Emotion labels come in layers. "Nervousness" is a kind of "fear", which is a negative emotion. Annotators also often disagree about which layer, or which neighbor, a sentence belongs to. Current fine-grained emotion classifiers capture the layering by placing each label as a single point in hyperbolic space, a curved geometry with room for tree-like structure. They then treat each sentence as another point and classify it by distance. We propose a different representation. Each sentence becomes a density matrix, a small positive semi-definite complex matrix with unit trace that describes a weighted mixture of directions rather than one direction. Each emotion label becomes a subspace. The model's probability for a label is the share of the sentence's mixture that falls inside that label's subspace. A coarse emotion is the subspace spanned by its finer emotions. Two consequences follow from linear algebra rather than from extra training. The model's predictions across levels of the hierarchy cannot contradict each other. And any new grouping of labels can be scored without training a new output head.
This matters because the gap between levels in current systems is large. On GoEmotions, a fine-tuned BERT reaches an average F1 of 0.46 over 27 categories. When the same data are grouped into Ekman's six emotions the F1 is 0.64, and when grouped by sentiment it is 0.69 (GoEmotions). Each of those numbers comes from a separately trained classifier. The question we test is whether a single model trained only on the 27 fine labels, with the hierarchy expressed as subspace inclusion, can match those coarse-level results, stay competitive with hyperbolic methods at the fine level, and expose annotator ambiguity through the spread of its density matrix.
Core Claim
Let $\rho$ be the utterance's density matrix in $\mathbb{C}^d$. Let $P_c$ be the orthogonal projector, the matrix that keeps only the part of a vector lying in a given subspace, onto label $c$'s learned subspace of rank $r_c$. The predicted probability of label $c$ is
$$p_c = \operatorname{Tr}(\rho P_c).$$
The coarse label for a set of fine labels $S$ uses $P_S$, the projector onto the span of the fine labels' subspaces. Whenever $S \subseteq T$ we have $P_S \preceq P_T$, so $p_S \le p_T$ for every sentence. The hierarchy is therefore consistent by construction. Unlike a classical mixture, $p_S$ is generally neither the sum nor the maximum of the fine probabilities. It depends on how much the fine subspaces overlap and on the off-diagonal entries of $\rho$. Overlapping emotions such as annoyance and anger therefore share probability mass instead of double-counting it.
Here the rank of a projector plays the role that distance from the origin plays in hyperbolic embeddings: broader concepts own bigger subspaces. This gives a concrete answer to a tension noted in the literature. In complex embeddings the amplitude of a vector has been read as meaning, while in hyperbolic embeddings the norm encodes depth in the hierarchy. In our model, generality is carried by subspace dimension, and the norm is freed from that job.
Our central claim is empirical and falsifiable. A density-matrix classifier with learned, possibly overlapping label subspaces, trained on fine labels alone, will match separately trained coarse classifiers on GoEmotions' Ekman and sentiment groupings. It will also stay within noise of the best hyperbolic methods at the fine level. If the off-diagonal (interference) structure does no work, a diagonal-only ablation will match it, and the claim fails.
Relation to Prior Work
Two hyperbolic methods motivate the hierarchy question, and they also supply our main competitor.
- HypEmo embeds labels in the Poincaré ball using Parrott's fixed emotion tree. It weights cross-entropy by the hyperbolic distance between text and label, and reports a 4.8% relative weighted-F1 gain over LCL on GoEmotions and EmpatheticDialogues.
- HyCoEM drops the fixed tree and learns Lorentz-model label points contrastively, gaining 1.3–1.9% accuracy over HypEmo on EmpatheticDialogues. It also shows that collapsing its geometry to a Euclidean one hurts performance.
Both represent every utterance and label as a point, so neither can express that a sentence is partly afraid and partly sad. Neither yields predictions across levels that are consistent by construction. HyCoEM's finding that learned label structure beats a fixed tree constrains our design: we learn the label subspaces rather than fixing them. Parrott's tree is used only to define which fine labels a coarse label spans, and only at evaluation time.
Quantum-inspired work supplies the representation. - End-to-End Quantum-like Language Models builds sentence density matrices analytically as weighted sums of outer products, which keeps them differentiable without iterative estimation. We reuse this construction.
- QMNN applies complex density matrices to conversational emotion. It reports 60.81% accuracy and 58.00% F1 on MELD, and a 1–3% drop when its unitary recurrence is replaced by a GRU. However, its emotion labels are flat categories, and its unitary evolution requires Riemannian optimization.
- QUIET treats sarcasm, sentiment and emotion as incompatible measurements on pure complex states.
We differ from all three by giving the measurement side structure. Label projectors form a nested lattice that mirrors the emotion hierarchy. We also avoid constrained optimization: $\rho$ is built as $AA^\dagger$ normalized by its trace, and projectors are built by orthonormalizing unconstrained matrices.
Two further papers inform the design.
- HYDEN shows that replacing hyperbolic points with distributions improves medical image–report alignment, reaching 0.856 zero-shot AUC on RSNA Pneumonia against 0.792 for the point-based hyperbolic model MERU. That is supporting evidence that mixtures beat points when concepts nest, but in a classical, real-valued setting.
- The Unit Ball Model embeds graphs where nodes can have several parents in complex hyperbolic space, whose curvature varies. It reports 91.95 MAP on Xiphophorus versus about 89.8 for standard hyperbolic models. This is evidence that rigid tree geometry struggles when a node has several parents, as compound emotions do. We address multiple parents differently: a fine subspace can lie inside several coarse spans at once.
Research Program
Encoder. A RoBERTa-base encoder produces token states. A complex projection head, with separate real and imaginary linear maps, turns $k$ attention-pooled vectors into the columns of $A \in \mathbb{C}^{d \times k}$. We then set
$$\rho = \frac{AA^\dagger}{\operatorname{Tr}(AA^\dagger)},$$
which is positive semi-definite with unit trace for any $A$. We plan $d = 64$ and $k \in \{1, 4, 8\}$, so the quadratic cost in $d$ is negligible next to the encoder.
Labels. Each fine label owns an unconstrained matrix $W_c \in \mathbb{C}^{d \times r_c}$. A QR factorization turns it into an orthonormal basis and hence the projector $P_c$. Coarse projectors are computed on demand from the span of the children's bases.
Training. We use binary cross-entropy between $p_c$ and the fine multi-label targets. A small regularizer discourages fine subspaces whose labels never co-occur from overlapping. Gradients for complex parameters are handled by treating real and imaginary parts as separate real parameters, which sidesteps the Wirtinger-calculus friction noted in the complex-embedding literature.
Analyses. Beyond accuracy, we study three things.
- Whether the von Neumann entropy of $\rho$ (how spread out the mixture is) tracks annotator disagreement on the items where several raters' labels can be compared.
- Whether the learned subspace overlaps, measured as $\operatorname{Tr}(P_a P_b)$ normalized by rank, recover the structure of Parrott's tree even though it was never used in training.
- How much of the off-diagonal mass in $\rho$ concentrates on emotion pairs that annotators confuse.
Baselines
Naive reference. RoBERTa-base with 27 independent sigmoid heads. Coarse predictions are obtained by taking the maximum, or alternatively the noisy-OR, over each group's fine outputs. This is the "do nothing" route to hierarchy. The reported reference for the BERT version is 0.46 average F1 on the fine labels, 0.64 on the Ekman grouping and 0.69 on the sentiment grouping (GoEmotions). The coarse numbers come from classifiers trained directly on grouped labels, which makes them the bar our fine-only model must reach.
Strongest existing methods. HyCoEM is the strongest competitor, with HypEmo as the established predecessor. Both directly address fine-grained confusion between related emotions with a hierarchy-aware geometry. HyCoEM reports gains of 1.3–1.9% accuracy and 1–1.7% weighted F1 over HypEmo on EmpatheticDialogues. The file gives only relative improvements, so: no reported absolute result found on Lacuna for HyCoEM or HypEmo weighted F1 on either GoEmotions or EmpatheticDialogues. We will therefore rerun both from their published descriptions under identical conditions: the same RoBERTa-base backbone, the same splits, the same number of epochs, and the same hyperparameter search budget of 20 trials per method on the validation set. Coarse-level scores for these baselines will be obtained with the same max and noisy-OR rules used for the naive reference.
Beating them. We count it as beating the baselines if three things hold together:
- fine-level weighted F1 within one standard deviation across five seeds of the better hyperbolic method;
- zero-shot coarse F1 significantly above max and noisy-OR aggregation, under a paired bootstrap;
- zero violations of hierarchy consistency, where the baselines do have violations under the same aggregation.
Ablations that isolate the key ingredient. - Diagonal-only $\rho$, with label subspaces restricted to disjoint sets of coordinate axes. This makes the model an ordinary classical mixture, so its coarse probabilities become sums of fine ones. It is the decisive ablation. - Real symmetric density matrices of the same size. This tests whether complex phase matters. - Rank-one $\rho$ ($k=1$, a single pure state). This tests whether mixtures matter. - A hyperbolic-point head on the same backbone. This separates the effect of geometry from the effect of the training pipeline.
Why not just…?
Why not just train coarse and fine heads jointly on a Euclidean model and enforce consistency with a max constraint? This gives consistency cheaply, and it is the most serious practical alternative. It has two weaknesses. It needs a head for every grouping known in advance. And its coarse score is a fixed function of the fine scores (the max), so it cannot express that two overlapping fine emotions jointly support their parent less than independent evidence would. The zero-shot grouping experiment tests this directly. If max or noisy-OR aggregation matches our span projectors on the Ekman and sentiment groupings, the subspace mechanism adds nothing beyond consistency, and we would say so.
Why not just use HyCoEM, since hyperbolic geometry already encodes hierarchy? HyCoEM's Euclidean ablation shows that geometry matters at the fine level. But a point cannot be split across sibling emotions, and distances between points do not define a probability for a new grouping of labels. HyCoEM may well remain better at the fine level. The proposal succeeds only if it stays close there while adding coarse-level transfer and calibrated ambiguity. A large fine-level loss would count against us.
Why not just use classical label-distribution learning, with a softmax mixture over fine labels and summed probabilities for parents? This is exactly what the diagonal-only ablation implements. It is a real threat: if it matches the full model, the off-diagonal "interference" story is decoration. Evidence that off-diagonal structure helps exists elsewhere, for example QMNN's 1–3% drop when its complex components are replaced. But that evidence comes from a different architecture and task, so this proposal has to settle the question itself.
Proposed Experiments
GoEmotions is the primary benchmark. We train on the 27 fine categories plus neutral using the standard splits, and report macro and weighted F1 at the fine level. The two decisive checks are zero-shot Ekman-grouped and sentiment-grouped F1, computed from span projectors with no grouped supervision. These are compared against the reported 0.64 and 0.69 and against the aggregation baselines. We also count violations of hierarchy consistency for every system.
EmpatheticDialogues, with 32 labels, is the second fine-grained test, run against rerun HypEmo and HyCoEM with accuracy and weighted F1. We also run a held-out-grouping test. We define coarse groupings by Parrott's tree, train on fine labels only, and score each coarse group zero-shot. Then we ask whether the learned subspace overlaps reproduce the tree, using the rank correlation between pairwise overlap and tree distance.
Ambiguity. On GoEmotions items with multiple raters, we correlate the entropy of $\rho$ with the rate of rater disagreement, and compare against the entropy of the naive baseline's normalized sigmoid outputs. The goal is to show that spread in the mixture carries information about ambiguity that a point model's confidence scores do not.
Optional extension. A conversational transfer to MELD, where QMNN's 60.81% accuracy and 58.00% F1 give a concrete target for a density-matrix model, will be attempted only if the GoEmotions results hold.
Decisive outcome. The full model must beat the diagonal-only ablation and max aggregation on zero-shot coarse F1 across five seeds, while losing little at the fine level to HyCoEM. If the diagonal ablation ties, we conclude that hierarchical consistency, not interference, explains any gains, and we recommend the simpler classical model.
References
- GoEmotions: A Dataset of Fine-Grained Emotions
- Label-Aware Hyperbolic Embeddings for Fine-grained Emotion Classification
- Semantic alignment in hyperbolic space for fine-grained emotion classification
- End-to-End Quantum-like Language Models with Application to Question Answering
- Quantum-inspired Neural Network for Conversational Emotion Recognition
- A Quantum Probability Driven Framework for Joint Multi-Modal Sarcasm, Sentiment and Emotion Analysis
- HYDEN: Hyperbolic Density Representations for Medical Images and Reports
- Unit Ball Model for Hierarchical Embeddings in Complex Hyperbolic Space