💬 NLP

SC-Taxo: Hierarchical Taxonomy Generation under Semantic Consistency Constraints using Large Language Models

The paper proposes SC-Taxo, a framework that leverages large language models with a bidirectional heading generation mechanism to address structural inconsistencies and semantic misalignment in scientific taxonomy generation by ensuring hierarchical semantic consistency.

Original authors: Shiqiang Cai, Nianhong Niu, Shizhu He, Kang Liu, Jun Zhao

Published 2026-05-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shiqiang Cai, Nianhong Niu, Shizhu He, Kang Liu, Jun Zhao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you walk into a massive library that is growing so fast every day that the librarians can't keep up. Books are being thrown onto shelves randomly. Some books about "Advanced Physics" are sitting next to "How to Bake a Cake," and others are buried so deep you can't find them. This is the problem scientists face with the explosion of research papers today. They need a way to organize this chaos into a clear, logical map—a taxonomy—so people can find what they need.

The paper introduces a new tool called SC-Taxo to solve this. Here is how it works, explained simply:

The Problem: The "Confused Librarian"

Existing methods for organizing these papers are like librarians who are either:

Too rigid: They just group books by how similar the words look on the cover, which often puts unrelated things together.
Too imaginative: They use powerful AI (Large Language Models) to write the labels, but the AI gets distracted. It might read a single sentence in a paper about "math" and decide the whole book belongs in a "Math" section, even if the book is actually about "Robotics." This creates a messy map where "Python Code" sits right next to "Faster R-CNN" (a complex algorithm), confusing the hierarchy.

The main issue is semantic consistency: The labels don't match the structure, and the structure doesn't match the meaning.

The Solution: SC-Taxo (The "Double-Check" System)

SC-Taxo is like hiring a team of two expert librarians who constantly check each other's work, rather than just one person guessing.

1. The Two Paths (The "Skeleton" and the "Brain")

Instead of asking the AI to build the whole map from scratch, SC-Taxo starts with two separate approaches:

The Skeleton (Structural Path): It uses math to group papers based on how similar they are, creating a rough "skeleton" tree. This is stable but has no names on the branches yet.
The Brain (Semantic Path): It uses a smart AI to read the papers and come up with a list of cool concepts and names. This is full of ideas but might be structurally messy.

2. The Deep Fusion (The "Four-Round Debate")

This is the magic part. The system forces the "Skeleton" and the "Brain" to talk to each other in four rounds of debate to fix mistakes:

Round 1 (Reality Check): The system asks, "Does this AI-generated concept actually exist in the group of papers we found?" If the AI invented a fake concept, it gets thrown out.
Round 2 (Naming): Now that we know the groups are real, the AI gives them proper names based on what's actually inside.
Round 3 (Parent-Child Check): This ensures the hierarchy makes sense. If a parent branch is "Supervised Learning," the child branch can't suddenly be "Mathematical Concepts." The AI is forced to look at the parent's name and the child's content to make sure they fit together perfectly.
Round 4 (Sibling Check): This looks at "brother" branches (sibling nodes) to make sure they aren't saying the same thing or missing a key topic. It ensures the map is balanced and complete.

3. The Quality Control (The Final Polish)

Before handing over the final map, the system does a final sweep:

It scores every label to make sure it's specific and useful.
It deletes duplicate labels (like "AI" and "Artificial Intelligence" appearing twice).
It checks that the tree isn't too deep or too shallow.

Why It Works (The Results)

The authors tested this on English scientific papers and a tricky set of Chinese papers.

Better Structure: The resulting maps looked much more like the "gold standard" maps made by human experts.
Fewer Mistakes: It stopped the AI from getting distracted by single words and putting "Python Code" next to high-level algorithms.
Language Proof: It worked just as well on Chinese papers as English ones, proving it understands the ideas, not just the specific words.

The Bottom Line

SC-Taxo is a framework that stops AI from "hallucinating" (making things up) when organizing scientific knowledge. By forcing the AI to constantly check its work against the actual data and its own structure, it creates a clean, logical, and reliable map of scientific knowledge that humans can actually use.

1. Problem Statement

The rapid expansion of scientific literature necessitates efficient organization of domain knowledge. While hierarchical taxonomies are crucial for navigation, trend analysis, and information retrieval, existing generation methods suffer from two primary failure modes:

Structural Distortion & Granularity Misalignment: Traditional methods often mix different levels of abstraction (e.g., placing specific implementation details like "Python Code" alongside high-level algorithms like "Faster R-CNN" under the same parent).
Semantic Drift & Hallucination: Pure Large Language Model (LLM) approaches often generate "out-of-domain" nodes by overfitting to local keywords in abstracts (e.g., generating a "Mathematical Concepts" node under "Supervised Learning" because an abstract mentions "mathematical concept," ignoring the global algorithmic context).
Root Cause: These issues stem from the inadequate modeling of hierarchical semantic consistency. Existing methods fail to explicitly enforce top-down conceptual constraints and bottom-up semantic abstractions simultaneously, nor do they coordinate sibling nodes to prevent redundancy.

2. Methodology: The SC-Taxo Framework

SC-Taxo proposes a framework that decouples structural induction from semantic generation, integrating them through a rigorous Deep Fusion mechanism. The pipeline consists of three main phases:

A. Dual-Path Initialization

To prevent ungrounded generation, the framework initializes two parallel paths:

Structural Path (Clustering): Papers are encoded into dense embeddings, reduced via UMAP, and partitioned into a structural Cluster Tree ( $T_{cluster}$ ) using a recursive Breadth-First Search (BFS). This provides a stable, language-agnostic topological scaffold.
Semantic Path (Concept Construction): An LLM extracts key technical concepts from abstracts, merges/synonyms, and induces a preliminary Concept Tree ( $T_{concept}$ ). This provides rich semantic meaning but lacks structural stability.

B. Deep Fusion (Four-Round Validation)

The core innovation is a four-round mutual validation process that aligns $T_{cluster}$ and $T_{concept}$ :

Round 1 (Concepts $\to$ Clusters): Validates if LLM-generated concepts are grounded in actual paper distributions. Concepts with dispersed paper associations are filtered out as hallucinations.
Round 2 (Clusters $\to$ Concepts): Assigns the most representative validated concept as the heading for each structural cluster. Clusters lacking clear support are flagged for re-clustering.
Round 3 (Parent-Child Alignment): Enforces Vertical Consistency. When generating a child node's title, the LLM is conditioned on: (1) the global root title, (2) the parent node's title, and (3) the child's specific paper subset. This prevents semantic drift.
Round 4 (Sibling Expansion): Enforces Horizontal Consistency. The LLM reviews existing sibling nodes to identify missing technical concepts while ensuring no conceptual overlap or granularity mismatch, reducing redundancy.

C. Quality Control

Before final output, the taxonomy undergoes:

Concept Scoring: LLMs score nodes (1-10) based on specificity and coverage; low-scoring nodes are pruned.
Redundancy Detection: Embedding similarity and LLM verification remove synonyms or subsumed concepts.
Structure Checking: Validates depth and branching factors to ensure a balanced hierarchy.

3. Key Contributions

SC-Taxo Framework: A novel architecture that explicitly mitigates structural distortion and context drift by decoupling topological clustering from LLM-driven semantic induction.
Deep Fusion Mechanism: Introduces a bidirectional heading generation mechanism for rigorous vertical alignment and peer-level semantic association for horizontal redundancy reduction.
Robust Cross-Lingual Generalization: Demonstrated effectiveness on both English and Chinese scientific literature, proving the framework captures universal conceptual regularities rather than language-specific surface patterns.

4. Experimental Results

The framework was evaluated on the TaxoBench benchmark (English) and a curated Chinese scientific literature dataset (LLMs + Reinforcement Learning).

Key Metrics:

CEDS (Catalogue Edit Distance Similarity): Measures structural and semantic isomorphism with ground truth.
HSR (Hierarchical Structure Retention): Evaluates the preservation of ancestral relationships.
Purity/NMI: Measures paper-level clustering consistency.

Performance Highlights:

English (TaxoBench): SC-Taxo achieved a state-of-the-art CEDS of 28.7 (outperforming the best baseline by ~5 points) and HSR of 75.3. It also achieved the highest Purity (64.8).
- Note: While NMI was slightly lower than some baselines (50.2 vs 60.1), this was a deliberate trade-off to avoid over-fragmenting papers into meaningless sub-clusters, prioritizing hierarchical coherence.
Chinese Evaluation: SC-Taxo maintained robust performance with a CEDS of 29.8 and HSR of 51.5, significantly outperforming pure generative methods (e.g., GoalEx dropped to CEDS 14.8). This highlights its ability to handle implicit topic boundaries and terminological ambiguity in Chinese.

Ablation Studies:

Removing the dual-path architecture (BU-only) caused a significant drop in CEDS (23.7), proving the necessity of LLM semantic guidance.
Removing bidirectional validation (w/o Bi) reduced CEDS to 25.1, confirming that simple concept injection is insufficient without cross-level constraints.
Removing peer-level expansion (w/o Peer) lowered HSR, showing the importance of horizontal consistency.

5. Significance

Solving Semantic Drift: SC-Taxo effectively addresses the "context drift" problem where LLMs lose global context, ensuring that generated taxonomies are logically coherent and structurally sound.
Interpretability: By enforcing strict parent-child and sibling constraints, the resulting taxonomies are more interpretable and useful for downstream tasks like trend analysis and ontology engineering.
Language Agnosticism: The framework's reliance on a structural scaffold (clustering) combined with semantic refinement allows it to generalize across languages, making it a viable solution for organizing non-English scientific corpora where standardization is often lacking.
Paradigm Shift: The work moves beyond treating LLMs as mere generators or post-hoc labelers, instead integrating them into a structured, iterative refinement loop that respects both statistical topology and conceptual logic.