一种基于词语抽象度的汉语隐喻识别方法

An Approach to Chinese Metaphor Identification Based on Word Abstractness

【目的】设计一种自动计算汉语词语抽象度的方法,并将其用在自然语言理解中的隐喻识别任务。【方法】以统计学习理论中逻辑回归为计算模型,把神经网络语言模型获取的词语词向量作为特征,通过构建抽象词库得到特征权重向量,计算汉语词语抽象度。提出一种基于词语抽象度的汉语隐喻识别算法,验证该方法的应用效果。【结果】通过与已有的方法进行实验对比,本文设计的汉语词语抽象度计算方法更接近于人的认知常识;并且在隐喻识别任务中,也体现出更好的准确率。【局限】词语词向量表示词语抽象程度有一些缺陷;抽象词语库的规模影响特征权重向量的学习。【结论】词语抽象度计算可以表现为人对概念的一种抽象分类能力,本文提出的汉语词语抽象度计算方法得到的结果能够较好地拟合人的认知,并且实验证明词语抽象度可有效提高隐喻识别的效果。

[Objective] Design a method to automatically compute Chinese word abstractness, and introduce it into metaphor identification task in natural language understanding. [Methods] The word abstractness is computed by logistic regression model. The features are the word vectors computed by neural network model and the feature weight vectors come from a hand coded abstractness dictionary. A metaphor identification algorithm based on word abstractness is proposed to demonstrate the validity of this method. [Results] By comparing with the existing methods of word abstractness computing, this method has better accordance with human cognition and is an effective method in metaphor identification task. [Limitations] The utilization of word vectors for word abstractness is defective. The scale of the abstract words affects the learning of feature weight vectors. [Conclusions] Word abstractness computing reflects the ability to concept classification, Chinese word abstractness computed by this method is better fitting the human cognition, and the experimental results show that word abstractness can improve the effect of metaphor identification.