The Dynamics of Emotional Chats with Bots: Experiment and Agent-Based Simulations

Quantitative research of emotions in psychology and machine-learning methods for extracting emotion components from text messages open an avenue for physical science to explore the nature of stochastic processes in which emotions play a role, e.g., in human dynamics online. Here, we investigate the occurrence of collective behavior of users that is induced by chats with emotional Bots. The Bots, designed in an experimental environment, are considered. Furthermore, using the agent-based modeling approach, the activity of these experimental Bots is simulated within a social network of interacting emotional agents. Quantitative analysis of time series carrying emotional messages by agents suggests temporal correlations and persistent fluctuations with clustering according to emotion similarity. {All data used in this study are fully anonymized.}

💡 Research Summary

The paper investigates how emotionally expressive chatbots influence both individual users and the collective dynamics of a social network. Three affective profiles—positive, negative, and neutral—were implemented in an Affect­ive Dialog System (ADS). In a controlled laboratory experiment, 91 participants each engaged in three seven‑minute chats (one per profile) with the bot. The conversation logs were processed using a suite of natural‑language‑processing tools, including a Support Vector Machine dialog‑act classifier, a lexicon‑based sentiment analyzer, and the ANEW (Affective Norms for English Words) dictionary, which provides valence, arousal, and dominance scores for each token.

Statistical analysis of the questionnaire responses (seven Likert items) revealed that the bot’s affective profile had a strong impact on participants’ self‑reported emotional change (items 5 and 6). The negative profile consistently lowered the valence and arousal of users’ utterances, increased the use of negative‑valence words, and reduced agreement expressions. Conversely, the positive profile raised valence, promoted positive emotion words and emoticons, and encouraged more user statements and fewer closed‑question replies. Basic interaction metrics such as response time and word count were unaffected by the profile, indicating that the emotional effect was not due to changes in chat length or speed.

To uncover patterns across participants, the authors constructed correlation matrices from the time series of bot‑ and user‑generated valence and arousal values (3 bots × 91 users × 2 dimensions). By thresholding these matrices and applying network‑theoretic measures (clustering coefficient, modularity), they identified distinct communities of users whose emotional responses were highly synchronized for a given bot profile. This network‑based clustering demonstrated that emotional similarity can give rise to mesoscopic structures even in a small, experimentally isolated setting.

Building on these empirical findings, the study extended the scenario to a large‑scale agent‑based model (ABM). The ABM placed 10,000 agents on a scale‑free social graph (average degree ≈4). Each agent carried a dynamic emotional state (valence, arousal) and exchanged messages with its neighbors. The three experimental bots were embedded as fixed nodes that periodically emitted messages reflecting their assigned affective profile. Agents updated their emotional state probabilistically based on the difference between their own valence and the incoming message, mimicking affective contagion.

Simulation results mirrored the laboratory data: a positive bot raised the network‑wide average valence by 0.3–0.5 units, while a negative bot depressed it by 0.4–0.6 units. Time series of collective emotion exhibited long‑range temporal correlations, with power‑spectral densities following a 1/f^β law (β≈0.7), indicative of persistent fluctuations. Emotional clusters emerged around the bot nodes and expanded through high‑degree hub agents, confirming that network topology critically shapes the speed and reach of affective spreading. The simulations also showed that the presence of an affective bot can induce a collective emotional state that persists long after the bot’s direct messages cease, highlighting the potential for bots to steer community mood.

Overall, the research demonstrates a two‑level impact of emotional chatbots: (1) they can reliably induce targeted emotional changes in individual users, and (2) through the social graph, these changes can cascade, producing correlated, long‑lasting collective emotions. The findings suggest practical applications in online community management—such as using positively‑toned bots to foster constructive discourse or deploying negatively‑toned bots to detect and mitigate harmful sentiment. Moreover, the work underscores the importance of integrating machine‑learning‑based affect detection with statistical‑physics models to capture the emergent dynamics of emotion‑laden online interactions. Future directions include validation on real‑world social‑media data, incorporation of additional affective dimensions (e.g., fear, surprise), and development of adaptive bot strategies that respond in real time to the evolving emotional landscape of the network.