The work, authored by Sayak Mondal of the School of Computer Science and Engineering at Vellore Institute of Technology in India, begins from an observation that will resonate with anyone who has worked with fine-grained emotion datasets. Real-world semantic structures, particularly those involving entangled emotional labels, exhibit anisotropic correlations and non-uniform separability — technical terms for the fact that meaning is not spread evenly in every direction. A sentence expressing both grief and gratitude, for example, does not sit neatly inside a single emotional category; it occupies a region of semantic space where multiple labels overlap and blur. Flat Euclidean geometry, with its uniform distances in all directions, struggles to capture such tangled structure. Mondal’s answer is to abandon the flat plane and work instead on a curved, deformable hypersphere governed by the mathematics of Riemannian geometry.

The framework unfolds in three stages, each with a distinct mathematical role. First, sentence embeddings produced by standard encoders are mapped into a learned hyperspace, a high-dimensional representation whose geometry is not fixed in advance but shaped during training. Second, the model performs what the paper describes as a Mahalanobis-style metric deformation. The Mahalanobis distance, a classical tool from statistics, generalizes ordinary Euclidean distance by allowing different directions in space to be weighted differently — effectively letting the model stretch the space along some axes and compress it along others. In HyperSpectrum Geometry, this deformation is controlled by a learnable global Riemannian metric tensor, meaning the curvature of the semantic space itself becomes a trainable parameter rather than a fixed property of the architecture.

The third stage carries out angular decision partitioning on a normalized, deformed hypersphere. After the metric deformation, embeddings are projected onto a hyperspherical manifold — the curved surface of a sphere embedded in high dimensions — where classification decisions are made based on angles rather than raw distances. This angular approach draws on ideas popularized in face recognition research, notably additive angular margin losses such as ArcFace, which showed that forcing class representations to occupy distinct angular regions of a sphere can dramatically sharpen decision boundaries. By combining angular partitioning with a deformable metric, HyperSpectrum Geometry effectively gives the model two independent dials: it can reshape the space to reflect anisotropic semantic correlations, and it can carve the resulting curved surface into angular sectors corresponding to different classes.

To test whether this geometric machinery actually matters, the study evaluated the framework on five structurally distinct datasets: GoEmotions, a collection of Reddit comments annotated with fine-grained emotion labels; AG News and DBpedia Classification, both of which organize text into well-separated topical categories; TweetEval Emotion, a benchmark for classifying emotional content in tweets; and Amazon Reviews Multi, a multilingual sentiment corpus. These datasets were chosen deliberately, because they span a spectrum of structural complexity. Topic and ontology datasets such as AG News and DBpedia feature classes that are largely disjoint — a story about sports is rarely also a story about science — whereas fine-grained emotion datasets are notorious for multi-label entanglement, where a single text can plausibly carry several overlapping emotional signals at once.

The empirical results reveal a striking pattern that the paper terms a deformation–separability relationship. On the fine-grained, multi-label emotional datasets, the model learned substantially higher global metric deformation, indicating that it needed to bend the semantic space aggressively to separate entangled categories. On the structurally separable topic and ontology datasets, the opposite occurred: classification relied more strongly on angular partitioning, with comparatively little metric deformation required. In other words, the geometry the model chose to learn was not arbitrary — it tracked the intrinsic difficulty and structure of the task. When classes overlap and blur, the space must be curved and stretched to pull them apart; when classes are naturally distinct, simple angular separation on a gently deformed sphere suffices.

This pattern carries implications that extend beyond benchmark scores. It suggests that the amount of metric deformation a trained model exhibits could serve as a diagnostic signal — a measurable fingerprint of how entangled a dataset’s label structure really is. The study also reports several interpretable geometric quantities computed over multiple random seeds, including the global metric deformation, the condition number of the learned metric tensor, and intra-class versus inter-class angular similarity. The condition number, a concept from linear algebra, quantifies how anisotropic the learned metric is: a value near one indicates a nearly uniform space, while larger values reveal strong directional stretching. Because these diagnostics remained stable across seeds, they offer a reproducible lens on representation learning that most black-box classifiers cannot provide.

Stability and efficiency were recurring themes in the evaluation. The author emphasizes that the model delivers lightweight, consistent performance across repeated runs with different random seeds, a nontrivial achievement given that learned metric tensors can be notoriously sensitive to initialization. The framework’s interpretability is presented not as an afterthought but as a core contribution: rather than treating the geometry as an opaque intermediate layer, the paper argues that learnable metric deformation constitutes a useful and underexplored component of interpretable semantic representation learning. In an era when concerns about the opacity of large language models dominate public discourse, approaches that make their internal geometry legible — and that tie that geometry to observable properties of the data — occupy valuable middle ground between pure performance engineering and post-hoc explanation.

The study also situates itself within a broader movement known as geometric deep learning, which has spent the past decade arguing that Euclidean spaces are only one member of a much larger family of possible representational geometries. Hyperbolic embeddings, for instance, have been used to capture hierarchical structures such as taxonomies, where tree-like data can be embedded far more compactly in negatively curved space than in flat space. Graph neural networks have likewise been adapted to non-Euclidean domains. HyperSpectrum Geometry contributes a complementary perspective: instead of committing to a single fixed curvature, it learns the curvature adaptively, letting the data dictate how much deformation is needed. This curvature-adaptive stance may prove especially relevant as language models are increasingly deployed on tasks — emotion recognition, mental-health screening, nuanced content moderation — where label boundaries are genuinely fuzzy rather than artificially clean.

Caveats remain, as with any single-author empirical study. The paper reports that implementation details, training scripts, random seeds, and hyperparameter settings will be made available as supplementary material or through a public repository on request, and independent replication will be needed to establish how broadly the deformation–separability pattern generalizes to other architectures, languages, and domains. The framework was also evaluated with specific encoder backbones, and the interaction between learned metric deformation and the geometry inherited from large pretrained transformers remains an open question. Nevertheless, the central insight stands on its own merits: the geometry of a model’s semantic space is not a fixed design choice but a learnable quantity, and the curvature a model discovers can tell us something honest about the structure of the problem it faces. As text classification systems are asked to navigate ever more nuanced shades of human meaning, frameworks like HyperSpectrum Geometry suggest that the future of machine understanding may depend less on bigger models than on giving them the freedom to bend the spaces in which they think.

Subject of Research: Curvature-adaptive Riemannian hyperspherical representation learning for neural text classification

Article Title: HyperSpectrum geometry: a Riemannian hyperspherical framework for curvature-adaptive text classification

Article References: Mondal, S. (2026). HyperSpectrum geometry: a Riemannian hyperspherical framework for curvature-adaptive text classification. Neural Computing and Applications, 38(17), Article 710. https://doi.org/10.1007/s00521-026-12332-4

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12332-4

Keywords: Riemannian metric learning, hyperspherical classification, geometric deep learning, Mahalanobis distance, curvature-aware representation learning, text classification, GoEmotions, semantic embeddings, angular margin loss, interpretable machine learning, natural language processing, Neural Computing and Applications

The framework, called CurvBi, short for Curvature-aware Bicomplex, is presented by Thanh Le, Vy Cao, Duong-Khang Au, Trong-Nghia Pham and Bac Le of the University of Science and Vietnam National University in Ho Chi Minh City. Writing in the journal Data Mining and Knowledge Discovery, the authors tackle the task known as multimodal knowledge graph completion: predicting missing links and entities in a graph by exploiting both its symbolic structure and the complementary information carried by textual and visual descriptions of its nodes. Their central insight is that the geometry used to fuse different modalities matters as much as the algebra used to model relationships between them.

To understand why, it helps to look at what existing approaches do. Most multimodal knowledge graph embedding methods treat the problem in two loosely connected stages. First, entity descriptions in each modality are projected into a shared Euclidean vector space, typically through simple concatenation, linear projection or a shallow fusion layer. Then, a relational model such as TransE, RotatE or a quaternion-based embedding scores candidate triples by comparing the vectors. This pipeline is convenient, but it flattens away the very structure that distinguishes modalities from one another. Text embeddings and image embeddings live on manifolds with different intrinsic curvature and different statistical character, and forcing them into a single flat space can blur or destroy the discriminative information they carry.

CurvBi departs from this convention on two fronts simultaneously. On the relational side, it adopts bicomplex embeddings, a representation built in a four-dimensional algebraic space formed by pairing two complex planes into a single number system. Bicomplex numbers are an extension of the complex numbers in which both the real and imaginary components are themselves complex, and their algebra is expressive enough to encode symmetric, asymmetric, inverse and cyclic relationships within one unified scoring function. This matters because real knowledge graphs contain all of these relational patterns at once: a relation like ‘is married to’ is symmetric, ‘is the parent of’ is asymmetric and invertible, and ‘is located in’ often participates in cyclic chains. Prior systems have typically needed separate mechanisms, or careful inductive-bias engineering, to cover this full catalogue of relation types.

The bicomplex formulation builds on a line of work the group has developed over several years, including quaternionic and bicomplex embeddings for temporal knowledge graphs and event prediction. In CurvBi, the four-dimensional algebraic structure plays a second, subtler role: it provides a natural home for the modality embeddings before they are fused. Rather than concatenating a text vector and an image vector end to end, CurvBi treats each modality-specific representation as a structured component within the bicomplex number, preserving the algebraic relationships between components while still allowing the model to learn how much weight each modality deserves for a given entity or relation.

The second front is geometric. Instead of projecting all modality embeddings into flat Euclidean space, CurvBi performs fusion on a Riemannian manifold, measuring the distance between modality-specific representations with the Fisher-Rao distance, a well-studied metric from information geometry that treats probability distributions as points on a curved manifold. Because the Fisher-Rao metric is curvature-aware, similarities between a text embedding and an image embedding are judged in a geometry that reflects their statistical structure rather than their raw coordinate differences. The result, according to the authors, is a fusion step that is geometry-sensitive in a way that Euclidean projection and shallow concatenation simply cannot be, aligning heterogeneous modalities while respecting the distinct manifold each of them inhabits.

This curvature-aware component connects CurvBi to a broader trend in machine learning. Hyperbolic and other non-Euclidean embedding spaces have been shown to represent hierarchies, cycles and scale-free structures far more efficiently than Euclidean space, and methods such as Poincaré embeddings and multicurvature adaptive schemes for temporal knowledge graphs have demonstrated measurable gains. What distinguishes CurvBi is the combination: the same framework unifies the algebraic machinery for relational reasoning and the differential-geometric machinery for modality alignment, so that neither stage of the pipeline is an afterthought bolted onto the other. The manifold structure informs how modalities are brought together, while the bicomplex algebra governs how the fused representation is used to score and complete triples.

The empirical case for the framework rests on four benchmark datasets for multimodal knowledge graph completion. Across all of them, CurvBi achieved consistently competitive results on standard evaluation metrics, including Hits at 1, Hits at 10 and mean reciprocal rank, which measure how often the correct missing entity appears at the top of the model’s ranked predictions. On certain benchmark settings the improvements were substantial: gains of up to 6 percent in Hits at 1 and 4.5 percent in mean reciprocal rank over state-of-the-art baselines. The authors also report that the framework scales well and generalizes across datasets with different sizes, relational vocabularies and modality configurations, an important practical consideration given that real multimodal knowledge graphs vary enormously in how much image and text data they attach to each entity.

The significance of the work extends beyond a leaderboard. Multimodal knowledge graphs are increasingly used in settings where text alone is ambiguous: in biomedical informatics, where molecular structures and microscopy images supplement gene and disease ontologies; in e-commerce, where product images carry information absent from catalog text; and in commonsense reasoning systems, where visual grounding helps language models resolve references. If the fusion geometry is wrong, the extra modalities can actively hurt, a phenomenon earlier studies have documented when visual context proved unhelpful for certain graph completion tasks. CurvBi’s results suggest that a principled geometric treatment of fusion, paired with an algebraically expressive relational model, is one way to make multimodal information reliably additive rather than subtractive.

The research, funded by the Vietnam National Foundation for Science and Technology Development under grant number 102.05-2025.75, also signals a maturing of geometric deep learning as applied to structured knowledge. Rather than treating curvature, complex numbers and manifold optimization as exotic mathematical decorations, the authors deploy them as load-bearing components of a system that demonstrably outperforms flatter alternatives. As knowledge graphs continue to absorb richer and more heterogeneous data streams, the lesson from CurvBi is likely to resonate: the shape of the space in which artificial intelligence reasons is not a technicality, but a first-class design decision that determines what that intelligence can and cannot see.

Subject of Research: Curvature-aware multimodal knowledge graph completion using bicomplex embeddings and Riemannian manifold fusion

Article Title: Learning curvature-aware multimodal representations with bicomplex embeddings

Article References: Le, T., Cao, V., Au, D.-K., Pham, T.-N., & Le, B. (2026). Learning curvature-aware multimodal representations with bicomplex embeddings. Data Mining and Knowledge Discovery, 40(5), Article 89. https://doi.org/10.1007/s10618-026-01257-0

Image Credits: AI Generated

DOI: 10.1007/s10618-026-01257-0

Keywords: knowledge graph completion, multimodal embeddings, bicomplex numbers, Riemannian geometry, Fisher-Rao distance, modality fusion, link prediction, representation learning, curvature-aware learning, quaternion algebra, data mining, artificial intelligence