Abstract

Transformer models and their variants have shown significant advantages in natural language processing tasks, but their high computational requirements limit their deployment on resource-constrained devices. To achieve the balance of computational cost and accuracy, the lightweight model pNLP-Mixer utilizes parameter-free projection to generate text embeddings. Nevertheless, the text embeddings generated by projection involve shallow semantic information and ignore the exploration of implicit semantic information. To overcome the challenges of high parameter cost and insufficient representation capabilities in existing models, we propose a method that incorporates topic information to enhance semantic richness. Our approach leverages the Latent Dirichlet Allocation (LDA) topic model to capture latent semantic relationships between words, thereby improving the expressiveness of text representations for downstream tasks. Building on this method, we propose a lightweight model named TEP-Mixer, which integrates multiple feature extraction modules to further enhance representation capabilities. Experimental results demonstrate that TEP-Mixer outperforms other lightweight models in accuracy while maintaining a lower parameter count across multiple benchmark datasets. It is suitable for resource-constrained devices.

Keywords

MSC

68T07

Cite this article

@article{cai-7311,
  author  = {Shuobin Zhang and Jianhua Sun and Huanling Tang and Wenhao Duan and Quansheng Dou and Mingyu Lu},
  title   = {A Lightweight Text Representation Method Integrating Topic Information},
  journal = {Computing and Informatics},
  volume  = {44},
  number  = {6},
  pages   = {1445--1467},
  year    = {2025},
  issn    = {1335-9150},
  doi     = {10.31577/cai_2025_6_1445}
}
Download .bib