cs.HCFeb 23, 2026
SavePuppetChat: Fostering Intimate Communication through Bidirectional Actions and Micronarratives
Organizations: Virginia Tech Blacksburg, Virginia, USA · Department of Computer Science City University of Hong Kong Hong Kong, SAR · Department of Computer Science George Mason University Fairfax, Virginia, USA
Abstract
As a primary channel for sustaining modern intimate relationships, instant messaging facilitates frequent connection across distances. However, today's tools often dilute care; they favor single tap reactions and vague emojis that do not support two way action responses, do not preserve the feeling that the exchange keeps going without breaking, and are weakly tied to who we are and what we share. To address this challenge, we present PuppetChat, a dyadic messaging prototype that restores this expressive depth through embodied interaction. PuppetChat uses a reciprocity aware recommender to encourage responsive actions and generates personalized micronarratives from user stories to ground interactions in personal history. Our 10-day field study with 11 dyads of close partners or friends revealed that this approach enhanced social presence, supported more expressive self disclosure, and sustained continuity and shared memories.
Figures & tables
Figure 1. Conceptual vision and interface implementation of the expressive messaging system. (Top) A conceptual story sequence illustrating the types of emotional support and reciprocal physical interactions the system aims to facilitate between users. (Bottom) Interface mockups demonstrating the actual translation of these concepts into mobile UI features. The center lists four key Design Goals (DG1–DG4). Arrows indicate how specific UI components, such as the recommended avatar actions under "Physicalness" (implementing DG1 & DG2) and the saveable interaction history under "Memories" (implementing DG4), directly address these design goals. Conceptual vision and interface implementation of the expressive messaging system. (Top) A conceptual story sequence illustrating the types of emotional support and reciprocal physical interactions the system aims to facilitate between users. (Bottom) Interface mockups demonstrating the actual translation of these concepts into mobile UI features. The center lists four key Design Goals (DG1–DG4). Arrows indicate how specific UI components—such as the recommended avatar actions under "Physicalness" (implementing DG1 and DG2) and the saveable interaction history under "Memories" (implementing DG4)—directly address these design goals.
Figure 2. The main interface of PuppetChat . (A) The contact panel supports user management. (A1) Relationship management through personalized icons. (B) The conversation panel for dyadic exchange. (B1) Personal narrative input for guiding and editing AI generated micronarratives with persona cues. (B2) The interactive puppet area, where puppets animate actions or remain idle in a resting pose (left: partner; right: self). (B3) Action recommendation button that suggests reciprocal actions based on conversational context. The main interface of PuppetChat. (A) The contact panel supports user management. (A1) Relationship management through personalized icons. (B) The conversation panel for dyadic exchange. (B1) Personal narrative input for guiding and editing AI-generated micronarratives with persona cues. (B2) The interactive puppet area, where puppets animate actions or remain idle in a resting pose (left: partner; right: self). (B3) Action recommendation button that suggests reciprocal actions based on conversational context.
| Attribute | Description |
|---|---|
| Name | The label of the action. |
| Description | A concise textual summary of the action’s content and expressive intent. |
| Keywords | A small set of terms that capture typical triggers or contexts. |
| Emotion | The associated affective category. |
| Interaction role | The situational function of the action within dyadic interaction. |
| Embedding array | A high dimensional vector enabling semantic fuzzy matching. |
Table 1. Attributes of each action in the library. This table lists the attributes defined for each action in our system’s library. Each action is represented as a structured object that includes: the Name, serving as the label of the action; the Description, a concise textual summary of the action’s content and expressive intent; Keywords, terms that capture typical triggers or contexts; the associated Emotion category; the Interaction role, situating the action within dyadic interaction; an Embedding array, a high-dimensional vector that enables semantic fuzzy matching; and ReactionCandidates, which link to complementary or reciprocal actions.
Figure 3. Workflow of PuppetChat . (A) Everyday chat view. Pressing the Actions button (Fig. 2 B3) reveals (A1) a recommendation strip with four context aware actions sampled from a 42-item action library (A2), and clicking any action in (A1) advances to (B). (B) Composition view showing (B1) a visual preview of the selected action and an automatically generated micronarrative; tapping (B2) opens the customization panel (C). In (C), users refine tags via (C1) likes and dislikes, (C2) habits, (C3) social interaction style, and (C4) emotion, and may edit (C5) My Story . Clicking "Use Tag to Regenerate" (C6) advances to regenerate the micronarrative based on the selected tags and enables sending. (A) Everyday chat view. Pressing the Actions button reveals (A1) a recommendation strip with four context-aware actions sampled from a 42-item action library (A2). (B) Composition view showing (B1) a visual preview of the selected action and an automatically generated micronarrative; tapping (B2) opens the customization panel (C). In (C), users refine tags via (C1) likes and dislikes, (C2) habits, (C3) social interaction style, and (C4) emotion, and may edit (C5) My Story. Clicking "Use Tag to Regenerate'' (C6) advances to regenerate the micronarrative based on the selected tags and enables sending.
Figure 4. Text to action interpretation. (a) Keyword extraction with negation handling aligns user text with action tags; (b) emotional alignment maps input valence (positive, negative, or neutral) to the action’s emotion labels; (c) embedding-based fuzzy semantic matching covers indirect or metaphorical expressions. Text-to-action interpretation. (a) Keyword extraction with negation handling aligns user text with action tags; (b) emotional alignment maps input valence (positive, negative, or neutral) to the action’s emotion labels; (c) embedding- based fuzzy semantic matching covers indirect or metaphorical expressions.
Figure 5. Contextual evaluation. (a) Reaction candidate matching: the partner’s most recent action queries the action library’s ReactionCandidates to prioritize complementary responses; (b) interaction role bias: actions labeled responsive (vs. self oriented ) are promoted—conditioned on the ongoing conversational state—to encourage reciprocal exchanges. Contextual evaluation. (a) Reaction-candidate matching: the partner’s most recent action queries the action library’s ReactionCandidates to prioritize complementary responses; (b) interaction-role bias: actions labeled responsive are promoted—conditioned on the ongoing conversational state—to encourage reciprocal exchanges.
Figure 6. Status views for action delivery and exchange. (a) Action Only : the selected action is dispatched without a micronarrative. (b) Send : the action and the edited micronarrative are delivered together. (c) Partner response: both puppets are co-present and their actions are rendered as an interaction. Status views for action delivery and exchange. (a) Action Only: the selected action is dispatched without a micronarrative. (b) Send: the action and the edited micronarrative are delivered together. (c) Partner response: both puppets are co-present and their actions are rendered as an interaction.
| Participant | Total | Action | Micronarr. | Edit ops | Preview opens |
|---|---|---|---|---|---|
| U01 | 26 | 20 | 12 | 19 | 15 |
| U02 | 18 | 14 | 9 | 13 | 11 |
| U03 | 22 | 17 | 10 | 21 | 12 |
| U04 | 15 | 12 | 5 | 9 | 10 |
| U05 | 19 | 16 | 11 | 18 | 14 |
| U06 | 13 | 10 | 6 | 12 | 8 |
Table 2. PuppetChat usage by participant (original logs). PuppetChat usage by participant (original logs).
| Scale | |||
|---|---|---|---|
| Co-presence & Affective Intimacy | .78 | 5.3 | 0.9 |
| Reciprocal Interaction | .82 | 5.1 | 1.0 |
| Personalized Micronarratives | .86 | 5.2 | 1.1 |
| Agency & Persistence | .83 | 5.4 | 0.9 |
| Total (16 items) | .93 | 5.26 | 1.0 |
Table 3. Reliability and descriptive statistics for each subscale and the overall instrument. Reliability and descriptive statistics for each subscale and the overall instrument.
Figure 7. Distribution of responses for the 16 items of the Basic scale (N=22). Each boxplot shows the median and interquartile range of item ratings on a 1–7 Likert scale. reverse coded items are labeled with “(R)”; for all other items, higher scores indicate stronger endorsement. Distribution of responses for the 16 items of the Basic scale (N=22). Each boxplot shows the median and interquartile range of item ratings on a 1–7 Likert scale. Reverse-coded items are labeled with “(R)”; for all other items, higher scores indicate stronger endorsement.
| Measure | Mean | SD | 95% CI Low | 95% CI High | |
|---|---|---|---|---|---|
| SUS | 22 | 86.59 | 17.77 | 68.71 | 84.47 |
| RTLX | 22 | 31.72 | 6.98 | 28.63 | 34.81 |
Table 4. SUS and NASA–TLX overview. SUS and NASA–TLX overview.
Figure 8. Box plots of the six NASA–TLX dimensions. Box plots of the six NASA–TLX dimensions.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
| Subscale | Facet | Item text |
|---|---|---|
| Co-presence and Affective Intimacy | Presence | When using PuppetChat, I often feel as if my partner is right in front of me. |
| Disclosure | In conversations, I am more willing to share my genuine feelings or updates. | |
| Rituals | We developed some expressions or rituals unique to our relationship. | |
| Warmth | Compared to our usual chatting style, our relationship feels warmer. | |
| Reciprocal Interaction | Pairs | We often exchange paired or complementary actions. |
| Understanding | I can easily understand the intention behind my partner’s actions and respond accordingly. |
Table 5. BASIC scale items by subscale and facet. (R) indicates reverse coded items. BASIC scale items by subscale and facet. (R) indicates reverse-coded items.
| ID | Gender | Age | Relationship | Occupation | Messaging Apps |
|---|---|---|---|---|---|
| P1-A | Female | 24 | Best friends | Student | WeChat; WhatsApp; Instagram |
| P1-B | Female | 23 | Best friends | Student | |
| P2-A | Female | 25 | Couple | HR Business Partner | WeChat; WhatsApp |
| P2-B | Male | 26 | Couple | Software Developer | |
| P3-A | Female | 25 | Best friends | Student | LINE; Instagram |
| P3-B | Female | 22 | Best friends | Student |
Table 6. Participants (11 dyads): demographics, relationship, occupation, and messaging apps Participants (11 dyads): demographics, relationship, occupation, and messaging apps
More in this topic
Emotional Support Conversation
Emotional support conversation studies how people and AI systems provide empathetic, safe help through dialogue, including counseling, peer support, and companionship.
See all 60 papers