EmoStance: Response-Side Affective-Orientation Control for Empathetic Response Generation via Emoji Weak Supervision¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/18277390221/EmoStance
Area: LLM Reasoning
Keywords: Empathetic Response Generation, Listener Stance, Weak Supervision, Continuous Prefix Tuning, Emotion Distribution
TL;DR¶
EmoStance formulates empathetic dialogue generation as response-side affective-orientation control, leveraging multi-annotator emoji distributions to induce a name-free continuous stance space that steers frozen instruction-tuned LLMs via continuous prefix embeddings, achieving a 62.2% decisive win rate in blind human evaluation.
Background & Motivation¶
Empathetic conversation requires dialogue agents to decide not only what topical information to convey, but fundamentally how the speaker should take up and position themselves relative to the interlocutor's emotional situation. In open-domain social interaction, an agent's response can be topically coherent and grammatically fluent yet still feel detached, overly cheerful, didactic, patronizing, or insufficiently responsive to the speaker's emotional state. Prior research has primarily modeled affective context using coarse situation-level discrete emotion categories, dimensional affect representations (valence, arousal, dominance), predefined support strategy taxonomies (such as questioning, reflection, or suggestions), or external commonsense inference. However, these variables predominantly characterize dialogue state transitions or external factual grounding, falling short of specifying whether the upcoming response should operationalize reassurance, shared excitement, gentle concern, cautious probing, or supportive validation.
The core tension lies in the fact that identical coarse emotion labels frequently demand divergent affective uptake and interpersonal distance, while response-side orientation is inherently subjective, context-dependent, and multi-faceted. When presented with the same conversational context, multiple human listeners may reasonably adopt different communicative attitudes, rendering hard single-label supervision fundamentally ill-suited for fine-grained response control. Furthermore, such nuanced, mixture-like interpersonal stances cannot be stably verbalized through short, fixed natural-language prompts without causing instruction drift or rigid template-like phrasing, highlighting the necessity of an intermediate continuous control representation.
The angle of attack in this work is to treat "response-side affective orientation" (an operational approximation of listener stance) as an intermediate latent control variable, leveraging multi-annotator emoji distributions as soft weak supervision rather than explicit output tokens or gold classification labels. Emoji naturally capture compact affective-attitudinal nuances; by aggregating votes and confidence scores across multiple diverse LLM annotators into soft probability distributions, the system preserves annotator disagreement and contextual ambiguity while denoising idiosyncratic surface variations. Core idea: disentangle empathetic generation into response-side affective-orientation prediction and surface realization, utilizing multi-annotator emoji soft distributions to induce a name-free continuous control space that steers a frozen instruction-tuned LLM via prefix embeddings and optional consistency reranking.
Method¶
Overall Architecture¶
The EmoStance framework operates through three core stages: (1) Inducing a Name-Free Affective-Orientation Space: collecting confidence-weighted soft emoji distributions over a screened candidate inventory of 136 emojis, building an affinity graph via contextual usage and co-selection similarities, and partitioning it via Leiden community detection into \(K=9\) latent regions; (2) Role-Aware Response-Orientation Prediction: encoding the dialogue context with DeBERTaV3, estimating source-side affective expression, combining it with role-transition empirical priors through uncertainty-aware gating, and reconstructing a continuous control vector from orientation prototypes; (3) Continuous Prefix-Control Generation and Consistency Reranking: mapping the continuous control vector into virtual prefix embeddings via a lightweight MLP to steer a frozen Mistral-7B-Instruct-v0.3 generator, followed by optional decoding-time consistency reranking across sampled candidate responses.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}}%%
flowchart TD
A["Dialogue Context & Role Markers<br/>xt = [SIT] s [CTX] ... <rt+1>"] --> B["Name-Free Latent Stance Space Induction<br/>Emoji Affinity Graph + Leiden Community Projection"]
B --> C["Role-Aware Response Orientation Prediction<br/>DeBERTa Context Encoding + Gated Transition Prior"]
C --> D["Prototype-Based Continuous Vector Reconstruction<br/>Mixture of Region Prototype Centroids"]
D --> E["Continuous Prefix Control Generation<br/>Lightweight MLP Projector + Frozen Mistral-7B"]
E --> F{"Enable Orientation Consistency Reranking"}
F -->|No (B=1)| G["Efficient Single-Pass Generation"]
F -->|Yes (B=4)| H["Consistency Reranking Selection<br/>Minimize Cross-Entropy with Intended Stance"]
H --> I["Final Empathetic Response"]
G --> I
Key Designs¶
1. Name-Free Latent Stance Space Induction: Bypassing Discrete Emotional Nomenclature
Directly conditioning generation on surface emoji symbols or predefined discrete emotion taxonomies introduces severe label sparsity, synonymy redundancy, and subjective categorization bias. EmoStance addresses this by screening an affective emoji candidate inventory of 136 symbols and collecting independent annotations with 1-5 confidence scores from four diverse LLMs (DeepSeek-V3.2, Claude-Sonnet-4.6, Gemini-2.5-Pro, and GPT-5.4), yielding confidence-weighted soft distributions \(q^E_t \in \Delta^{|E|}\). Crucially, relying purely on data-internal relational structures—combining utterance contextual usage similarity \(S_{ctx}\) and annotator co-selection similarity \(S_{conf}\)—an emoji affinity matrix \(W\) is constructed, sparsified, and partitioned via the Leiden community detection algorithm into \(K=9\) latent regions. Each emoji is assigned a soft membership row in matrix \(A \in [0, 1]^{|E| \times K}\), mapping utterance-level emoji distributions into a smooth latent distribution:
This name-free induction transforms sparse, noisy symbolic labels into a compact, continuous probabilistic manifold, effectively neutralizing annotator-specific idiosyncrasies while preserving nuanced interpersonal ambiguity.
2. Role-Aware Response Orientation Prediction: Harmonizing Contextual Semantics with Interactional Priors
In natural dyadic communication, the listener's interpersonal stance is shaped not only by the immediate semantic situation but also by regular turn-taking and affective-uptake patterns across speaker transitions (e.g., speaker distress typically invites reassurance, whereas playful self-deprecation invites gentle tease or encouragement). EmoStance encodes the serialized context \(h_t = \text{Enc}_\theta(x_t)\) using DeBERTaV3, predicting source-side affective expression \(\hat{q}^Z_t\), and computes an empirical transition prior \(\pi_{t+1} = \hat{q}^Z_t T^{\rho_t}\) conditioned on the ordered role transition \(\rho_t = r_t \to r_{t+1}\). To prevent the prior from dominating when source-side expression is ambiguous, the system integrates an uncertainty-aware gate based on normalized entropy \(\bar{H}_t\):
The final next-response distribution \(\hat{q}^Z_{t+1} = \text{softmax}(\ell^0_{t+1} + \lambda_{tr} \gamma_t \log(\pi_{t+1} + \epsilon))\) dynamically balances local contextual semantics with interpersonal dialogue regularities.
3. Prototype-Based Continuous Vector Reconstruction: Mitigating High-Dimensional Regression Collapse
Directly predicting a high-dimensional continuous control vector (e.g., 256 dimensions) via end-to-end regression from text representations frequently leads to severe variance collapse and noise overfitting under weak supervision. Instead, EmoStance pre-computes mean prototype vectors \(\mu_k \in \mathbb{R}^{256}\) for each latent region over the training corpus, reconstructing the continuous orientation vector as a soft convex combination of prototypes:
During generation, the base language model (Mistral-7B-Instruct-v0.3) remains completely frozen. A lightweight two-layer MLP projector \(R_\omega(\cdot)\) maps \(\hat{v}_{t+1}\) into \(m\) continuous prefix token embeddings \(P_{t+1} \in \mathbb{R}^{m \times d_\Omega}\) prepended to the input sequence. This modular separation preserves the foundational linguistic and commonsense capacities of the LLM while achieving precise, non-invasive continuous steering.
4. Orientation-Consistency Candidate Reranking: Test-Time Fidelity Verification
Because soft prefix conditioning does not strictly guarantee that every autoregressively sampled sequence adheres to the intended stance, EmoStance incorporates an optional decoding-time verification mechanism. Given predicted orientation \(\hat{q}^Z_{t+1}\) and control vector \(\hat{v}_{t+1}\), the generator samples \(B\) candidate continuations. Each candidate is temporarily appended to the dialogue context and evaluated by the internal orientation scorer to derive its realized stance distribution \(\tilde{q}^{Z,(b)}_{t+1}\). The system selects the candidate minimizing cross-entropy divergence with the target distribution alongside a length penalty:
This test-time filtering requires no external models or auxiliary annotations, ensuring high stance fidelity at decoding time.
Loss & Training¶
The architecture is trained in two strictly decoupled stages: 1. Orientation Predictor Training: Optimizing the encoder and prediction heads using a composite loss over training pairs: $\(\mathcal{L}_{orient} = \lambda_{next} \text{CE}_w(q^Z_{t+1}, \hat{q}^Z_{t+1}) + \lambda_0 \text{CE}_w(q^Z_{t+1}, \hat{q}^{Z,0}_{t+1}) + \lambda_{cur} \text{CE}(q^Z_t, \hat{q}^Z_t) + \lambda_{vec} \left(\|\hat{v}_t - v_t\|_2^2 + \|\hat{v}_{t+1} - v_{t+1}\|_2^2\right)\)$ where \(\text{CE}_w\) denotes weighted soft cross-entropy compensating for latent region class imbalance. 2. Prefix Projector Training: Freezing the generator weights \(\Omega\), the projector parameters \(\omega\) are updated via standard autoregressive language modeling loss conditioned on the observed ground-truth orientation prefix: $\(\mathcal{L}_{gen} = -\sum_{j=1}^{|u_{t+1}|} \log p_\Omega(u_{t+1,j} \mid P_{t+1}, x_t, u_{t+1,<j})\)$
Key Experimental Results¶
Main Results¶
Primary evaluation is based on an expanded blind pairwise human evaluation (20 annotators, 800 independent judgments across two balanced batches, reporting decisive win rates excluding neutral ties) alongside reference-based and diversity metrics on the full EmpatheticDialogues test set (5,255 examples). All baselines are reproduced under the identical Mistral-7B-Instruct-v0.3 backbone.
| Evaluation Metric / Dimension | EmoStance (Ours) | LLM-only | LLM-prompt | LLM-SFT | EmPO-DPO | CASE (Reprod.) | APTNESS (Reprod.) | Sibyl (Reprod.) |
|---|---|---|---|---|---|---|---|---|
| Human Decisive Win% | 62.2% (Overall) | 54.4% | 56.5% | 62.6% | 58.1% | 78.4%* | 73.5%* | 42.6% |
| - 95% Wilson CI | [58.4, 65.9] | [44.2, 64.3] | [46.3, 66.2] | [52.8, 71.5] | [47.9, 67.6] | [69.5, 85.3] | [64.0, 81.2] | [31.0, 55.1] |
| - Two-sided sign test \(p\) | \(p < .001\) | \(p = 0.461\) | \(p = 0.251\) | \(p = 0.015\) | \(p = 0.146\) | \(p < .001\) | \(p < .001\) | \(p = 0.306\) |
| BERTScore-F1 ↑ | 0.6523 | 0.6348 | 0.6353 | 0.6440 | 0.6479 | 0.6240 | 0.6187 | 0.6492 |
| ROUGE-L ↑ | 0.1453 | 0.0908 | 0.0867 | 0.1310 | 0.1305 | 0.1354 | 0.0768 | 0.1115 |
| BLEU-2 ↑ | 0.0399 | 0.0168 | 0.0152 | 0.0329 | 0.0345 | 0.0309 | 0.0107 | 0.0243 |
| METEOR ↑ | 0.1594 | 0.2042 | 0.2199 | 0.1624 | 0.1694 | 0.1379 | 0.1906 | 0.2027 |
| Distinct-2 ↑ | 0.2450 | 0.2947 | 0.2404 | 0.2624 | 0.3489 | 0.0282 | 0.2292 | 0.2716 |
| Generic Rate ↓ | 0.4186 | 0.0228 | 0.0057 | 0.2634 | 0.1412 | 0.4725 | 0.0021 | 0.1743 |
(Note: Asterisks () denote Holm-corrected significance \(p < .001\). Generic Rate reflects rule-based matching of fixed generic templates and very short utterances \(\le 4\) tokens.)*
Ablation Study¶
Focused judgment-level human ablations (300 judgments per ablation pair across 100 contexts) and component diagnostics systematically validate each design:
| Ablation Comparison | Win | Tie / Neutral | Lose | Decisive Win% | 95% Wilson CI | Significance \(p\) |
|---|---|---|---|---|---|---|
| EmoStance Full vs. w/o Reranking | 156 | 71 | 73 | 68.1% | [61.8, 73.8] | \(< .001\) |
| EmoStance Full vs. w/o Role-Aware Transition | 138 | 84 | 78 | 63.9% | [57.3, 70.0] | \(< .001\) |
| EmoStance Full vs. Zero Control | 222 | 41 | 37 | 85.7% | [80.9, 89.5] | \(< .001\) |
| Overall Human Ablation | 516 | 196 | 188 | 73.3% | [69.9, 76.4] | \(< .001\) |
Continuous Vector Construction & Target Supervision: - Vector Construction: Prototype reconstruction yields a target cosine similarity of 0.9236 compared to only 0.3220 for direct 256-dimensional vector regression (\(\Delta = +0.6015\)), reducing MSE from 0.001058 to 0.000022. - Target Supervision Form: Soft distribution supervision substantially outperforms argmax hard labels, reducing cross-entropy from 1.4450 to 1.3792 and raising Macro-F1 from 0.3067 to 0.3260.
Key Findings¶
- Gains Are Concentrated in Contextual Uptake: Human dimension breakdowns reveal decisive win rates of 75.9% on Context Specificity and 73.5% on Felt Responded, while gains in surface Naturalness (57.7%) and Emotion Appropriateness (56.0%) are modest with intervals crossing parity. The approach enhances genuine interpersonal responsiveness rather than superficial lexical polish.
- Complementarity with Commonsense Reasoning: Comparing EmoStance against the future-aware commonsense baseline Sibyl yields an inconclusive win rate (42.6%, 95% CI [31.0, 55.1]), indicating that internal stance orientation and external knowledge reasoning target complementary cognitive dimensions.
- Efficiency vs. Quality Trade-off: Benchmarked on an RTX 4090, single-candidate generation (\(B=1\)) incurs 331.7ms latency (3.015 samples/s), whereas \(B=4\) reranking requires 1333.4ms (0.750 samples/s). Generating candidates accounts for 99.48% of the latency, while orientation scoring accounts for only 0.52%, confirming \(B=1\) as an efficient deployment mode and \(B=4\) as a quality-driven mode.
Highlights & Insights¶
- Reframing Emojis as Weak Structural Evidence: Rather than treating emojis as surface generation tokens or ground-truth emotion labels, EmoStance uses multi-annotator emoji probability distributions as soft weak supervision, bridging conversational semantics and continuous latent listener stances.
- Data-Driven Name-Free Space Induction: By discovering latent clusters through contextual usage and co-selection graph affinity rather than imposing rigid linguistic or psychological taxonomies, the framework remains highly flexible and robust to annotator disagreements.
- Stability via Prototype Convex Combination: Replacing brittle high-dimensional regression with soft mixture prototype reconstruction solves representation collapse, enabling seamless continuous steering of frozen 7B LLMs.
Limitations & Future Work¶
- LLM-Annotation Bias and Cultural Specificity: The weak supervision relies on four commercial LLM annotators operating on English dialogue benchmarks. Emojis inherently carry cultural, demographic, and platform-dependent nuances that may not generalize universally.
- Residual Template-Like Phrasing: Automatic metrics exhibit a relatively high Generic rate (41.86%), showing that under conservative predicted stances, the model occasionally defaults to safe, concise turns. Future work should investigate diverse decoding regularizers.
- Short-Horizon Dyadic Scope: Evaluation is restricted to brief dyadic interactions in EmpatheticDialogues, without testing long-horizon therapeutic support, group dialogue, or persistent persona memory.
Related Work & Insights¶
- vs MojiTalk (Zhou & Wang, 2018): MojiTalk directly conditions generation on single discrete Twitter emojis; EmoStance aggregates multi-annotator distributions to induce an internal latent continuous stance space without generating emojis at test time.
- vs CASE (Zhou et al., 2023) / APTNESS (Hu et al., 2024): Previous empathetic dialogue systems rely on hardcoded cognitive theories or complex multi-stage strategy trees; EmoStance learns an implicit, data-driven listener stance space that steers frozen LLMs via parameter-efficient prefix tuning.
- vs Sibyl (Wang et al., 2025): Sibyl relies on external commonsense inference regarding future events, whereas EmoStance governs internal interpersonal orientation; their mechanisms are orthogonal and combinable.
Rating¶
- Novelty: ⭐⭐⭐⭐☆ (Innovative use of emoji distributions as weak supervision to induce a name-free continuous listener-stance space)
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ (800 blind human judgments, multiple statistical corrections, extensive component ablations, and detailed profiling)
- Writing Quality: ⭐⭐⭐⭐⭐ (Exemplary clarity, precise problem formulation, and exceptionally honest boundary articulation)
- Value: ⭐⭐⭐⭐☆ (Provides an actionable, parameter-efficient paradigm for controllable affective alignment in LLMs)