.png){kind=icon}
Return
DynamAlign: A fine-grained label dynamic alignment framework for complex semantic text classification
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2026.132703.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Fine-grained label dynamic alignment (FLDA) is an advanced natural language processing technique that dynamically assigns fine-grained labels to text based on semantic information without predefined label sets. FLDA provides fundamental support for text information fusion and complex semantic text classification. The advancement of Large Language Models (LLMs) provides new opportunities for this task, but challenges remain with LLM memory decay and LLM hallucination, often resulting in redundant labels and extremely long-tail labels during the alignment process. Moreover, existing methods rely primarily on pre-trained language models or fine-tuned LLMs, making it difficult to regulate model outputs. To address these issues, we propose DynamAlign, a fine-grained label dynamic alignment framework incorporating Structured Prompt Engineering, External Knowledge Base, and Dual-Dimensional Critical Mechanism. To solidify LLMs’ reasoning paths and regulate their outputs, we propose Structured Prompt Engineering. To effectively reduce redundant labels and alleviate LLM memory decay, External Knowledge Base is constructed to provide structured storage of dynamically generated labels for model retrieval during alignment tasks. To reduce the occurrence of extremely long-tail labels caused by LLM hallucination, we propose Dual-Dimensional Critical Mechanism that evaluates outputs from two perspectives. We evaluate the performance of DynamAlign by instantiating it with 8 distinct LLMs as backbones across two datasets (AGNC, UCNSC). Furthermore, we conduct a comprehensive comparison against pre-trained language model methods, contrastive learning methods, open-source LLM fine-tuning method, and state-of-the-art alignment method. Experimental results show that compared to state-of-the-art alignment method, DynamAlign improves label generation accuracy by 4.03% and 4.01% on two datasets respectively, while reducing label semantic redundancy by 6.63% and 5.51%, demonstrating its effectiveness in achieving fine-grained label dynamic alignment for complex semantic texts within the evaluated domains, establishing a solid foundation for future broader cross-domain validation.
Keywords:
Fine-grained label dynamic alignment
Large Language Models
Structured Prompt Engineering
External Knowledge Base
Dual-Dimensional Critical Mechanism
Journal
IF:
7.5
Papers:
3.0W
Citations:
10.2W
Organization
View more organizations
Cited Papers
No cited papers available