Abstract: Existing research work on metaphor detection task mostly adopted deep learning techniques and did not utilize linguistic rules in depth, which was mainly manifested in the defective representation of semantic and basic meanings of the target words involved in the rules, resulting in the model's inability to focus on the differences between the target words and more relevant contextual words, and the boundaries between the basic meanings and the contextual meanings were still fuzzy. To address the above problems, a metaphor detection model MeRL (Metaphor Detction for Improving Representation in Linguistic Rules) was proposed. First, it enhanced the semantic representation of target words involved in both Selectional Preference Violation (SPV), and Metaphor Identification Procedure (MIP) rules; second, it characterized the basic meaning of target words in Metaphor Identification Procedure rules; and finally, it fused the rule-based design of SPV and MIP modules to jointly identify metaphors. The experimental results on the benchmark datasets VUA-18, VUA Verb, and MOH-X showed that compared with MelBERT (Metaphor-aware late interaction over BERT) and other models, the F1 values of the proposed model were improved by 0.6, 0.9, and 1.2 percentage points respectively, and the detection of metaphors was more accurate. Also, the results of doing zero-shot transfer learning on the TroFi dataset showed that the F1 value was improved by 0.7 percentage points and the generalization ability of the model was stronger.

Key words: metaphor, metaphor detection, pre-training model, linguistic rule, external resource

摘要： 现有隐喻识别任务研究工作多采用深度学习技术，并未深入利用语言学规则，主要表现为规则中涉及的目标词的语义与基本义的表征存在缺陷，导致模型无法聚焦目标词与更相关上下文词之间的差异，基本义与上下文含义界限仍然模糊。针对上述问题，提出了一种改进语言规则中的表示的隐喻识别模型MeRL(Metaphor Detction for Improving Representation in Linguistic Rules)。首先，增强选择偏好违反、隐喻识别过程规则都涉及的目标词的语义表示，其次，表征隐喻识别过程规则中目标词的基本义；最后，融合基于规则设计的SPV(Selectional Preference Violation)与MIP(Metaphor Identification Procedure)模块共同识别隐喻。在基准数据集VUA-18、VUA Verb、MOH-X上的实验结果表明，相比MelBERT(Metaphor-aware late interaction over BERT)等模型，所提模型的F1值分别提高了0.6、0.9、1.2个百分点，识别隐喻更准确。另外在TroFi数据集上做zero-shot迁移学习的结果显示，F1值提高了0.7个百分点，模型的泛化能力也更强。

关键词: 隐喻, 隐喻识别, 预训练模型, 语言学规则, 外部资源