SC-Taxo: Hierarchical Taxonomy Generation under Semantic Consistency Constraints using Large Language Models
| Authors | Shiqiang Cai et al. |
| Year | 2026 |
| Field | NLP |
| arXiv | 2605.00620 |
| Download | |
| Categories | cs.CL |
Abstract
Scientific literature is expanding at an unprecedented pace, making it increasingly challenging to efficiently organize and access domain knowledge. A high-quality scientific taxonomy offers a structured and hierarchical representation of a research field, facilitating literature exploration and topic navigation, as well as enabling downstream applications such as trend analysis, idea generation, and information retrieval. However, existing taxonomy generation approaches often suffer from structural inconsistencies and semantic misalignment across hierarchical levels. Through empirical analysis, we find that these issues largely stem from inadequate modeling of hierarchical semantic consistency. To address this limitation, we propose a semantic-consistent taxonomy generation (SC-Taxo) framework that leverages large language models (LLMs) with hierarchy-aware refinement stages to ensure semantic consistency. Specifically, SC-Taxo introduces a bidirectional heading generation mechanism that jointly performs bottom-up abstraction and top-down semantic constraint, while further capturing peer-level semantic dependencies to enhance horizontal consistency. Experiments on multiple benchmark datasets demonstrate consistent improvements in hierarchy alignment and heading quality, and additional evaluation on Chinese scientific literature validates its robust cross-lingual generalization.
Engineering Breakdown
Plain English
This paper addresses the problem that scientific taxonomies—hierarchical structures organizing research domains—are difficult to generate automatically with consistent quality across levels. Existing methods produce taxonomies with structural inconsistencies and semantic misalignment where parent-child relationships don't maintain coherent meaning. The authors identified that the root cause is inadequate modeling of hierarchical semantic consistency, meaning the algorithms don't properly enforce that concepts at different levels of the hierarchy remain semantically aligned. They propose SC-Taxo (semantic-consistent taxonomy generation), a framework designed to generate taxonomies where terms at each hierarchical level maintain consistent semantic relationships with their parents and children, enabling better organization of rapidly expanding scientific literature.
Core Technical Contribution
The core innovation is the introduction of hierarchical semantic consistency as an explicit modeling objective in taxonomy generation. Unlike prior work that treats hierarchy construction as a local or greedy problem, SC-Taxo enforces global semantic consistency constraints across all hierarchical levels simultaneously. The key technical novelty appears to be a constraint-based or loss-function approach that penalizes semantic misalignment between parent concepts and their child clusters, ensuring that child nodes are genuinely subconcepts of their parents rather than semantically distant terms. This shifts the problem from purely structural hierarchy learning to a joint optimization of structure and semantics, which is a meaningful departure from existing approaches that handle these aspects independently.
How It Works
The framework takes raw scientific literature (abstracts, titles, or full texts) and their associated domain metadata as input. Step one extracts semantic embeddings for all candidate terms using pretrained language models, capturing the meaning of each potential taxonomy node. Step two constructs hierarchical relationships by grouping semantically similar terms while enforcing that child clusters maintain semantic alignment with parent concepts—likely using a hierarchical clustering approach with consistency constraints. Step three iteratively refines the taxonomy by checking that the semantic distance between parent terms and their children clusters stays within acceptable bounds, adjusting assignments or cluster boundaries when violations are detected. The output is a multi-level taxonomy where traversing from root to leaf maintains semantic coherence, with each level representing increasingly specific research subtopics.
Production Impact
For teams building literature management systems, this directly solves the problem of auto-generated taxonomies producing nonsensical hierarchies where unrelated papers end up grouped together. Adopting SC-Taxo would replace current bag-of-words or simple clustering approaches with a semantically-aware system that produces taxonomies suitable for direct user navigation without manual curation. In production, this enables better downstream applications: researchers can reliably explore literature by following the hierarchy, recommendation systems can traverse parent-child relationships with confidence, and trend analysis becomes more accurate when the taxonomy structure reflects actual semantic boundaries. The trade-off is increased computational cost during taxonomy generation (likely O(n²) or higher due to consistency checking), but since this is typically a one-time or infrequent batch operation, it's acceptable. Integration would require swapping the taxonomy generation module while maintaining compatible output formats with existing downstream systems.
Limitations and When Not to Use This
The paper assumes that semantic embeddings from pretrained models adequately capture domain-specific meaning, which may not hold in niche or interdisciplinary fields where terminology differs from general text. The approach likely requires careful tuning of consistency thresholds—too strict and valid hierarchies are rejected, too loose and semantic alignment is meaningless—and the paper doesn't clearly specify how to set these hyperparameters across different domains. The method assumes a clear hierarchical structure exists in the data; in reality, many research relationships are multiparent (a concept could reasonably fit under multiple parents) or cross-cutting, and forcing a strict tree structure may lose important conceptual connections. Additionally, as an abstract, the paper doesn't specify scalability characteristics or how performance degrades with extremely large taxonomies (millions of terms), which limits applicability to very large domains like biology or medicine.
Research Context
This work builds on decades of taxonomy and ontology learning research, improving upon methods like TAXI and recent neural hierarchy construction approaches. It directly addresses known failure modes in automatic taxonomy generation, where prior work (likely cited in the full paper) demonstrated that ignoring semantic consistency produces hierarchies that don't reflect actual domain structure. The contribution fits into the broader trend of making taxonomy generation less reliant on manual curation or distant supervision, moving toward fully automatic, semantically-aware systems. This opens research directions in hierarchical semantics, multiparent taxonomy structures, and cross-domain taxonomy alignment—areas that remain relatively unexplored in the automatic generation literature.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::