Research on Chinese Social Media Metaphor Text Detection Model based on Multi-Feature Fusion

DOI: https://doi.org/10.62517/jike.202604228

Author(s)

Jingxiang Zhang

Affiliation(s)

Computer Science and Technology, Beijing University of Technology, Beijing, China

Abstract

In this paper, a lightweight metaphor text detection model based on multi-feature fusion is proposed to address the challenge of identifying non-literal emotional expressions, such as metaphor and irony, in Chinese social media[1], which remain difficult to detect due to their non-literal semantics, dynamic linguistic patterns, and the limited generalization ability of existing methods. The proposed model incorporates two parallel dynamic feature extractors. CPDI (Contextual PMI-based Dynamic Incongruity) is designed to capture dynamic changes in emotional polarity within specific contexts, while templateGAN[2] is employed to discover and learn emerging online linguistic patterns through adversarial generation. These features are encoded using a dual-channel BiLSTM network and further enhanced via an attention mechanism. Subsequently, the representations are concatenated and nonlinearly transformed through an adaptive fusion layer before being fed into a classifier for prediction. Experimental results demonstrate that the proposed model achieves significant improvements over baseline models on Chinese metaphorical text detection tasks. The overall accuracy reaches 0.9033, with macro-averaged Precision, Recall, and F1 scores of 0.9036, 0.9032, and 0.9033, respectively, indicating its effectiveness in identifying metaphorical expressions.

Keywords

Metaphor Detection; Multi-Feature Fusion; Attention Mechanism; biLSTM; templateGAN; cPDI

References

[1] Deng Dailing, Gao Wenxuan. Research hotspots and development trends of network buzzwords-visual analysis based on cnki database [J].Chinese character culture, 2024, (14): 29-32.DOI: 10.14014 / j.cnki.cn11-2597/g2.2024.14.052. [2] Huang Guizhen, Liang Ting, Zheng Meng, et al. Review of Generative Adversarial Networks [ J ].Intelligent Internet of Things Technology, 2025,57(04):17-20.DOI: 10.26921/j.cnki.2096-6059.2025.04.003. [3] Чирвоний, Олександр Сергійович (2024) The Evolution of Social Media Language: A Sociolinguistic Analysis of Recent Neologisms Закарпатські філологічні студії (35). pp. 127-132. ISSN ISSN 2663-4899 [4] HUANG X X,LIU G F,LIU X Y,et al. Sentiment classification depth model based on word2vec and bi-directional LSTM[J]. Application Research of Computers,2019,36(12):3583-3587,3596. [5] LI Y S,WANG L M,CHAI Y M,et al Research on the construction method of dynamic emotion dictionary based on Bi-LSTM[J]. Journal of Chinese Computer Systems,2019,40(3):503-509. [6] YIN W,SCHÜTZE H,XIANG B,et al. ABCNN：attention-based convolutional neural network for modeling sentence pairs[J]. Transactions of the Association for Computational Linguistics,2016,4：259-272. [7] Joel Philip Thekkekara,Sira Yongchareon & Veronica Liesaputra.(2024).An attention-based CNN-BiLSTM model for depression detection on social media text.Expert Systems With Applications,249(PC),123834-.https://doi.org/10.1016/J.ESWA.2024.123834.