基于Transformer和图卷积神经网络的隐喻识别

Identifying Metaphor with Transformer and Graph Convolutional Network

【目的】提出一种基于图卷积神经网络和Transformer的隐喻识别模型,既能有效识别单词隐喻,又能解决多个词语共同引发的隐喻表达。【方法】利用图卷积神经网络提取句法依存树中的句法结构信息,联合从Transformer结构学习的深层语义表示,通过SoftMax计算目标词隐喻表达的概率。【结果】在英文数据集UVA VERB和UVA ALLPOS上F1值分别提高1.9%和1.7%,TOEFL VERB和TOEFL ALLPOS上F1值分别提高1.1%和1.9%;在中文数据集CCL上F1值提高1.2%。【局限】如果句子中存在歧义或者指代信息不明确的现象,则不能有效识别句子中的隐喻现象。【结论】图卷积神经网络和句法依存树的引入确实能在一定程度上丰富目标词语义信息,提高单词和多词隐喻的识别效果。

[Objective]This paper proposes a metaphor identification model based on graph convolutional neural network and Transformer,aiming to effectively find metaphor expressions with multiple words.[Methods]We used the graph convolutional neural network to extract the structure information from the syntactic dependency tree.Then,we combined the structure with deep semantic representation by the Transformer.Finally,we calculated the probability of metaphorical expression for the target words through SoftMax.[Results]Compared with the existing algorithms,the F1 values of our model increased by 1.9% and 1.7% on UVA VERB and UVA ALL POS datasets.The F1 values were also improved by 1.1% and 1.9% on TOEFL VERB and TOEFL ALL POS.The F1 value increased by 1.2% on the Chinese data CCL.[Limitations]If there is ambiguity or ambiguous referential information in the sentence,our model will not effectively identify the metaphor expressions.[Conclusions]Graph convolutional network and syntactic dependency tree can enrich the semantics of target words,which improves the recognition of single and multi-word metaphors.