relying on the model. However,for most non-abstract nominal metaphorical sentences,it canachieve a better recognition effect without consuming human resources. Through experimental comparison,it is found that the accuracyrate of CB model is 0. 904 7,the recall rate is 0. 936 2,and the F1 value is 0. 926 2,all of which are higher than the existing optimal deeplearning model.