.png){kind=icon}
.png){kind=icon}
Return
Discrete linguistic structures learning for entity relation extraction
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2026.116051.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Although large language models (LLMs) have excelled in diverse natural language processing tasks, they continue to struggle with relation extraction, particularly for sentences containing multiple or overlapping relation instances with shared contextual features. Thus, effective relation extraction requires accurate identification of the linguistic structures of sentences. However, existing works on this topic assume complete connectivity between linguistic units (e.g., tokens or words) and only model their intensities as continuous real-valued variables. This is inconsistent with the theory of cognitive linguistics, which argues that discrete conscious perception follows unconscious continuous processing in the human cognitive system. To account for the absence of discrete conscious perception, we propose an Adaptive Discrete Convolutional Network (AdaDCNet). By implementing a novel discrete convolutional operation, AdaDCNet adaptively quantizes weights corresponding to different linguistic units in a sentence, enabling the construction of task-specific linguistic structures inspired by human cognition. Evaluated on six entity relation extraction datasets, AdaDCNet effectively captures discrete linguistic structures and significantly enhances performance. Notably, our model outperforms both conventional deep learning architectures and several powerful LLMs, including ChatGPT-4o (zero-shot) as well as instruction-tuned Qwen2 and Llama 3 (open-source). Furthermore, comprehensive complexity analysis and reliability calibrations confirm that AdaDCNet achieves a favorable balance between computational efficiency and model robustness. These findings validate the effectiveness and practical viability of our approach, providing a solid foundation for its application in broader structure prediction tasks.
Keywords:
Entity relation extraction
Discrete linguistic structures
Adaptive Discrete Convolutional Network
Large language models
Cognitive linguistics
Journal
K
IF:
7.6
Papers:
1.2W
Citations:
4.5W