This dissertation investigates how emotions unfold, are influenced, and can be modeled in text-based social interaction. We focus on the domains of customer service dialogues and conflict resolution in the face of negative stimuli (specifically microagressions). We argue that emotions are not static labels but dynamic, relational processes tied to events, social roles, and regulation strategies. To this end, data resources and models were developed that enable both in-context detection and real-time forward-looking inference of affective states.
Three complementary corpora are introduced. First, a large-scale Dutch Twitter resource (EmoTwiCS) contains 9,489 business-oriented customer service dialogues annotated for fine-grained emotions (28 categories and valence–arousal–dominance), explicit causes, and operator response strategies, enabling the study of within-conversation trajectories and strategy–emotion interactions at scale. Second, a bilingual Wizard of Oz corpus (EmoWOZ-CS) collects 2,148 human–wizard chat dialogues with operator-steered end valence, multilabel strategy tags, and both self-reported and third-party emotion annotations. It supports analyses of perspectivist divergences and proactive, strategy-aware modeling of future customer emotions from prior turns. Third, a role‑playing corpus (COPING) operationalizes coping theory by linking discrete emotions to action tendencies (attack, contact, distance, reject) in responses to adverse verbal stimuli, providing behavioral grounding for affect modeling beyond surface labels.
Analyses across the three corpora yield several key insights. In social media complaints, emotion trajectories generally shift from negative or neutral openings toward a more neutral or occasionally positive tone. While anger and annoyance tend to persist and are slow to dissipate, gratitude and joy remain relatively stable throughout the interaction. These patterns highlight the importance of modeling disagreement and uncertainty in affective interpretation. Moreover, operator strategies play a crucial role in shaping emotional dynamics: problem‑focused tactics (such as providing explanations, requesting information or action, and offering assistance) are associated with increased neutrality and expressions of gratitude. Cheerful communication fosters positive reciprocity, whereas other affective strategies such as apologies or expressions of empathy sometimes prove less effective. In contrast, suboptimal replies (such as miscomprehensions, non‑collaborative tones, or ironic remarks) tend to amplify anger, annoyance, disappointment, desire, and confusion. These findings support the recommendation to prioritize problem‑focused interventions before attempting emotion‑focused repair.
Self‑reported and third‑party emotion annotations show notable divergence: agreement is highest for gratitude and neutral states, mixed for anger, and lowest for more subtle emotions. Mapping emotions to behavioral responses shows interpretable but uneven patterns: contact is most reliably inferred, while rejection is the least. Distinct lexical anchors are associated with attack and contact. Self‑reports link anger and frustration to attack, hope to contact, and fear and distress to distancing behaviors.
In terms of machine learning, the dissertation investigates context-aware emotion detection and forward-looking inference. Sequence-sensitive decoding yields small, yet consistent gains over turn-isolated baselines for emotion trajectories. However, forecasting the customer’s next emotional state from only prior context remains challenging, with valence being easier to predict than discrete emotion categories. In this setting, zero-shot large language models underperform supervised encoders in accuracy, but the former appear sensitive to latent negative affect that may remain underlexicalized, pointing to an intrinsic gap between internal affect and its textual realization in real-time settings. Finally, coping-aware modeling proves feasible: contact and distancing tendencies are recoverable in text, though overall ambiguity and subjectivity remains.
Together, the resources, analyses, and models presented contribute a process-centric foundation for emotion-aware NLP: they connect events to affective trajectories, quantify strategy–emotion interactions, expose perspectivist variation, and enable proactive guidance in time-sensitive dialogues. The findings argue for integrating behavioral grounding, perspectivist supervision, and forward-looking inference in future systems that aim not only to detect, but to skillfully regulate emotions in social interaction.