Zhu Dongmei
School of Foreign Studies, Zhongyuan University of Technology, Zhengzhou, 450007, China.
Sci Rep. 2025 May 29;15(1):18947. doi: 10.1038/s41598-025-04171-5.
This study aims to enhance computers' ability to understand and generate metaphors, offering a novel perspective and technical approach in the field of natural language processing. It proposes a metaphor recognition algorithm that combines a Convolutional Neural Network (CNN) with a Support Vector Machine (SVM). First, the text is transformed into numerical features using a pre-trained word embedding model. Then, local contextual features are extracted through a multi-layer CNN. These features are subsequently input into the SVM for classification, enabling optimal metaphor recognition. In English verb metaphor recognition tasks, the model-when combined with the SVM classifier-achieves an accuracy of 85%, an F1 score of 85.5%, and a recall of 86%. In Chinese metaphor recognition experiments, the integration of the SVM classifier significantly improves performance, yielding an F1 score of 81.5%, an accuracy of 81%, and a recall of 82%. In conclusion, the proposed model effectively integrates the CNN's powerful feature extraction capabilities with the SVM's superior classification performance. Additionally, it incorporates part-of-speech features to enhance semantic analysis. This integrated approach enables more accurate identification of complex textual semantics, particularly in interpreting metaphorical language that requires deeper understanding.
本研究旨在提高计算机理解和生成隐喻的能力，为自然语言处理领域提供一种新颖的视角和技术方法。它提出了一种将卷积神经网络（CNN）与支持向量机（SVM）相结合的隐喻识别算法。首先，使用预训练的词嵌入模型将文本转换为数值特征。然后，通过多层CNN提取局部上下文特征。这些特征随后被输入到SVM中进行分类，从而实现最佳的隐喻识别。在英语动词隐喻识别任务中，该模型与SVM分类器相结合时，准确率达到85%，F1分数为85.5%，召回率为86%。在中国隐喻识别实验中，SVM分类器的集成显著提高了性能，F1分数为81.5%，准确率为81%，召回率为82%。总之，所提出的模型有效地将CNN强大的特征提取能力与SVM卓越的分类性能相结合。此外，它还纳入了词性特征以增强语义分析。这种集成方法能够更准确地识别复杂的文本语义，特别是在解释需要更深入理解的隐喻性语言时。
