taxonomy-constrained learning

This work addresses the limitations of existing concept-based models, which rely on fine-grained annotations and treat concepts as flat, independent units, thereby hindering the construction of interpretable, hierarchical concept structures. To overcome this, the authors propose Multi-Level Concept Segmentation (MLCS) and Deep Hierarchical Concept Embedding Models (Deep-HiCEMs), which require only coarse-grained top-level supervision to automatically discover multi-layered, human-interpretable concept hierarchies. The framework supports concept interventions across abstraction levels and successfully uncovers novel, explainable concepts absent from training data across multiple benchmarks. While maintaining high predictive accuracy, the method significantly enhances task performance through test-time interventions, marking the first approach capable of automatically constructing a multi-granular concept system from coarse-grained labels alone.

To address the limitation of prior-dependent constraints undermining generalizability in hierarchical multi-label classification (HMC), this paper proposes EDR, a prior-free error-driven constraint discovery framework. EDR automatically identifies model misprediction patterns to induce interpretable, structured logical constraints; it further integrates a constraint-driven post-processing mechanism with a neuro-symbolic joint modeling architecture to jointly perform error detection, constraint recovery, and multi-level consistency verification. For the first time, EDR achieves fully automated, interpretable knowledge discovery and robust cross-domain constraint transfer without predefined constraints—enabling effective constraint learning even under label noise. Evaluated on multiple public benchmarks and a newly constructed military vehicle recognition dataset, EDR achieves an error detection F1-score exceeding 0.89 and constraint recovery accuracy above 92%, significantly improving both hierarchical consistency and overall classification performance.

Large language models (LLMs) struggle to represent unipolar categorical concepts (e.g., “is an animal”) due to the absence of natural antonyms and inherent nonlinearity in their embeddings. Method: We propose the Extended Linear Representation Hypothesis—modeling unipolar categories as direction vectors and their semantic scope as convex polyhedra—and establish the first rigorous mathematical mapping between conceptual hierarchy depth and representation geometry (i.e., decreasing inter-vector angles and polyhedral containment). Using WordNet, we construct a hierarchical concept set of 900+ categories and estimate direction vectors in Gemma and LLaMA-3, followed by joint convex-geometric and semantic analysis. Results: Empirically, deeper-category vectors exhibit significantly smaller pairwise angles; moreover, parent-category polyhedra approximately contain those of their children. These findings confirm that LLMs internally organize semantic categories via interpretable geometric structures—offering a novel paradigm for interpretable and hierarchically grounded semantic representation in LLMs.

To mitigate catastrophic forgetting in prompt-based continual learning, this paper proposes a cognitively inspired confusion-aware method. First, it constructs a dynamic hierarchical label tree to model inter-class semantic relationships. Second, it leverages optimal transport to analyze the implicit knowledge structure of pretrained models and identify fine-grained discriminative regions prone to confusion. Finally, within a prompt-tuning framework, it introduces a novel class-relation-aware regularization loss that explicitly constrains prompt vector updates in confusion-sensitive regions. This work is the first to jointly integrate a dynamic hierarchical taxonomy with optimal-transport-driven knowledge relationship mining into prompt tuning. Extensive experiments on multiple continual learning benchmarks demonstrate that the proposed method significantly outperforms state-of-the-art approaches, effectively alleviating forgetting and improving overall accuracy across both old and new tasks.

Addressing the challenge of multi-label automatic annotation for large-scale, hierarchical classification systems in software requirements engineering, this study proposes a sentence-level zero-shot classification paradigm to circumvent the high annotation costs associated with supervised training. We introduce the first industrial-scale requirements annotation benchmark comprising 769 taxonomy labels and systematically demonstrate a strong negative correlation between the number of taxonomy leaf nodes and classification recall. We further propose a zero-shot multi-label classification method leveraging SBERT sentence embeddings, achieving significant improvements in recall. Empirical evaluation reveals that hierarchical strategies yield no consistent performance gain across settings. Our work validates the effectiveness and feasibility of zero-shot learning for large-scale requirements classification, offering a scalable, low-human-effort automation solution for requirements tracing. (138 words)

This work addresses the inconsistent performance of large multimodal models in hierarchical visual recognition due to their lack of explicit taxonomic knowledge. To this end, the authors propose HiR², a plug-and-play hierarchical representation regularization method that introduces, for the first time, a dual-objective regularization mechanism into multimodal models. Specifically, it constructs a semantics-aware visual hierarchy via hyperbolic entailment cones embedded in Lorentz space to preserve radial hierarchical structure, while simultaneously imposing an angular separation loss on the unit hypersphere to enhance the discriminability of semantically similar embeddings. By effectively fusing intermediate linguistic features with semantics-guided visual representations, HiR² significantly improves hierarchical recognition consistency across various state-of-the-art multimodal architectures and fine-tuning paradigms, while better capturing taxonomic structures across tasks.

This study addresses the limitation that classification trees generated by generative AI, despite appearing plausible, often suffer from leaf-node redundancy and cross-branch leakage, thereby lacking practical operational utility. To overcome the constraints of conventional local naming checks, this work proposes a holistic tree-level evaluation framework incorporating a structural discriminator and a team partitionability metric. Furthermore, an agent-based framework is designed to process proprietary customer feedback corpora, combining statistical analysis with semantic deduplication for multidimensional quantitative assessment. Experimental results reveal that 97.7% of generated leaf nodes duplicate ancestor names, accompanied by severe cross-branch leakage, demonstrating that surface plausibility alone cannot guarantee taxonomy quality. These findings establish a new paradigm for evaluating the practical utility of classification systems.

本文针对大规模概念集分类体系构建问题，提出SPARROW框架，通过结构保留的分区和约束引导的合并方法，有效解决了现有方法在规模扩大时性能下降的问题。

This work addresses the hierarchical inconsistency problem in multi-level visual classification, where fine-grained predictions often conflict with their parent categories. To mitigate this issue, the authors propose a hierarchy-constrained contrastive learning mechanism that performs contrastive optimization exclusively within the same semantic level, thereby eliminating interference from cross-level false negatives. Additionally, a group-balanced optimization strategy is introduced to ensure adequate training across all hierarchy levels. Built upon the BioCLIP framework, the method jointly optimizes representations in both Euclidean and hyperbolic spaces, significantly improving hierarchical consistency and classification performance. Evaluated on benchmarks including iNaturalist 2021, the approach achieves a 30.47% average improvement in cross-level accuracy over baseline methods and demonstrates notably enhanced consistency under zero-shot settings.

研究探讨了在冻结的DINOv2特征上进行严格显式分类图像检索时，分类学对齐、目标选择和几何选择（欧几里得-双曲几何）对层次检索性能的影响。