When a text classifier is trained on a main task alongside a helper task, the helper is usually switched on or off on a schedule: kept constant, decayed, or phased out. The best schedule is fragile. It changes with dataset size, it must be found by grid search, and it rarely transfers to a new dataset. This note argues that in multi-label emotion classification, where one sentence can carry several emotions at once, much of this fragility comes from a mismatch. Each emotion label behaves like its own task, with its own amount of data and its own relationship to the helper task, yet a single global schedule serves all of them. A helper that rare emotions like grief need throughout training may start to distract frequent emotions like approval after a few epochs. The global schedule can only compromise between the two.
We propose to give each emotion label its own auxiliary weight and to set it from a cheap training signal: whether the helper task's gradient for that emotion still points the same way as that emotion's classification gradient. The helper we use is emotion-definition modeling, which already exists in per-label form. We then test the claim that matters most for the curriculum question: that per-label pacing does not just raise the mean score but makes results much less sensitive to schedule hyperparameters, and lets a configuration tuned on one dataset transfer to others without retuning.
Core Claim
The main evidence comes from Scheduled Multi-Task Learning: From Syntax to Translation. It controls how often a syntax helper task is sampled during translation training and finds that exponential decay wins on large data, while a constant schedule wins on small data (+2.7 BLEU on low-resource WIT). Label-Similarity Curriculum Learning appears to contradict this. It anneals soft label targets into hard ones and gains most in the scarcest regimes. Our reading is that "resource level" is not a property of the dataset in multi-label emotion. GoEmotions contains both data-rich and data-starved labels. So the schedule question should be answered per label, and a global schedule's sensitivity is the variance produced by averaging over labels that want opposite things.
Fine-Grained Emotion Prediction by Modeling Emotion Definitions supplies the testbed. It trains BERT jointly on GoEmotions classification and Class Definition Prediction (CDP), in which the model judges whether a candidate emotion definition matches the sentence. It uses a fixed sampling ratio throughout training and reports test F1 rising from 46% to 52.34%, with grief doubling from 25% to 50%. Because CDP is posed per emotion class, its loss can be split by class and weighted per class without any architectural change. That paper motivates the idea (the helper clearly helps rare labels) but never tests whether it hurts frequent ones late in training. That untested question is our first measurement.
Research Program
Let $g^{\text{cls}}_k$ be the gradient of emotion $k$'s binary cross-entropy loss with respect to the top encoder layers. Let $g^{\text{aux}}_k$ be the gradient of the CDP loss restricted to pairs involving emotion $k$'s definition. Every $N$ steps, on a small probe batch, we compute their cosine similarity. We keep an exponential moving average $a_k$ of that cosine and set emotion $k$'s auxiliary weight as
$$w_k = \mathrm{clip}\!\left(a_k / \tau,\ 0,\ 1\right),$$
where $\tau$ is a single scale constant shared across labels. The total loss is the sum of the per-label classification losses plus $\sum_k w_k \mathcal{L}^{\text{aux}}_k$. Under this rule, a label keeps its helper while the two gradients agree and loses it once they conflict. The rule stays per label and schedule-free: no start proportion, no decay rate, no transition step. It still has hyperparameters ($\tau$, the averaging rate, and $N$). We do not claim zero hyperparameters. We claim the outcome is flat across them, and we measure that flatness directly.
The program has three stages. The first is diagnostic. With Singh et al.'s static CDP setup, we log per-label cosine over training. We then run global constant, exponential, and sigmoid schedules in the style of Kiperwasser and Ballesteros, and record per-label F1 under each. The hypothesis fails if every label prefers the same schedule, or if cosine trajectories do not differ systematically between rare labels and labels that are frequent or easily confused. The second stage is the method above, plus a gradient-free variant that sets $w_k$ from label frequency alone. The third stage is the sensitivity and transfer study, described under experiments.
We restrict gradients to the top layers for cost reasons, a choice also made by RotoGrad: Gradient Homogenization in Multi-Task Learning, which works in a reduced subspace to stay tractable. We will report wall-clock overhead and check on a subsample that top-layer cosines track full-model cosines.
Baselines
The naive reference is BERT-base with independent per-emotion binary cross-entropy heads and no helper task. On GoEmotions, GoEmotions: A Dataset of Fine-Grained Emotions reports an average F1 of 0.46 across the 27 categories. We rerun it with our seeds and thresholds so that every comparison uses the same evaluation code.
The strongest competitor is static BERT+CDP from Fine-Grained Emotion Prediction by Modeling Emotion Definitions, with reported GoEmotions test F1 of 52.34% and grief F1 of 50%. It is the right competitor because it uses exactly our helper task, and our method changes only when and for which labels that helper is active. We reproduce it and give its sampling ratio the same number of tuning trials we give our $\tau$. Beating it means matching or exceeding its overall F1 while doing two further things: showing a statistically significant gain on frequent labels, where we predict static CDP causes late-training distraction, without losing its rare-label gains; and showing a markedly narrower spread of scores across its tuning grid.
The second strongest competitor is global scheduling from Scheduled Multi-Task Learning: From Syntax to Translation. We port its constant, exponential, and sigmoid samplers to CDP, with each schedule's rate and midpoint tuned under the matched budget. This baseline isolates whether the per-label part of our method matters or whether any schedule would do.
The third is Conflict-Averse Gradient Descent for Multi-task learning. CAGrad treats classification and CDP as two tasks and resolves their gradient conflict at every step. It reduced the per-task drop on NYU-v2 to 0.20%, against 3.97% for PCGrad, though it has never been tested on text emotion. Its conflict radius $c$ is tuned under the same budget. It represents the claim that global conflict handling suffices without per-label decisions.
The ablation that isolates our key ingredient is the frequency-bucket variant, which sets $w_k$ from each label's training count through a fixed monotone map instead of from gradient agreement. The frequency variant is the ablation that carries most weight.
For the secondary benchmark we use SpanEmo, which on SemEval-2018 Task 1 E-c English reports Micro F1 0.713 and Jaccard 0.601 (SpanEmo: Casting Multi-label Emotion Classification as Span-prediction). Our method is applied on top of SpanEmo's input format with CDP added, so the comparison is SpanEmo against SpanEmo plus paced CDP.
Why not just…?
Why not just tune the static sampling ratio carefully? This is the most direct threat. Singh et al.'s 52.34% already comes from a fixed ratio, and Text Emotion Distribution Learning via Multi-Task Convolutional Neural Network found a single good static weight ($\lambda = 0.7$) for its joint loss. If a well-tuned constant matches us on mean F1, that alone does not settle the question. What settles it is the per-label breakdown and transfer regret. A single ratio cannot be optimal for labels that the diagnostic shows want opposite treatment. If the diagnostic finds no such divergence, the premise is wrong, and we will report that rather than push the method.
Why not just use a gradient-conflict method like CAGrad and skip schedules entirely? CAGrad resolves conflict between the averaged classification gradient and the averaged helper gradient. Conflict on rare labels can be hidden in that average by agreement on frequent labels, or the reverse. The result in When Attributes Disagree: Gradient Conflict in Image Aesthetic Assessment points the same way: failures concentrate in rare attribute combinations whose gradients are overwhelmed by majority patterns. Still, this is an empirical question, and the matched-budget CAGrad baseline answers it. If CAGrad matches per-label pacing on rare and confusable labels, the per-label claim loses.
Why not just schedule by label frequency, without any gradient signal? This alternative could plausibly work as well, and we treat it as a real candidate outcome. Our prediction is that frequency explains most of the rare-label behavior. It should fail on labels that are frequent but semantically confusable, such as anger versus disgust or sadness versus fear, where definitions may keep helping long after data is plentiful. It should also need re-thresholding on a new dataset whose label counts differ. If the frequency variant ties on GoEmotions and on transfer, the honest conclusion is that the simple heuristic is the contribution, and we would publish it as such.
Why not just reweight rare classes in the classification loss? The rare-label gains from CDP might come from extra gradient pressure on rare classes, not from definition semantics. We therefore include a control that reweights per-class binary cross-entropy by inverse frequency with no helper task. If it recovers grief-level gains, the helper task is not doing what Singh et al. suggest.
Proposed Experiments
The main benchmark is GoEmotions, with 28 classes and a severe long tail. We report macro F1, micro F1, and per-label F1 grouped into frequency terciles, plus a pre-registered confusable set (anger, disgust, sadness, fear, and nervousness). Secondary benchmarks are SemEval-2018 Task 1 E-c (English) and EthioEmo, where Evaluating the Capabilities of Large Language Models for Multi-label Emotion Understanding reports weighted F1 of 72.6% for Afan Oromo and 69.9% for Amharic with fine-tuned AfroXLMR-76L. On EthioEmo we use that encoder with English definitions of Ekman's six emotions. Every configuration runs five seeds.
The first decisive check is sensitivity. For each method, we sample 20 configurations from its own hyperparameter space: sampling ratio for static CDP, rate and midpoint for the global schedules, $c$ for CAGrad, and $\tau$ and averaging rate for ours. We compare the interquartile range and the worst-decile macro F1. Our claim is supported only if per-label pacing has a clearly narrower spread than the global schedules and static CDP at comparable median. A higher best score alone does not count.
The second check is transfer regret. We tune each method on GoEmotions, freeze its hyperparameters, and train on SemEval-2018 E-c and on the two EthioEmo languages. We then measure the gap between that score and each method's own per-dataset oracle tuning. This also bears directly on the SMTL-versus-LCL tension. If the same frozen configuration works on low-resource EthioEmo, where Kiperwasser and Ballesteros would predict a constant schedule is needed, then per-label pacing is absorbing the resource effect as we hypothesize.
The third check is mechanistic. We correlate each label's final $w_k$ trajectory with the per-label F1 difference between the constant and decayed global schedules from stage one. A strong correlation means the gradient signal is detecting the per-label preference we claim exists. A null correlation means any gains come from somewhere else.
References
- Scheduled Multi-Task Learning: From Syntax to Translation
- Label-similarity Curriculum Learning
- Fine-Grained Emotion Prediction by Modeling Emotion Definitions
- RotoGrad: Gradient Homogenization in Multi-Task Learning
- GoEmotions: A Dataset of Fine-Grained Emotions
- Conflict-Averse Gradient Descent for Multi-task learning
- SpanEmo: Casting Multi-label Emotion Classification as Span-prediction
- Text Emotion Distribution Learning via Multi-Task Convolutional Neural Network
- When Attributes Disagree: Gradient Conflict in Image Aesthetic Assessment
- Evaluating the Capabilities of Large Language Models for Multi-label Emotion Understanding