PaperlayerView exact revision on arXiv

Source-linked AI summary

Empath: Tracing Multi-Level Emotion Dynamics in Crisis Counseling Dialogues

Ziwei Gong, Yuchen Huang, Wen Liang, Nicholas Deas, Melanie Subbiah, Kathleen McKeown, Julia Hirschberg

arXiv:2609.29056v1cs.CLcs.AI

TL;DR

Computational accounts of mental-health dialogue often label emotions utterance by utterance, leaving emotional change over time underexplored. EMPATH traces emotion from utterance labels to transitions and conversation-level patterns; in grief conversations with Black texters, it finds persistent distress alongside hope-ward movement and varied recovery trajectories.

  • Problem

    Much computational work treats mental-health dialogue emotion as static utterance-level labels, leaving the temporal structure of emotional change underexplored.

    Source: §1 · p. 1

  • Method

    EMPATH analyzes emotion dynamics through utterance-level emotions, turn-level transitions, and conversation-level trajectory archetypes.

    Source: §1 · p. 1, §4 · p. 3

  • Results

    In Crisis Text Line grief conversations with Black texters, EMPATH identifies persistent distress, gradual movement toward hope, distinct texter–volunteer emotional roles, and heterogeneous recovery trajectories.

    Source: §1 · p. 2, §7 · p. 10

  • Takeaways & Limitations

    The findings support understanding crisis support as a dynamic, interactional process rather than a simple shift from negative to positive emotion.

    Source: §7 · p. 10

    The analyses characterize emotional expression, not longitudinal grief change or clinical recovery, and the findings should not be generalized beyond the described samples and inclusion criteria.

    Source: Limitations · p. 10

Abstract

from arXiv · show

Emotion dynamics are critical for understanding crisis-support conversations, yet most computational work treats emotion as static utterance-level labels. We introduce EMPATH, a framework for understanding affective dynamics in mental health dialogues across three granularities: turn-level labels, transition probabilities, and global conversation archetypes. Applying EMPATH to text-based crisis conversations with self-identified Black texters discussing grief, we find persistent negative affect, gradual hope-ward transitions, distinct texter-volunteer emotional roles, and heterogeneous recovery trajectories. These results highlight the informative patterns that emerge from computationally understanding crisis support and expressions of grief as dynamic processes within conversations, as well as the overall value of emotion-dynamic analysis for analyzing and comparing affect in dialogues.

1 Introduction

EMPATH addresses the limited treatment of emotion as a static utterance label by tracing emotional change in crisis-support dialogue. In grief conversations with Black texters, it identifies persistent distress, hope-ward shifts, role differences, and varied trajectories.

1 Introduction

  • EMPATH analyzes crisis-support emotion at utterance, transition, and conversation levels, moving beyond aggregate emotion distributions.Its locally runnable labeling pipeline is designed for privacy-restricted data; its other levels examine turn-level transitions and conversation-level strategies and trajectories.Source: §1 · p. 1
  • In Crisis Text Line conversations with Black texters discussing grief, EMPATH finds repeated negative expressions, shifts toward hope, distinct texter–volunteer emotional roles, and heterogeneous patterns.The analysis focuses on texters who self-identify as Black or African American and conversations tagged as discussing grief.Source: §1 · pp. 1–2
  • These patterns suggest that crisis support is better understood as a dynamic process than as a terminal shift from distress to resolution.The timing and structure of transitions reveal aspects not visible from utterance-level labels or conversation endpoints alone.Source: §1 · p. 2
  • In synthetic crisis-support dialogues, EMPATH finds some aggregate emotional trends similar to authentic dialogues but gaps in fine-grained patterns.The authors use this comparison to show the framework also applies beyond Crisis Text Line conversations.Source: §1 · p. 2

2 Related Work

Prior work treats emotion recognition as a tool for understanding crisis intervention, while research on therapeutic dialogue, support strategies, and grief points to the value of examining emotional change over time.

2 Related Work

  • Emotion Recognition in Conversation (ERC) is described as a foundational NLP task and a tool for tracking crisis intervention trajectories.Source: §2 · p. 2
  • Studies of therapeutic dialogue report tradeoffs across model types, including stronger targeted clinical-task performance from domain-specific models (Ji et al., 2022).The review also describes crisis-focused continuous valence scoring for tracking turn-level de-escalation.Source: §2 · p. 2
  • Research on crisis de-escalation and emotional support dialogue examines interactional trajectories, support strategies, and changes in distress over time.Source: §2 · p. 2
  • The paper frames digital grief counseling as a domain for emotion-dynamics analysis because grief is complex and its expression has moved into online spaces.The discussion also notes evolving perspectives including dual process models (Margaret Stroebe, 1999) and meaning-making approaches.Source: §2 · p. 2

3 Crisis Dialogue Data

The study analyzes Crisis Text Line text conversations, focusing on grief discussions from texters who self-identified as Black or African American. Researchers annotate a 100-conversation sample to validate emotion models.

3 Crisis Dialogue Data

  • Crisis Text Line connects people seeking crisis support over text with volunteer counselors trained and supervised by mental health professionals.Source: §3 · p. 2
  • The corpus contains 2,478 de-identified grief-focused conversations from texters who self-identified as Black or African American in an optional survey.Researchers randomly sampled 100 conversations for human annotation and emotion-model validation.Source: §3 · p. 2
  • CTL data were de-identified, accessed under a signed Data Use Agreement, and processed locally without sending data to API-based LLMs or other online services.Source: §3 · p. 3
  • The 100-conversation annotation sample spans December 2016 to October 2023 and averages 56.3 dialogue turns per conversation.Two researchers annotated each conversation, resolving disagreements through discussion; the study focuses on 37 emotions identified by social work researchers.Source: §3 · p. 3

4 EMPATH Framework

EMPATH analyzes dialogue from utterance-level emotion labels through turn-level dynamics to conversation-level strategies and trajectory archetypes. Its emotion-labeling pipeline uses a locally runnable model, while subsequent analyses characterize transitions, distress, and conversation patterns.

4 EMPATH Framework

  • EMPATH processes dialogue through utterance-level emotion detection, micro-dynamics, and macro-dynamics to study crisis and therapeutic conversations.The three levels cover emotion labels, transitions and polarity, and strategies, archetypes, and diversity.Source: §4 · p. 3

4.1 Level 1: Emotion Detection

  • At the utterance level, a language model selects one of 37 emotion labels using the utterance and optional conversational context; the main analysis uses context.Context-free models score slightly higher on single-utterance accuracy, while context produces the temporally coherent trajectories needed for conversation-dynamics analysis.Source: §4.1 · pp. 3–4
  • Semantic similarity scores of 0.97–0.98 from Llama-3.2-3B-Instruct were comparable to larger models, and generative models outperformed BERT-based models across taxonomy levels.The authors selected this model for primary analyses as a trade-off between performance and computational efficiency.Source: §4.1 · p. 4

4.2 Level 2: Micro-Dynamics (Transitions and Polarity)

  • Micro-dynamics track adjacent transitions among 37 emotions and project those labels onto positive, negative, or neutral polarity using valence thresholds.Source: §4.2 · p. 4, Table 2 · p. 4

4.3 Level 3: Macro-Dynamics (Strategies and Archetypes)

  • Macro-dynamics derive turn-level distress from negated valence, compare downstream distress after volunteer strategies, and cluster normalized trajectories into five archetypes.The strategy analysis uses three-turn windows; trajectory archetypes are named by mean distress, total decline, and when that decline occurs.Source: §4.3 · pp. 4–5

5 Results: Emotion Dynamics in CTL Grief Conversations

Across CTL grief conversations with self-identified Black texters, EMPATH reveals persistent distress alongside hope-ward shifts and distinct texter-volunteer emotional profiles. Counselor strategies are associated with different short-term distress trajectories, while conversations follow heterogeneous recovery paths.

5.1 Level 1: Utterance-Level Emotion Profiles

  • Texter utterances are led by hopeless (25.0%), while volunteers’ are led by hopeful (46.4%); hopeful also ranks third for texters (12.6%).The texter top five labels account for 66.8% of utterances, while the full emotion inventory includes 37 labels.Source: §5.1 · p. 5
  • The emotion model may misinterpret African American Language in some cases, although dense AAL features are uncommon in this corpus.AAL was the most likely classifier label for about 6% of texts and had probability above .8 for less than 1%; the authors note emotion labels on AAL texts may be unreliable.Source: §5.1 · p. 6

5.2 Level 2: Micro-Dynamics

  • Six of the ten most frequent texter emotion transitions are self-transitions, and hopeless→hopeful is the most frequent recovery transition, ranking eighth overall.Hope-ward movement also appears across conversations: those beginning hopeless frequently end hopeful or grateful, while other distress-related starts disperse across positive and negative end states.Source: §5.2 · p. 6, Table 4 · p. 6, Figure 2 · p. 7
  • Texters show 84.1% negative-state persistence, compared with 64.6% for volunteers, while volunteers’ positive-state persistence is 72.4%.CTL conversations mostly begin in negative states (90.9%), while 70.9% end in positive states.Source: §5.1 · p. 5, §5.2 · pp. 6–7
  • Neutral states are transient for both roles (P(stay) ≤25.0%); texter neutral states transition to positive emotion in 41.4% of cases, versus 14.9% from negative states.This pattern suggests neutral states can function as intermediate points in broader affective movement.Source: §5.2 · pp. 6–7, Table 6 · p. 7

5.3 Level 3: Macro-Dynamics

  • Volunteer strategies differ in their association with next-turn distress: Information, Providing Suggestions, and Affirmation and Reassurance are associated with reductions, with Information showing the largest mean decrease.Question and Restatement or Paraphrasing are followed by small mean increases, which the authors say may reflect continued disclosure rather than ineffectiveness; Information estimates have wider uncertainty intervals.Source: §5.3 · pp. 7–8, Figure 3 · p. 8, Table 8 · p. 8
  • The five conversation-level trajectory archetypes each account for approximately 19–22% of the corpus, spanning early resolution, late recovery, high-distress recovery, and unresolved high distress.The distribution indicates heterogeneous recovery paths rather than a single average de-escalation curve.Source: §5.3 · p. 8

6 Secondary Analysis: Do Synthetic Dialogues Preserve Emotion Dynamics?

The secondary analysis applies EMPATH’s emotion, transition, strategy, and trajectory analyses to zero-shot and dual-agent synthetic crisis dialogues. Synthetic conversations show distress reduction but do not consistently reproduce CTL’s finer-grained dynamics or balanced trajectory distribution.

Micro Dynamics

  • Synthetic dialogues show substantial distress reduction, but distress often remains elevated longer or changes more abruptly than in CTL, depending on generation setup.The figure compares mean texter distress trajectories for the full CTL corpus and synthetic conditions.Source: §6 · p. 8, Figure 4 · p. 9
  • Synthetic texters show greater negative-state persistence than CTL texters (89.4% vs. 84.1%) and fewer direct negative→positive transitions (10.3% vs. 14.9%).No cross-polarity recovery transition appears in the synthetic top ten, unlike CTL’s recurring hopeless→hopeful pivots.Source: §6 · pp. 8–9

Contextual Grounding

  • Synthetic support can use the same broad Information strategy as CTL while offering generic resources rather than a grief-specific, context-matched resource.The comparison illustrates that strategy labels alone do not capture how support is adapted to the preceding interaction.Source: §6 · p. 9, Table 9 · p. 9

Macro Dynamics

  • Neither synthetic condition reproduces CTL’s relatively even distribution across five trajectory archetypes, each accounting for 19–22% of CTL conversations.Zero-shot conversations overrepresent Steady De-escalation (high distress) (34% vs. 19% in CTL); 64% of dual-agent conversations are Unresolved High Distress, partly due to the termination procedure.Source: §6 · p. 9, Figure 5 · p. 9
  • Synthetic dialogues can reproduce some aggregate emotional properties but do not consistently recover CTL’s gradual, locally patterned, contextually grounded, and heterogeneous dynamics.EMPATH characterizes these properties directly rather than relying on surface plausibility alone.Source: §6 · p. 9

7 Conclusion

EMPATH tracks crisis-support emotion at utterance, turn-transition, and conversation-trajectory levels. In CTL grief conversations with Black texters, it reveals persistent distress alongside hope-ward movement and varied recovery paths.

7 Conclusion

  • EMPATH analyzes utterance-level emotions, turn-level transitions, and conversation-level trajectory archetypes.The framework is presented as a tool for studying affective dynamics in crisis-support dialogue.Source: §7 · p. 10
  • CTL grief conversations show persistent distress, gradual hope-ward movement, distinct texter–volunteer emotional roles, and heterogeneous recovery trajectories.The paper characterizes crisis support as a dynamic, interactional process rather than a simple shift from negative to positive emotion.Source: §7 · p. 10
  • EMPATH is also positioned for studying broader counseling dialogues and comparing affective dynamics in synthetic or simulated settings.The authors identify peer support, therapy dialogue, counselor-training simulations, and mental-health dialogue systems as possible extensions.Source: §7 · p. 10

Limitations

The analyses describe emotional expression in sampled text-based crisis conversations, not clinical recovery or longitudinal grief change. Their interpretations are bounded by the data, model predictions, and synthetic-comparison conditions.

Limitations

  • The analyses do not measure longitudinal changes in grief or establish clinical recovery.They characterize emotional expression within text-based crisis-support conversations.Source: Limitations · p. 10
  • Findings are specific to the study’s samples and inclusion criteria and should not be generalized to all CTL texters or Black communities broadly.Source: Limitations · p. 10
  • Emotion labels are model predictions, single-label outputs may miss simultaneous emotions, and volunteer labels may reflect texter emotions rather than volunteers’ own states.Source: Limitations · p. 10
  • Associations between support strategies and subsequent label-derived distress do not establish causal effects or counseling effectiveness.Source: Limitations · p. 10
  • Synthetic comparisons are sensitive to topic composition, prompt design, context-window length, and conversation termination criteria.Source: Limitations · p. 10

Ethics

The paper describes privacy protections for CTL data and an emotion-labeling setup built around prompts, filtering, and staged prediction. The setup uses different context windows for CTL and synthetic conversations.

Ethics

  • CTL data were de-identified, processed locally under a signed agreement, and withheld from API-based LLMs and other online services.The paper also states that its example dialogue is simulated and CTL excerpts are paraphrased.Source: Ethics · p. 10

A Emotion Detection Prompt, Label Inventory, and Experimental Set-Ups

  • Preprocessing removes system messages, single-character Y/N feedback, and trivial first texter messages; sequence analyses additionally filter non-substantive neutral openers.Mid-conversation neutral labels are retained, and the opener filter is not applied to label-frequency analyses.Source: A Emotion Detection Prompt, Label… · p. 14
  • The CTL pipeline uses Llama-3.2-3B-Instruct and a five-preceding-utterance context window, while synthetic data use context length 10.The reported CTL inference took approximately 27.5 hours for 117,248 items across 1,832 batches.Source: A Emotion Detection Prompt, Label… · p. 14
  • The emotion prompt supplies 37 labels and directs the model to weigh the current message most heavily, using preceding turns as supporting context.It requests a structured label-and-reason response.Source: A Emotion Detection Prompt, Label… · p. 14
  • A three-stage cascade accepts valid generated labels, scores first tokens when parsing fails, and applies self-consistency voting to uncertain predictions.When no local model is available at a fallback stage, the utterance is assigned neutral and flagged.Source: A Emotion Detection Prompt, Label… · pp. 14–15

B Model Validation

Validation compares emotion-labeling models and taxonomy mappings, then specifies the representations used for polarity and conversation-trajectory analyses. The selected context configuration favors temporally coherent sequences for dynamics analysis.

B Model Validation

  • 0.682 vs. 0.667 Ekman 7-category accuracy separates Mistral-7B-Instruct-v0.3 from Llama-3.2-3B-Instruct, which the authors use for primary analyses.At sentiment level, their accuracy is 0.844 vs. 0.842, respectively; generative-model semantic similarity is 0.975–0.983 across mapped levels.Source: B Model Validation · p. 16

D Multi-Level Emotion Taxonomy Mapping

B Model Validation

  • The with-context configuration is used for primary analyses because it produces more coherent trajectories, despite lower single-utterance accuracy.Without context, the reported change rate is 84.2% vs. 63.9% with context, and the role distinction collapses.Source: B Model Validation · pp. 17–18, Table 12 · p. 18

C Emotion to Polarity Mapping

E Archetype Analysis Details

  • Texter distress sequences with at least four labeled turns are interpolated to 20 normalized-progress points, then pooled across sources for five-cluster K-means archetypes.Pooling shares the archetype labels across CTL and both synthetic sources.Source: E Archetype Analysis Details · p. 18
  • Archetype names are based on mean distress, total decline, timing of decline, and volatility, with high-distress suffixes set by the centroid’s mean distress.The paper reports that no centroid met its relative-volatility criterion for Rupture-Repair or Oscillatory Distress.Source: E Archetype Analysis Details · pp. 18–19
  • Inter-conversation diversity is the mean pairwise Euclidean distance among length-normalized distress trajectories, with bootstrap intervals used to assess CTL-versus-synthetic differences.The procedure subsamples each source to at most 1,000 conversations and uses 1,000 bootstrap resamples.Source: E Archetype Analysis Details · p. 20

F.1 Speaker Role Differentiation

With conversational context, texters and volunteers show distinct emotion profiles, while removing context largely erases that differentiation. Texter emotions change more often, and volunteers more often persist in positive states.

F.1 Speaker Role Differentiation

  • Texter utterances persist in negative states at 84.1%, while volunteer utterances persist in positive states at 72.4%.Texters also begin in a negative state in 90.9% of conversations.Source: §F.1 · p. 20
  • Texter emotion labels change more often than volunteer labels, at 62.8% versus 52.8% per conversation.Table 15 reports per-conversation emotion-change rates, defined as the fraction of adjacent utterance pairs with different labels.Source: §F.1 · p. 20, Table 15 · p. 20, §F.2 · p. 20
  • The volunteer’s most frequent cross-emotion transition is hopeless→hopeful, while texter labels are more variable.Without context, dominant self-transitions converge on low-specificity labels, and the roles’ emotional differentiation largely collapses.Source: §F.1 · p. 20
  • Without conversational context, negative persistence gaps narrow from 22.0 pp to 12.0 pp and positive persistence gaps from 8.9 pp to 3.1 pp.Both roles end positive at identical rates of 37.6%.Source: §F.1 · p. 20
  • The findings support using the with-context configuration for primary analyses because context helps distinguish speaker roles and recover the therapeutic arc.The section states that the full with- versus without-context comparison appears in §G.Source: §F.1 · p. 20

G With- vs. Without-Context Comparison

With conversational context, EMPATH preserves texter–volunteer role distinctions that largely collapse when utterances are classified alone. The section also describes dual-agent and zero-shot setups for generating synthetic crisis-support dialogues.

G With- vs. Without-Context Comparison

  • The comparison tests utterance-only prediction against a sliding window of recent utterances from both speakers plus the target utterance.The window contains the N most recent preceding utterances.Source: G With- vs. Without-Context Comparison · p. 20
  • With context, texters have a 22.0 pp higher negative-persistence rate than volunteers, while their starting negative-state rates are 92.2% versus 51.8%.Without context, the negative-persistence gap narrows to 12.0 pp, and empathetic counselor messages can read as negative in isolation.Source: G With- vs. Without-Context Comparison · pp. 20–21
  • Without context, emotion-change rate rises to 84.2% from 63.9% with context, and both roles end positive at 37.6%, collapsing role distinctions.With context, sequences shift from 92.2% negative starts to 63.7% positive endings.Source: Table 16 · p. 21, G With- vs. Without-Context Comparison · p. 21

H Synthetic Generation Conditions

  • Dual-agent generation uses separate role-specific LLM agents turn by turn, while zero-shot generation produces a complete conversation in one pass.No CTL messages or CTL-derived materials were used in either setup.Source: H Synthetic Generation Conditions · p. 21
  • Synthetic dialogues use three instruction-tuned models and seeds covering grief-related and non-grief concerns, with prompted texter openings and volunteer counseling behaviors.Texter prompts adapt simulated-patient prompts; volunteer prompts encode counselor-oriented behaviors and asymmetric role grounding.Source: H Synthetic Generation Conditions · pp. 21–22
  • The same EMPATH pipeline is applied to CTL and synthetic conversations, from emotion labeling and transition analysis through strategy classification and archetype assignment.Source: H Synthetic Generation Conditions · p. 22
Loading 2609.29056v1…