QuanTaxo: A Quantum Approach to Self-Supervised Taxonomy Expansion

TL;DR

QuanTaxo introduces a quantum-inspired approach to taxonomy expansion by embedding entities in a complex Hilbert space and using density-matrix representations to capture hierarchical polysemy. It combines a complex-valued entity projector, a quantum representation module, and a joint representation for scoring anchor candidates, trained with self-supervised data from seed taxonomies. Across five benchmarks and nine strong baselines, QuanTaxo achieves substantial improvements in accuracy ($12.3\%$), MRR ($11.2\%$), and Wu\&Palmer ($6.9\%$), with ablations validating the benefits of complex embeddings and entangled features. The work demonstrates the practical potential of quantum-inspired semantics for scalable, accurate taxonomy modeling and expansion in dynamic knowledge graphs.

Abstract

A taxonomy is a hierarchical graph containing knowledge to provide valuable insights for various web applications. However, the manual construction of taxonomies requires significant human effort. As web content continues to expand at an unprecedented pace, existing taxonomies risk becoming outdated, struggling to incorporate new and emerging information effectively. As a consequence, there is a growing need for dynamic taxonomy expansion to keep them relevant and up-to-date. Existing taxonomy expansion methods often rely on classical word embeddings to represent entities. However, these embeddings fall short of capturing hierarchical polysemy, where an entity's meaning can vary based on its position in the hierarchy and its surrounding context. To address this challenge, we introduce QuanTaxo, a quantum-inspired framework for taxonomy expansion that encodes entities in a Hilbert space and models interference effects between them, yielding richer, context-sensitive representations. Comprehensive experiments on five real-world benchmark datasets show that QuanTaxo significantly outperforms classical embedding models, achieving substantial improvements of 12.3% in accuracy, 11.2% in Mean Reciprocal Rank (MRR), and 6.9% in Wu & Palmer (Wu&P) metrics across nine classical embedding-based baselines.

Figures (7)

• Figure 1: An illustration of taxonomy expansion using QuanTaxo on the Environment seed taxonomy, where the query terms Soil Pollution and Used Oil are to be inserted.
• Figure 2: An illustration of the QuanTaxo framework. Given parent and query entities, their complex representations are generated, followed by outer products on the word and sentence embeddings to generate quantum representations. The joint representation of parent and child entities is then computed to score the semantic relationship.
• Figure 3: Performance comparison of real and complex-valued embeddings across 'Env', 'Sci' and 'PSY'.
• Figure 4: Performance comparison of direct and weighted sum in real-valued and complex-valued mixture models across 'Env', 'Sci', and 'WIKI' benchmarks.
• Figure 5: Comparison of parent-child joint representation and concatenation for scoring on 'Env' and 'Sci'. 'Sup' $\rightarrow$ superposition module, while 'Mix' $\rightarrow$ mixture module. $\mathbb{R}$$\rightarrow$ real embedding while $\mathbb{C}$$\rightarrow$ complex embedding.

Theorems & Definitions (2)

• Definition 1
• Definition 2