ReLTEx: A New Framework for Reliable LLM-Based Taxonomy Expansion
A novel framework named ReLTEx has been developed to tackle the difficulties associated with automated taxonomy expansion using Large Language Models (LLMs). This framework integrates candidate generation driven by LLMs with structure-aware validation and recursive expansion control, aiming to minimize hallucinations while enhancing the consistency and quality of the resulting taxonomies. The research, which can be found on arXiv with the identifier 2608.10970, assesses ReLTEx through benchmark taxonomies in a masked taxonomy expansion context, comparing various validation techniques. Experimental findings, bolstered by tailored evaluation metrics and human assessments, reveal that ReLTEx yields more dependable expansions compared to direct LLM generation. This advancement is vital for knowledge organization, providing a stronger approach to enriching taxonomies essential for applications like information retrieval and data management.
Key facts
- ReLTEx is a framework for reliable LLM-based taxonomy expansion.
- It combines LLM-driven candidate generation with structure-aware validation and recursive expansion control.
- The framework aims to reduce hallucinations and improve consistency and quality of generated taxonomies.
- Evaluation was conducted using benchmark taxonomies under a masked taxonomy expansion setting.
- Multiple validation strategies were compared.
- Results were supported by adapted evaluation metrics and human evaluation.
- The paper is available on arXiv with identifier 2608.10970.
- The research addresses limitations of directly relying on LLM-generated expansions, which often lead to noisy or hierarchically inconsistent structures.
Entities
Institutions
- arXiv