Emotional Annotation Function

• Emotional annotation functions are rigorously defined mappings that convert unstructured data into structured emotion labels, ratings, or descriptors.
• They use various methodologies including categorical, multi-label, and continuous mappings to capture complex affective dimensions across texts and signals.
• Protocols emphasize inter-annotator reliability and hybrid human-LLM approaches to ensure valid and scalable emotion analysis.

An emotional annotation function is a rigorously defined mapping—typically realized as a function $f$ or $A$—that assigns structured emotion labels, ratings, or descriptors to linguistic, acoustic, or multimodal units (e.g., words, sentences, utterances, spans, or continuous intervals) based on human judgment or algorithmic procedures. These functions operationalize the process of transforming unstructured data into formalisms suited to emotion modeling, analysis, and downstream tasks such as classification, regression, or generation. The function’s design encodes choices about granularity, label space (categories, dimensions, or continuous spectra), annotator perspective, protocol, and aggregation.

1. Formal Definitions and Label Spaces

The formal structure of emotional annotation functions varies by research objective and modality:

• Categorical mapping: A prototypical function maps each input unit $x$ (e.g., text segment, timeline point) from space $\mathcal{X}$ to a discrete set of labels $\mathcal{Y}$:

$f: \mathcal{X} \rightarrow \mathcal{Y}$

or for multi-label or probabilistic tasks,

$$f: \mathcal{X} \rightarrow \mathcal{P}(\mathcal{Y}) \quad \text{or} \quad f(x) = \{y \in \mathcal{Y} : P(y \mid x) \geq \tau \}$$

as in multilabel emotion analysis or filtering with LLMs (Niu et al., 2024).

• Multi-label and fine-grained mapping: Functions such as in PO-EMO map poetic lines to up to two “aesthetic” emotion labels per line:

$$f: L \;\longrightarrow\;\{\,E' \subseteq E\mid 1 \le |E'|\le 2\}$$

with $E$ a set of empirically motivated categories (e.g., Beauty/Joy, Sadness, Uneasiness, etc.) (Haider et al., 2020).

• Continuous/dimensional mapping: Continuous ratings are reflected by

$f: \mathcal{X} \rightarrow [a,b]^d$

where $A$0 is the number of emotion dimensions (e.g., Valence–Arousal–Dominance) and $A$1 specifies the permitted range, typically as in EmoBank (Buechel et al., 2022).

• Event/role-structured mapping: Some schemes (e.g., cognitive analysis, emotion carrier extraction) yield rich tuples:

$A$2

with $A$3 a token or span, $A$4 a semantic or functional label (cue, cause, experiencer, etc.), and $A$5 an emotion (Cortal et al., 2022, Tammewar et al., 2020).

Label sets are derived from psychological theory (e.g., Ekman, Plutchik, cognitive appraisal), aesthetic reception studies, or task-driven empirics. Dimensional models use scalar ratings, e.g., on SAM scales (Buechel et al., 2022). Some multi-dimensional frameworks (AffectSpeech) directly produce structured annotation objects spanning categorical, dimensional, prosodic, and open-vocabulary descriptors (Qi et al., 5 Apr 2026).

2. Annotation Protocols and Schemes

Annotation protocols are critical for validity and reliability. Key elements include:

• Unit of analysis: Lines (poetry), sentences, stanzas, utterances, time windows (e.g., 4s event windows in physiological studies (Tang et al., 3 Jul 2025)), or free-form spans (narrative “emotion carriers” (Tammewar et al., 2020)).
• Annotation context: Expert annotators access full context (e.g., stanza, poem (Haider et al., 2020)), and sometimes entire documents prior to segment-level labeling.
• Label assignment constraints: Multi-label constraints (e.g., max two per line), frequency-based pruning of under-specified or confusable categories (Haider et al., 2020), and annotation redundancy or ranking for primary and secondary emotions.
• Calibration and training: Repeated, iterative batch annotation with feedback rounds, use of gold standards or adjudication (INCEpTION workflow (Haider et al., 2020)), and workshops for harmonizing criteria in clinical contexts (DementiaBank-Emotion (Jeong et al., 4 Feb 2026)).

Crowdsourcing protocols incorporate pre-filtering for annotator location, redundancy levels, and task-specific cognitive load restrictions (Haider et al., 2020, Niu et al., 2024). LLM-assisted pipelines, as in AffectSpeech and recent LLM-centric studies, employ hybrid human-LLM adjudication, style diversification, and adversarial verification (Qi et al., 5 Apr 2026, Niu et al., 2024).

3. Statistical and Reliability Evaluation

Emotional annotation functions are evaluated for reliability and statistical agreement using established metrics:

• Inter-annotator agreement:
  • Cohen’s kappa ($A$6): For binary or multi-labels, as in PO-EMO, where per-label $A$7 ranged from 0.50 to 0.84, with average $A$8 across expert annotators (Haider et al., 2020).
  • Krippendorff’s alpha ($A$9) and multi-$x$0: Used for ordinal ratings, span-based schemes, or empathy scores (Wambsganss et al., 2021), and for capturing moderate to substantial agreement.
  • Fleiss’ kappa: Employed in multi-rater clinical settings for categorical label harmonization (Jeong et al., 4 Feb 2026).
  • Positive agreement (F1): For span-based "carrier" annotations where true negatives are ill-defined (Tammewar et al., 2020).
• Aggregation and adjudication: Majority vote, adjudication with a “reference-if-in-doubt” standard, and weighting by annotator confidence are commonly used (Haider et al., 2020, Jeong et al., 4 Feb 2026).
• Evaluation metrics for models: Macro-F1, micro-F1, accuracy, Pearson r, MAE for regression, and Jaccard index for set-based agreement (Haider et al., 2020, Buechel et al., 2022, Niu et al., 2024).

Empirical studies reveal agreement challenges in complex, subjective annotation settings (e.g., narrative emotion carriers, Krippendorff’s $x$1 for evaluative French industrial corpus (Noblet, 1 Sep 2025)), partially addressed by protocol design, explicit guidelines, and hybrid annotation workflows.

4. Automated, LLM-Based, and Quantization Approaches

Recent advances leverage LLMs and machine-centric techniques both as annotators and as assistants:

• LLM annotation and filtering: LLMs such as GPT-4 have been shown to produce annotation sets with higher human preference rates (>60%) compared to traditional annotator labels in forced-choice studies. LLMs can be used both for pre-filtering candidate labels (label-level) and post-filtering corpus items (sample-level), reducing annotator workload and improving training data utility (Niu et al., 2024, Niu et al., 2024). Filtering by intersection of LLM–human label sets consistently yields higher downstream model performance per sample than randomly downsampled or full-size sets (Niu et al., 2024, Niu et al., 2024).
• Energy-quantized annotation: The Expansion Quantization Network (EQN) formalizes all-label regression functions:

$x$2

with subsequent regression providing continuous micro-emotion scores for every label—extracting fine-grained, energy-level annotation suitable for data-intensive machine learning (Zhou et al., 2024).

• LLM-instructed multi-task annotation: EmoLLMs are instruction-following LLMs with an emotional annotation function

$x$3

mapping input $x$4 to discrete emotion label(s) $x$5 and real-valued intensity $x$6, supporting a variety of classification and regression tasks with accuracy/macro-F1 competitive or superior to supervised baselines and ChatGPT (Liu et al., 2024).

Physiological paradigms (e.g., immediate recall with video-induced EEG/GSR/ECG/PPG) reflect increased annotation precision, better alignment with objective affective signatures, and measurable improvement in emotion recognition models (+9.7% accuracy over whole-trial baselines) (Tang et al., 3 Jul 2025).

5. Best Practices, Transfer, and Limitations

Cross-domain and cross-genre transfer, as well as annotation best practices, have been identified:

• Iterative, theory-driven definition of label sets: Labels should be motivated by psychological/aesthetic theory, adjusted through pilot batches, and pruned via observed annotator confusion or infrequency (Haider et al., 2020, Buechel et al., 2022).
• Fine-grained unit and context presentation: Present annotators with full documents or stanzas for context, but enforce granular annotation units (e.g., lines, spans, windows) to maximize local specificity and mixed-emotion capture (Haider et al., 2020, Tang et al., 3 Jul 2025).
• Explicit documentation of guidelines: Iterative guideline development with detailed edge cases, calibration workshops, and adjudication protocols remains essential for inter-rater reliability (Jeong et al., 4 Feb 2026).
• Hybrid annotation models: LLMs offer scalable, high-throughput annotation, but human adjudication is essential for controlling for systematic misalignments and domain idiosyncrasies (Niu et al., 2024, Qi et al., 5 Apr 2026).
• Cognitive load and fatigue: Label-set size and interface design affect annotator fatigue and agreement; LLM pre-filtering offers a practical compromise (Niu et al., 2024).
• Evaluation diversities: Multiple axes (agreement statistics, human preference, downstream regression/classification) must be considered, as mutual exclusivity or subjectivity can limit single-metric reliability (Niu et al., 2024, Niu et al., 2024).

Limitations persist, including subjectivity in narrative span selection, ambiguity in implicit affect or experiencer attribution, and protocol transfer gaps outside annotated language or genre (Tammewar et al., 2020, Zad et al., 2022). Protocols are recommended to calibrate for such effects via pilot annotation, targeted guideline adjustment, and ongoing review.

6. Exemplary Functions Across Modalities and Frameworks

The table below summarizes representative emotional annotation functions, their formalisms, and key characteristics:

Paper/Corpus	Formal Function	Label Space/Unit
PO-EMO (Haider et al., 2020)	$x$7	Nine aesthetic categories, lines in poems
EmoBank (Buechel et al., 2022)	$x$8	VAD scores, two perspectives per sentence
EQN (Zhou et al., 2024)	$x$9	Continuous “micro-emotion” energy per label
AffectSpeech (Qi et al., 5 Apr 2026)	$\mathcal{X}$0 multi-dim. structured object	Categories, open text, prosody, segment, etc.
ABBE (Zad et al., 2022)	$\mathcal{X}$1 spans $\mathcal{X}$2 labels $\mathcal{X}$3 experiencers	Plutchik’s 8, minimal spans, explicit experiencer
Carrier Extraction (Tammewar et al., 2020)	$\mathcal{X}$4	Narrative spans, carrier vs. non-carrier
Cognitive Appraisal (Hofmann et al., 2021)	$\mathcal{X}$5	6–7 binary appraisal variables per event
LLM Pre/Post-filter (Niu et al., 2024)	$\mathcal{X}$6	Any fixed set $\mathcal{X}$7 (multi-label)
DementiaBank-Emotion (Jeong et al., 4 Feb 2026)	$\mathcal{X}$8	Ekman + neutral per utterance, with prosody

The diversity of formalizations reflects the breadth of emotion analysis research, ranging from subjective art appreciation to clinical affective science, from continuous time-aligned traces to multi-dimensional narrative structures.

---

In summary, emotional annotation functions constitute the algorithmic, protocol, and statistical foundations for all empirical work on emotion detection, modeling, and generation. Their design reflects accumulating evidence from psychology, linguistics, and computational modeling, as well as pragmatic tradeoffs in annotation cost, reliability, and domain specificity. Best practice is increasingly hybrid, leveraging statistical rigor, human-LLM synergy, and theory-driven scheme selection (Haider et al., 2020, Niu et al., 2024, Zhou et al., 2024).