From Expressions to Needs: Hierarchical Interpretive Reasoning for Emotional Support Conversation¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/43zxj/HierINT
Area: LLM Reasoning
Keywords: Emotional Support Conversation, Cognitive Appraisal Theory, Self-Determination Theory, Hierarchical Interpretive Reasoning, Collaborative Reasoning
TL;DR¶
Addressing the premature intervention and psychological misalignment of existing emotional support conversation systems that directly react to surface-level distress expressions, this paper proposes HierINT, a psychologically grounded hierarchical interpretive schema, along with the HierINT-CoRe collaborative reasoning framework and the HierINT-Dlg dataset, substantially boosting conversational empowerment and personalization.
Background & Motivation¶
Emotional Support Conversation (ESC) aims to provide empathetic, psychologically meaningful assistance to individuals experiencing emotional distress. However, existing conversational agents powered by large language models (LLMs) predominantly operate under a reactive recognize-and-respond paradigm: once negative affect or support-seeking cues are identified, models immediately trigger reassurance, praise, or generic advice based largely on superficial textual tokens. In authentic interpersonal counseling, skilled practitioners rarely respond directly to surface-level venting; instead, they engage in a deliberate interpretive process to understand why the distress arose, what cognitive meaning the seeker attaches to the situation, and what unmet psychological needs underlie their emotional turmoil before deciding how to intervene.
This reactive paradigm suffers from fundamental gaps at both the representation and reasoning levels. At the representational level, prior work typically categorizes seeker utterances using flat communicative intention labels. This flattens the critical distinction between what seekers explicitly articulate and what they fundamentally need. For instance, two seekers expressing identical workplace frustration and feelings of inadequacy might stem from radically different psychological origins: one may suffer from threatened competence needing self-efficacy affirmation, while the other experiences thwarted autonomy or longing for relational validation. Lacking a deep interpretive layer, systems produce well-intentioned yet psychologically misaligned platitudes (such as empty "keep trying, you've got this" tropes). At the reasoning level, recent intention-aware models and chain-of-thought frameworks focus almost exclusively on deciding what supportive actions to take, failing to reason about why a specific intervention is psychologically appropriate for the seeker's underlying condition.
To bridge this crucial gap between observation and intervention, emotional support must be reframed from immediate reaction to a psychologically guided interpretive reasoning pipeline. Core idea: Introduce an explicit, psychologically grounded hierarchical interpretive schema (HierINT) that chains Event Identification, Primary Appraisal, Secondary Appraisal, and Fundamental Needs before intervention planning, transforming emotional support from surface pattern matching into a need-aligned problem-solving process.
Method¶
Overall Architecture¶
The HierINT-CoRe (Hierarchical Interpretive Collaborative Reasoning) framework operationalizes the HierINT schema into a scalable, multi-stage pipeline. It comprises three coordinated modules: Seeker Psychological Interpretation, Supporter Intention Reasoning with causal explanation, and Strategy Planning & Multi-Intention Response Generation. The pipeline enforces a complete cognitive loop—reasoning inwards from observable linguistic cues to latent appraisals and psychological needs, before reasoning outwards to formulate supportive intentions, select communication strategies, and craft cohesive responses.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
In["Dialogue Context D_t & Current Seeker Utterance"] --> L1["Seeker Surface Intention Mining<br/>7 CBS communicative categories"]
L1 --> L2["Seeker Psychological Interpretation<br/>Cognitive Appraisal (CAT) & Needs (SDT)"]
L2 --> L3["Supporter Intention Reasoning & Explanation<br/>14 supportive intentions + rationale generation"]
L3 --> L4["Need-Aligned Strategy Planning & Response Generation<br/>Intention grouping & multi-strategy synthesis"]
L4 --> Out["Empathetic & Need-Aligned Response"]
Key Designs¶
1. Seeker Surface Intention Mining: Identifying Communicative Goals and Behavioral Roles To parse explicit conversational behavior without conflating it with deep psychological distress, the first module maps the seeker's latest utterance \(u_t^s\) to a set of surface communicative goals \(I_{\text{surf}}^t\). Grounded in Hill's Client Behavior System (CBS), this design defines 7 operational categories: Emotional Expression, Recounting, Support Request, Cognitive Exploration, Emotional Insight, Change Declaration, and Resistance. By categorizing what the seeker is performing at the conversational surface, this module establishes a reliable operational anchor for subsequent interpretive depth.
2. Seeker Psychological Interpretation: Cascading Cognitive Appraisals to Fundamental Needs To prevent premature intervention, this module constructs a four-stage Chain-of-Thought reasoning chain \(I_{\text{upi}}^t = \{\Psi_t, \text{UN}_t\}\) integrating Lazarus's Cognitive Appraisal Theory (CAT) and Ryan & Deci's Self-Determination Theory (SDT). First, Event Identification (EI) isolates the objective external trigger without subjective bias. Second, Primary Appraisal (PA) evaluates Self-Relevance and Self-Congruence. Third, Secondary Appraisal (SA) dissects cognitive coping evaluations across Blame/Credit, Coping Potential, and Future Expectations. Finally, these appraisal insights are mapped directly to unmet fundamental psychological needs (UN) under SDT: Autonomy, Competence, or Relatedness. This cascading design enables the system to uncover the true psychological driver behind surface expressions, preventing misaligned responses to complex distress.
3. Supporter Intention Reasoning & Explanation: Need-Driven Multi-Goal Inference with Rationale Supervision To ensure supportive interventions directly address the identified psychological vulnerability, this module translates the inferred needs into a multi-intentional goal set \(I_p^t \subset I_p\) (spanning 14 counseling-grounded intentions such as Promote Insight, Enhance Self-Efficacy, Facilitate Emotional Release, and Inject Hope). Crucially, the module simultaneously generates a natural language Intention Explanation (\(IE_t\)) justifying why this combination of goals addresses the seeker's specific cognitive appraisals and unmet needs. Enforcing explanation generation during training acts as an auxiliary inductive bias, forcing the model's latent representation to attend to core psychological deficits rather than superficial patterns.
4. Need-Aligned Strategy Planning & Response Generation: Intention Clustering and Cohesive Multi-Strategy Synthesis To reconcile the practical tension between validating raw emotion and providing structured guidance, the final module groups co-occurring supporter intentions, maps each cluster to corresponding counseling strategies \(S_{t,i}\) (from 9 strategies such as Reflection of Feelings, Affirmation and Reassurance, and Open Questions/Probes), and produces targeted intermediate response segments \(R_{t,i}\). These segments are then harmonized into a globally coherent final response \(R_t\): $\(R_t = f_{\text{gen}}(D_t, I_{\text{surf}}^t, I_{\text{upi}}^t, I_p^t)\)$ This structured assembly ensures that responses avoid abrupt topic switches or intrusive solution-pushing, delivering warm emotional validation while gently opening avenues for insight and autonomous action.
Loss & Training¶
HierINT-CoRe employs a parameter-efficient fine-tuning paradigm using LoRA on top of LLaMA3.1-8B-Instruct. The Seeker Psychological Interpretation module is trained separately with a learning rate of \(1 \times 10^{-4}\) for 2,400 steps. In contrast, the Supporter Intention Reasoning and Strategy–Response Generation modules are jointly optimized within a unified model using a learning rate of \(5 \times 10^{-4}\) for 2,300 steps. Both modules utilize LoRA configurations with rank \(r = 16\), scaling factor \(\alpha = 32\), dropout \(0.05\), per-device batch size 2, and maximum sequence length 4,096 tokens on NVIDIA A6000 GPUs. Joint training of supporter-side tasks reduces cumulative inference latency and achieves superior generation quality over cascaded independent supporter models.
Key Experimental Results¶
Main Results¶
Evaluation was conducted on both the crowd-sourced benchmark ESConv and the newly constructed HierINT-Dlg dataset. Responses were evaluated across basic conversational quality (Fluency, Coherence, Consistency, Safety) and core emotional support dimensions (Empathy, Informativeness, Empowerment, Personalization, Overall) using DeepSeek-V4-Flash on a 5-point Likert scale.
| Dataset | Model | Flue. | Cohe. | Cons. | Safe. | Empa. | Info. | Empo. | Pers. | Overall |
|---|---|---|---|---|---|---|---|---|---|---|
| ESConv | BlenderBot-Joint | 4.08 | 2.98 | 3.32 | 2.69 | 2.03 | 1.20 | 1.33 | 0.67 | 1.31 |
| MultiESC | 3.56 | 2.71 | 3.07 | 2.64 | 1.85 | 1.12 | 1.29 | 0.61 | 1.22 | |
| ESCoT | 4.84 | 4.22 | 4.48 | 3.49 | 3.33 | 1.59 | 2.19 | 1.22 | 2.08 | |
| IntentionESC | 4.83 | 3.82 | 4.18 | 3.12 | 2.41 | 1.35 | 1.57 | 0.82 | 1.54 | |
| HierINT-CoRe (Ours) | 4.70 | 4.71 | 4.80 | 3.69 | 3.72 | 1.95 | 2.53 | 1.68 | 2.47 | |
| HierINT-Dlg | BlenderBot-Joint | 2.03 | 1.98 | 2.36 | 2.41 | 2.41 | 1.19 | 1.65 | 0.89 | 1.54 |
| ESCoT | 5.00 | 4.20 | 4.63 | 3.49 | 3.69 | 2.08 | 2.83 | 1.86 | 2.62 | |
| IntentionESC | 4.98 | 3.66 | 4.03 | 3.29 | 2.92 | 1.65 | 2.25 | 1.45 | 2.07 | |
| HierINT-CoRe (Ours) | 4.98 | 4.73 | 4.92 | 4.00 | 4.15 | 2.64 | 3.53 | 2.72 | 3.26 |
Ablation Study¶
To isolate the contribution of each hierarchical component, an ablation study was performed on HierINT-Dlg using the LLaMA3.1-8B backbone (SI: Seeker Intention & Interpretation, PI: Supporter Intention):
| Configuration | Cohe. | Cons. | Safe. | Empa. | Info. | Empo. | Pers. | Overall | |---|---|---|---|---|---|---|---| | Full HierINT-CoRe | 4.73 | 4.92 | 4.00 | 4.15 | 2.64 | 3.53 | 2.72 | 3.26 | | w/o SI(Interp.) (remove underlying psychological needs) | 4.73 | 4.93 | 3.78 | 4.07 | 2.58 | 3.52 | 2.53 | 3.18 | | w/o SI (remove all seeker intention & interpretation) | 4.75 | 4.91 | 3.87 | 4.09 | 2.63 | 3.43 | 2.53 | 3.17 | | w/o PI (remove supporter intention modeling) | 4.64 | 4.81 | 3.83 | 4.08 | 2.49 | 3.44 | 2.67 | 3.17 | | w/o ALL (remove all intermediate reasoning stages) | 4.71 | 4.88 | 3.76 | 4.04 | 2.50 | 3.32 | 2.39 | 3.06 |
Key Findings¶
- Underlying psychological needs drive personalization and empowerment: Ablating the psychological interpretation layer (w/o SI(Interp.)) causes the sharpest drops in Personalization (from 2.72 to 2.53) and Safety (from 4.00 to 3.78), while removing all reasoning stages degrades Empowerment (from 3.53 to 3.32) and Overall score (from 3.26 to 3.06). This statistically confirms that modeling latent psychological needs, rather than surface breadth, is the primary driver of supportive quality.
- Robust cross-backbone and multi-turn performance: Switching to Qwen3-8B achieves a competitive Overall score of 3.50 (comparable to LLaMA3.1's 3.57 under GPT-4.1 evaluation), confirming that the framework's gains are architecture-agnostic. In multi-turn interactive roleplay via ESC-Eval (across 5, 8, 10, and 15 turns), HierINT-CoRe ranks first in Expression (2.95), Skill (2.95), and Overall score (1.98).
- Expert-validated psychological depth: Human evaluation of seeker state inference reveals that HierINT significantly outperforms baselines in Psychological Depth (2.97 vs. ESCoT 2.33 / IntentionESC 1.33) and Actionability for Support (2.93 vs. ESCoT 2.30 / IntentionESC 1.82) with an inter-annotator agreement of Krippendorff's \(\alpha = 0.734\).
Highlights & Insights¶
- Operationalizing CAT and SDT into an executable computational pipeline: By grounding latent inference in established psychological frameworks (5 appraisal dimensions and 3 basic needs), the framework avoids ad-hoc LLM speculation and establishes a disciplined causal pathway from external triggers to inner emotional blockages.
- Explanation-augmented intention reasoning: Generating natural language rationales for supporter intentions boosts intention prediction accuracy from 0.45 to 0.51, proving that verbalizing the psychological link between seeker vulnerability and supporter action enforces stronger cross-attention alignment.
- Preventing multi-turn error propagation and strategy lock-in: In cases where seekers disclose deeper self-doubt across multiple turns (e.g., thesis procrastination rooted in fear of intellectual inadequacy), baselines often remain trapped in productivity advice formulated during turn 1, whereas HierINT dynamically updates its appraisal and shifts interventions from task management to competence affirmation.
Limitations & Future Work¶
- Text-only modality and low-risk scope: The current framework relies solely on textual utterances, omitting critical non-verbal signals such as vocal prosody, hesitations, and facial expressions. Furthermore, extreme high-risk crises (e.g., suicidal ideation and self-harm) were intentionally excluded, meaning specialized triage and crisis referral mechanisms must precede clinical deployment.
- Defensive communication vulnerabilities: When seekers deliberately obscure emotional distress through deflection, cynicism, or polite compliance ("I'm fine, just tired"), event identification may fail, causing downstream appraisal misattributions.
- Lack of dynamic need prioritization: When multiple fundamental needs (e.g., competence and relatedness) are concurrently threatened, the framework currently lacks an explicit mechanism to arbitrate conversational urgency and temporal sequencing.
Related Work & Insights¶
- vs ESCoT (Zhang et al., 2024b): ESCoT applies chain-of-thought prompting to select support strategies, but its reasoning chain directly maps dialogue context to strategy labels without exploring why distress occurred or what psychological need exists. HierINT introduces an explicit psychological interpretive layer, providing the missing cognitive foundation for strategy choice.
- vs IntentionESC (Zhang et al., 2025): While IntentionESC models supporter intentions, it still treats intentions as communicative actions rather than need-fulfillment mechanisms. HierINT links seeker surface behaviors, cognitive appraisals, and unmet needs directly to supporter intentions, eliminating the gap between superficial empathy and true psychological alignment.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ Grounding ESC in Cognitive Appraisal Theory and Self-Determination Theory via a structured four-stage reasoning chain is theoretically rigorous and conceptually novel.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive evaluation across static benchmarks, module ablations, multi-turn roleplaying, multiple LLM backbones, and human psychological depth auditing.
- Writing Quality: ⭐⭐⭐⭐⭐ Exceptionally structured, insightful error propagation analysis, and lucid theoretical exposition.
- Value: ⭐⭐⭐⭐⭐ Sets a new standard for conversational AI in mental health, pivoting the community from surface text mimicry to genuine cognitive interpretation.