Writing the Missing Middle: Synthetic Blended-Emotion Text for Low-Data Intensity Prediction

Most text emotion datasets treat feelings as one discrete category per sentence (this one is "anger," that one is "fear"). People rarely write that way. A complaint can be mostly angry, a little afraid, and faintly sad at once. Benchmarks that score emotion intensity per category, with several annotators per sentence, show this blending directly. When labeled data is scarce, as it is for most of the 30-plus languages in the SemEval-2025 emotion task, these blended and low-intensity sentences are exactly what a small training set lacks. Its few examples cluster around clear, single-emotion cases. We propose to ask a large language model to write new training sentences aimed at specific blends: target intensity vectors over several emotions that sit in thinly populated parts of the label space. We then relabel each sentence according to what it actually expresses, not according to what the prompt asked for, and keep only the sentences whose text carries real information about that label.

The two-sentence version is this. Generative augmentation for text has so far been used to add more examples of discrete classes. Emotion distribution learning in low-data settings has so far only reshuffled the real examples it already had. We generate text into the ambiguous interior of the emotion space, then filter and relabel it with a continuous version of an information-based data-valuation score. Our prediction is specific and falsifiable. The gains should concentrate on test sentences where human annotators disagree, and they should shrink as real data grows.

Core Claim

In low-data emotion intensity prediction, the useful part of synthetic text is the coverage it adds in blended regions of the label space. That coverage only helps if each synthetic sentence's label is re-estimated rather than copied from the generation prompt. The claim splits into two separately testable parts. The first is about where to generate: targets upweighted toward under-covered blends versus targets matched to the seed label distribution. The second is about how to label: the prompt's target versus a re-estimated soft label passed through a continuous information filter.

The closest work sits on two sides of this claim. Selective In-Context Data Augmentation for Intent Detection using Pointwise V-Information generates few-shot training utterances with OPT-66B and filters them with pointwise V-information (PVI). PVI measures how much seeing the input makes a model more confident in the correct label, compared with a model that sees no input. That method works with single, mutually exclusive intent labels. It raised BANKING77 accuracy from 79% to 84.01%, but it has no notion of a label distribution or of ambiguity. On the other side, Text Emotion Distribution Learning from Small Sample: A Meta-Learning Approach learns emotion distributions from 10% of SemEval-2007 headlines. It builds meta-learning tasks from nearest-neighbor clusters of real sentences and never synthesizes new ones. We take the generation-plus-valuation recipe from the first paper and the distributional target from the second. The new parts are generation conditioned on intensity blends and a filter defined over distributions.

Research Program

Choosing targets. Each training sentence carries a per-emotion intensity vector. In SemEval-2025 Task 11 Track B this is a 0–3 score for each emotion; in SemEval-2007 Task 14 it is a 0–100 score. We normalize these vectors into a distribution $d$ over the $C$ emotions, keeping total intensity as a separate scalar. We fit a kernel density estimate to the seed vectors in this space. Generation targets $d^*$ are then drawn with probability weighted toward low seed density and multi-emotion mass. We require at least two emotions above a threshold and cap the share of single-emotion targets. This follows from the observation in Toward Understanding Generative Data Augmentation that synthetic data mainly helps by filling sparse regions when the real sample is small.

Generating. An instruction-tuned open model writes sentences given three things: a target blend described in words ("mostly anger, moderate fear, slight sadness"), a few seed sentences from the same language and genre as style anchors, and the total-intensity level. The design reuses the in-context generation of the PVI paper, but conditions on intensity blends instead of a class name.

Relabeling. Generators miss their targets, especially for subtle states. Proisl et al. show that anger and disgust, or fear and surprise, share contextual footprints. We therefore discard $d^*$ as a label. Each synthetic sentence $x$ receives a re-estimated distribution $\hat d(x)$, the average of two independent judges. One judge is a regressor trained on seed data only. The other is a prompted LLM asked for per-emotion intensities. This is the text analog of the soft labels that SAFLEX learns for noisy image augmentations. We avoid SAFLEX's bilevel optimization. It costs about 42% more compute per batch and needs a clean validation set, and a 116-example development split like the one in QMUL at SemEval-2025 Task 11 is too small for that. QMUL's own tuned thresholds overfit that split.

Filtering with a continuous information score. We extend PVI from class labels to distributions. Let $g_x$ be a predictor fine-tuned on the seed data that sees the sentence, and $g_\varnothing$ the same architecture fine-tuned with empty inputs, so it can only predict the average label. A synthetic pair $(x, \hat d)$ receives

$$ V(x,\hat d) \;=\; \mathrm{KL}\big(\hat d \,\|\, g_\varnothing\big) \;-\; \mathrm{KL}\big(\hat d \,\|\, g_x(x)\big), $$

which measures how much reading $x$ moves the prediction toward the sentence's label, in nats. When $\hat d$ is one-hot, $V$ reduces to ordinary PVI. We keep sentences with $V$ above a threshold that is set per region of the label space, not per class. We also drop sentences where the two judges disagree by more than the seed annotators typically disagree. The relabeling and the filter both reuse the seed-trained regressor. To avoid circularity, the final model is a fresh encoder trained on seed plus kept synthetic data, and it is evaluated only on held-out real test data.

Where domain-specific embeddings fit. This direction is a lens, not the core. Both the judge regressor and the filter's predictor $g_x$ need an encoder whose geometry reflects emotion and not only topic. QMUL found that pre-fine-tuning BERT on the DAIR-AI emotion tweets helped. We compare a general encoder against one adapted to emotion text, both inside the filter. This tests whether a domain-specialized scorer selects better synthetic data. It answers a narrower and more checkable question than building an affective knowledge graph.

Baselines

The naive reference is a pretrained multilingual encoder fine-tuned on the real seed data only, using the same training recipe as every other condition, with a KL loss on normalized intensities and a mean-squared-error loss on total intensity. This is the bar every augmentation must clear. On SemEval-2007 at the 10% split, the strongest published low-data distribution learner is EDL-Meta from Text Emotion Distribution Learning from Small Sample. It reports Euclidean distance 0.3949 and fidelity 0.8319, against 0.4450 and 0.7794 for a plain CNN. We reproduce its 10-fold protocol exactly. Beating it means lower Euclidean distance and higher fidelity than 0.3949 and 0.8319 on the same folds, with the gap exceeding the across-fold standard deviation.

For Task 11 Track B, the file contains no reported result on Lacuna for the top intensity correlation. We therefore re-run strong participant-style systems ourselves. For Track A's English split, the reported numbers are macro F1 0.5638 for QMUL's lexicon-augmented BERT (QMUL) and 0.7249 for prompted Llama-3.1-8B with the best verbalizer (NLP-Cimat). We report Track A as a secondary check after thresholding predicted intensities. For SemEval-2018 EI-reg, the top English system reached Pearson about 0.799 with full data (SemEval-2018 Task 1). That serves as a ceiling reference, not a low-data competitor.

The strongest augmentation competitor is discrete in-context augmentation with ordinary PVI filtering, following Lin et al.. It generates sentences per dominant emotion, labels them one-hot, and filters with class PVI. This is what a reviewer would reach for first. We give it the same generator, the same number of generated candidates, the same kept-set size, and the same final encoder and tuning budget.

Two ablations isolate our ingredients. In the first, targets are drawn from the seed label distribution instead of the upweighted blends; this removes the "where" ingredient. In the second, synthetic sentences keep their prompt target $d^*$ as the label and skip the filter; this removes the "how to label" ingredient. A third control, blend generation without filtering, measures how much the filter alone contributes. We count the method as winning only if the full system beats discrete PVI augmentation at matched synthetic volume, and only if each ablation loses ground.

Why not just…?

Why not just have an LLM soft-relabel the real training data, or distill its intensity judgments, with no new text? This is the most serious threat. Our relabeling step already uses such a judge, so the gain might come entirely from better labels on existing sentences. We run it as its own condition: real seed sentences plus LLM-estimated soft labels mixed with the gold ones, and the same downstream training. If it matches the full method, then new text adds nothing, and we will report that the contribution reduces to label enrichment. Our expectation is that relabeling cannot add coverage of blends that are absent from the seed set. The split by annotator disagreement described below is where the two should come apart.

Why not just use mixup or label smoothing to create interpolated blends cheaply? Mixing two sentence embeddings with mixed labels produces points in the blended interior with no generator at all. However, Proisl et al. and Jian et al. show that blended or confusable emotions are expressed through specific contextual cues, not through averages of clear cases. A linear mix of an angry and a fearful embedding need not resemble a sentence that is both. Embedding-space mixup at matched training steps is a baseline. If it equals generation on the high-disagreement subset, the expensive generator is unjustified.

Why not just LoRA-fine-tune a bigger model on the seed data? A Systematic Evaluation of LLM Strategies for Mental Health Text Analysis reports 91% accuracy for LoRA-tuned LLaMA-3 8B on DAIR-AI emotion, far above prompting (49%) and RAG (47%). Capacity therefore matters. That result used 16,000 training tweets and discrete labels. Our regime is hundreds of examples in languages where Task 11 systems fall below 0.30 F1 (SemEval-2025 Task 11). We include a LoRA 8B model trained on seed data only, and the same model trained with our augmentation. If augmentation helps the small encoder but not the 8B model, we will say so: the method would then substitute for capacity rather than add to it.

Proposed Experiments

We simulate low data on English Task 11 Track B, SemEval-2018 EI-reg, and SemEval-2007 Task 14 by subsampling 50, 100, 250, 500, and 1,000 seed sentences, with five seeds each. We also use the genuinely low-resource Task 11 languages at their full sizes. The metrics are Pearson correlation per emotion for intensity, plus KL divergence, Euclidean distance, and fidelity for distributions. Following Toward Understanding Generative Data Augmentation, whose theory predicts a constant-size gain at small sample sizes that can turn negative once real data is plentiful, we expect the advantage to shrink across the subsample ladder. A flat or growing advantage would contradict our account, not support it.

The decisive check uses Task 11's multiple annotators per instance. We split each test set into terciles by annotator variance and report gains per tercile. If our account is right, blend-targeted, relabeled augmentation should beat discrete PVI augmentation and soft relabeling mostly on the high-disagreement tercile, and roughly tie on the low-disagreement tercile. A uniform gain would suggest the method is simply adding data, not adding ambiguity coverage.

Three further checks probe failure modes. First, we audit 300 kept synthetic sentences per language with native speakers, measuring fluency, whether the realized emotion blend matches $\hat d$, and the rate of judge–human disagreement. Second, we run an Equity Evaluation Corpus probe (SemEval-2018 Task 1) to test whether the generator's stereotypes add demographic bias beyond the seed-only model. Third, we swap the filter's encoder (general versus emotion-adapted) to see whether a specialized scoring geometry keeps measurably better data, judged by downstream correlation at fixed kept-set size.