Dual-Aspect Empathy (DAE) Framework

• Dual-Aspect Empathy (DAE) is a framework that models both cognitive understanding and emotional sharing, distinguishing between the originator’s intent and the recipient’s perception.
• Empirical annotation protocols within DAE use multidimensional scales and taxonomies to capture self- and other-reported empathy in settings like dialogue and therapeutic contexts.
• DAE underpins advanced AI systems and human-centric applications by integrating algorithmic modeling with measurable outcomes, thereby enhancing empathy detection and response generation.

Dual-Aspect Empathy (DAE) is a conceptual and computational framework that formalizes empathy as an interplay between two complementary, yet distinct, processes: understanding others (cognitive empathy) and sharing or responding affectively (affective/emotional empathy). DAE has become foundational in both empirical social-interaction research and the development of artificial intelligence systems that seek to operationalize, evaluate, or simulate empathy in dialogue, therapeutic, support, and even adversarial contexts. The explicit modeling of both the originator’s intent and the recipient’s perception (or mutual sensemaking and adaptation in dyads) reveals subtle divergences between the experience and the recognition of empathy, offering richer analytic and generative capability than single-aspect perspectives.

1. Theoretical Foundations and Definitions

DAE is rooted in multidisciplinary theories of empathy. Classical definitions distinguish between “affective empathy” (the capacity to feel or share another’s emotional state) and “cognitive empathy” (the ability to infer, understand, or reason about another’s thoughts, feelings, or intentions) (Sabour et al., 2021, Hu et al., 2024). Recent theory extends DAE into bidirectional social-cognitive frameworks, recognizing empathy failures as mutual—rather than unidirectional—mismatches, a notion encapsulated in Milton’s double empathy problem (Tang et al., 21 Feb 2026). DAE thus encompasses:

• Cognitive Empathy: Understanding or reconstructing another’s affective state or situation, operationalized as perspective-taking, emotion recognition, or goal/inference modeling.
• Affective (Emotional) Empathy: Emotional attunement, sharing resonance, or affective response, operationalized as emotional expression, comfort, or affective mirroring (Sabour et al., 2021, Hu et al., 2024).

A parallel distinction exists in empirical annotation: “self-report” empathy (actor’s felt/claimed empathic engagement) and “other-report” empathy (observer’s or interlocutor’s perception of empathy expressed) (Omitaomu et al., 2022, Xu et al., 2024).

2. Annotation Protocols and Quantitative Scoring

Large-scale annotation efforts in DAE research decompose empathy into multiple axes for tractable measurement. In the “Empathic Conversations” dataset, DAE is defined at the turn level as paired 7-point Likert self-report and other-report scores per dialogue turn. For a turn $t$:

• $e_\text{self}^t \in \{1, ..., 7\}$: speaker’s self-assessment
• $e_\text{other}^t \in \{1, ..., 7\}$: receiver’s assessment of empathy
• Normalization: $$\hat{s}_\text{self}^t = (e_\text{self}^t - \mu_\text{self})/\sigma_\text{self}$$, $$\hat{s}_\text{other}^t = (e_\text{other}^t - \mu_\text{other})/\sigma_\text{other}$$

Similarly, in customer-service and peer-support domains, DAE is operationalized by multi-label intent taxonomies (expressed empathy: 16 fine-grained strategies) and by perceived empathy ratings (engagement, understanding, sympathy, helpfulness; 1–5 scales) (Xu et al., 2024).

Empirical distributions consistently show moderate but not perfect correlation between facets (e.g., per-turn $r = +0.45$ between self- and other-report), with demographic and personality moderators (e.g., agreeableness, gender) (Omitaomu et al., 2022). Annotation reliability achieved Cohen’s $\kappa \approx 0.6$ for both axes.

Table 1. Dual-Aspect Empathy Annotations (Example: (Omitaomu et al., 2022))

Aspect	Rating Metric	Example Item
Self	7-pt Likert (norm.)	“How empathic did you feel in this response?”
Other	7-pt Likert (norm.)	“How empathic was your partner’s previous response?”
Expressed	16-label taxonomy	“Did agent express sympathy, apology, explanation?”
Perceived	4-dim 5-pt scale	“Did agent seem understanding? Helpful? Sympathetic?”

3. Algorithmic Modeling and Representation

DAE’s dual-factor nature is central to model architectures aiming at empathy prediction or generation.

• Empathetic Response Generation: Methods such as CEM (Sabour et al., 2021) and APTNESS (Hu et al., 2024) architecturally separate cognitive and affective streams. CEM encodes affective commonsense ($xReact$) and cognitive commonsense ($xWant$, $xNeed$, $xIntent$, $xEffect$) using a Transformer-based fusion:

$$\mathbf{H}_\text{Refine}[i]=\mathbf{H}_\text{Aff}[i]\oplus\mathbf{H}_\text{Cog}[i], \quad \widetilde{\mathbf{H}}_\text{CTX} = \mathbf{MLP}(\sigma(\mathbf{H}_\text{Refine})\odot\mathbf{H}_\text{Refine})$$

Knowledge selection, gating, and separate loss components for emotion recognition ($\mathcal{L}_\text{emo}$), response likelihood ($\mathcal{L}_\text{nll}$), and lexical diversity ($\mathcal{L}_\text{div}$) ensure both aspects are represented and measurable.

• Retrieval-Augmentation and Strategy Integration: APTNESS operationalizes cognitive empathy via semantic retrieval of contextually relevant emotional situations (appraisal framework) and affective empathy via explicit emotional support strategies (LoRA adapter; strategy definitions for comforting, validation, etc.). Empathy is quantitatively evaluated along identification (cognitive) and comforting (affective) axes.
• Multi-Agent and Human-in-the-Loop Systems: NeuroWise (Tang et al., 21 Feb 2026) structures DAE as multi-agent LLM pipelines (interpreting hidden internal states, externalizing cognitive states through explanations and visualization, and coaching adaptive response), embodying DAE’s requirement for both sensemaking and adaptive modulation.
• Automatic Empathy Classification: Instruction-finetuned LLMs (e.g., Flan-T5-XL) are employed for DAE evaluation; multi-task heads predict expressed intents and perceived empathy (Xu et al., 2024).

4. Empirical Findings and Moderating Factors

Research consistently finds a moderate alignment between DAE aspects: self-reported and perceived empathy exhibit positive yet clearly distinct correlation ($r \sim 0.45$ per-turn). The gap is accentuated by personality (agreeableness, extraversion), demographics (gender, age), and content factors (e.g., response specificity) (Omitaomu et al., 2022, Xu et al., 2024). For example:

• High agreeableness: stronger coupling of self–other empathy ($r=0.54$) vs. low ($r=0.31$).
• Gender: female responses perceived as more empathic ($\mu_\text{female}=4.01$, $\mu_\text{male}=3.72$, $p=0.005$).
• Response specificity (inverse perplexity): more predictive of other-perceived empathy ($\beta=-0.22$, $p<0.01$) than self-reported empathy.

Ablation and regression analyses show subject age moderates distortion (e.g., younger participants overestimate self, older underestimate) (Omitaomu et al., 2022).

Supervised fine-tuning and enhanced instruction context significantly improve performance in both empathy prediction and classification tasks over zero-shot or encoder-only methods, with Macro F1 gains of $+2$–$7$ points and high correlation (Spearman’s $\rho=0.58$ for helpfulness perception and satisfaction) (Xu et al., 2024).

5. Practical Applications

DAE frameworks are implemented in:

• Dialogue and Conversational AI: Both generation (e.g., empathetic chatbots (Sabour et al., 2021, Hu et al., 2024)) and evaluation (multi-dimensional analysis, real-time monitoring, instruction-finetuned classifiers (Xu et al., 2024)).
• Psychological Support and Peer Counseling: Human–AI interaction enhanced by retrieval-augmented and strategy-guided LLMs improves both short- and long-context empathy quality (Hu et al., 2024).
• Human–Human Mediation: Systems like NeuroWise (Tang et al., 21 Feb 2026) provide visual and textual cues to support double-DAE in neurodivergent–neurotypical communication, yielding measurable reductions in deficit-based attributions ($\Delta = -0.63$ NeuroWise vs. $+0.30$ baseline, $p=0.02$) and increased communication efficiency (median turns: 8 vs. 11, $p=0.03$).

Table 2. DAE Applications and Outcomes

Domain	Model/System	Key Outcome (DAE-relevant)
Dialogue	CEM, APTNESS	Improved empathy/emotion accuracy, diversity, coherence
Coaching	NeuroWise	Reduced deficit framing, <em>efficient mutual repair</em>
Evaluation	Flan-T5-XL FT	Macro F1 ↑ (+2–7 pts vs. prior), context critical

6. Evaluation, Metrics, and Limitations

Empirical evaluation of DAE uses multi-axis metrics:

• Regression/classification (MSE, cross-entropy, focal, LDAM) for self/other-reports, intent, and aspect-wise perceived empathy (Omitaomu et al., 2022, Xu et al., 2024).
• Dialogue-level preference (human/automatic): Empathy, coherence, informativeness, diversity, and explicit DAE-aligned metrics (identification, comforting) (Sabour et al., 2021, Hu et al., 2024).
• Real-world satisfaction: Spearman’s $\rho$ links perceived empathy to user satisfaction (e.g., helpfulness $\rho=0.58$) (Xu et al., 2024).

Limitations of current DAE implementations include reliance on static or external commonsense sources (e.g., COMET), coverage gaps in emotional-state databases, occasional instability in strategy prediction for long-context dialogues, and the challenge of generalizing attributional change out-of-domain (Sabour et al., 2021, Hu et al., 2024, Tang et al., 21 Feb 2026).

7. Emerging Directions and Open Questions

Recent research highlights several open problems:

• Full bidirectionality: Developing interfaces and models that support DAE guidance for both modes of mismatch (e.g., autistic and neurotypical partners simultaneously (Tang et al., 21 Feb 2026)).
• Personalization: Tuning DAE systems to individual profiles (across neurotypes, emotional palettes, or support needs).
• Mechanism optimization: Determining optimal types, complexity, and timing of interpretive feedback for attributional learning.
• Data and generalization: Expanding DAE resources to encompass rare/mixed emotions, and validating real-world transfer/persistence of empathetic adaptation.
• Architectural innovation: Partitioning cognitive/affective heads in decoders, integrating multimodal/multilingual datasets, and incorporating iterative human-in-the-loop refinement (Hu et al., 2024).

Ongoing work continues to strengthen DAE’s empirical reliability and its value as a paradigm for both human-centric AI and the analytic understanding of empathy in naturalistic social settings.