Enhancing text classification with neural label embedding and weakly-supervised learning

Zhe Li

doi:10.1016/j.eswa.2025.129569

Enhancing text classification with neural label embedding and weakly-supervised learning

2025-09-12

0

PRE

AI

DOI：10.1016/j.eswa.2025.129569

Original

Original request for help

Save

Abstract

En 中文

Recent years have witnessed the widespread adoption of deep-learning-based models in a range of linguistic tasks including the fundamental text classification. These deep neural networks, however, often face challenges due to the limited availability of large-scale training data with high-quality label annotations. Furthermore, while supervised learning has proven to be superior in training sentence representations for downstream tasks like text classification, this aspect has received relatively little attention. In this study, a novel model named Label Embedding joint with Weakly-supervised Classification (LemWec) is proposed, which aims to establish a unified framework by combining supervised sentence embedding with multiclass classification. For supervised sentence embeddings, the model incorporates seed information such as label names and designs an encoder network with a new pooling layer. Additionally, the model adopts a pseudo-labeling approach to leverage a substantial amount of unlabeled samples. This approach specifically addresses the drawback of generating pseudo-labels with the highest confidence and introduces a noise adaptation method to mitigate this issue. The results of extensive experiments conducted on four real-world datasets demonstrate that the proposed LemWec model can significantly enhance the performance of text classification when compared to a comprehensive set of baselines.

Enhancing text classification with neural label embedding and weakly-supervised learning

Abstract

Journal

Organization