The increasing demand for intelligent systems that can interpret and respond to human emotions has driven significant advancements in Natural Language Processing (NLP) and conversational agents. Such technology is crucial in mental health applications, where recognizing emotions and sustaining empathetic responses support effective diagnosis and therapy. However, capturing the complexity of human emotions and dialogue structures in therapeutic settings remains challenging, requiring sophisticated, adaptable models. This thesis introduces a novel framework to model and verify therapeutic dialogues using Stochastic Hybrid Automata (SHA) and Statistical Model Checking (SMC). The framework aims to classify dialogue acts, recognize emotional states, and maintain conversational coherence specific to therapy. Grounded in empirical data, it integrates probabilistic transitions and real-time simulation, along with therapeutic strategies tailored to different emotional dynamics. The contributions include: (1) a data-driven model generation approach in UPPAAL, automating dialogue and emotional patterns; (2) the development of strategy-specific variants (Baseline, Data-Driven, High Valence, and Low Valence) to simulate diverse therapeutic approaches; and (3) validation of the model’s alignment with real-world therapeutic interactions via UPPAAL probabilistic queries and simulations. Evaluation confirms the model's accuracy in reflecting conversational flow and emotional fluctuations typical in therapy. These findings highlight the potential of the UPPAAL-based framework for analyzing therapeutic dialogues and as a foundation for future developments in real-time emotional feedback and empathy-focused metrics. This thesis advances the intersection of conversational AI and mental health, suggesting pathways to improve digital therapeutic solutions.
La crescente richiesta di sistemi intelligenti capaci di interpretare e rispondere alle emozioni umane ha stimolato importanti progressi nell’elaborazione del linguaggio naturale e nei modelli conversazionali. Queste tecnologie sono essenziali in ambito psicologico e psichiatrico, dove riconoscere accuratamente le emozioni e rispondere con empatia è fondamentale per supportare diagnosi e terapie. Tuttavia, rappresentare la complessità emotiva e dialogica dei discorsi in contesti terapeutici richiede modelli sofisticati e flessibili. Questa tesi propone un framework avanzato per la modellazione e verifica dei dialoghi terapeutici, basato su automi ibridi stocastici e tecniche di verifica statistica dei modelli. Il framework ha l’obiettivo di identificare atti dialogici, riconoscere stati emotivi e garantire coerenza nelle interazioni terapeutiche. Integrando dati empirici, transizioni probabilistiche e simulazioni in tempo reale, il modello supporta anche strategie terapeutiche mirate, in grado di modulare le dinamiche emotive nel dialogo. I principali contributi includono: (1) un processo di generazione automatica di modelli basato su dati reali, che riproduce schemi dialogici e dinamiche emotive in UPPAAL; (2) lo sviluppo di varianti strategiche (Baseline, Data-Driven, High Valence e Low Valence) per simulare approcci terapeutici diversi; e (3) la validazione del modello rispetto a dati osservativi, tramite query probabilistiche e simulazioni in UPPAAL. I risultati confermano l’efficacia del modello nel rappresentare il flusso conversazionale e le dinamiche emotive tipiche del contesto terapeutico. Questi risultati mostrano il potenziale del framework come strumento di analisi per i dialoghi terapeutici, e suggeriscono sviluppi futuri quali il feedback emotivo in tempo reale e metriche di analisi dell’empatia. Questa tesi rappresenta un avanzamento nell'intersezione tra intelligenza artificiale conversazionale e salute mentale, suggerendo percorsi per migliorare le soluzioni terapeutiche digitali.
Formal modeling of patient-therapist dialogues and emotions for NLP applications
DETTORI, FRANCESCO
2023/2024
Abstract
The increasing demand for intelligent systems that can interpret and respond to human emotions has driven significant advancements in Natural Language Processing (NLP) and conversational agents. Such technology is crucial in mental health applications, where recognizing emotions and sustaining empathetic responses support effective diagnosis and therapy. However, capturing the complexity of human emotions and dialogue structures in therapeutic settings remains challenging, requiring sophisticated, adaptable models. This thesis introduces a novel framework to model and verify therapeutic dialogues using Stochastic Hybrid Automata (SHA) and Statistical Model Checking (SMC). The framework aims to classify dialogue acts, recognize emotional states, and maintain conversational coherence specific to therapy. Grounded in empirical data, it integrates probabilistic transitions and real-time simulation, along with therapeutic strategies tailored to different emotional dynamics. The contributions include: (1) a data-driven model generation approach in UPPAAL, automating dialogue and emotional patterns; (2) the development of strategy-specific variants (Baseline, Data-Driven, High Valence, and Low Valence) to simulate diverse therapeutic approaches; and (3) validation of the model’s alignment with real-world therapeutic interactions via UPPAAL probabilistic queries and simulations. Evaluation confirms the model's accuracy in reflecting conversational flow and emotional fluctuations typical in therapy. These findings highlight the potential of the UPPAAL-based framework for analyzing therapeutic dialogues and as a foundation for future developments in real-time emotional feedback and empathy-focused metrics. This thesis advances the intersection of conversational AI and mental health, suggesting pathways to improve digital therapeutic solutions.File | Dimensione | Formato | |
---|---|---|---|
2024_12_Dettori_Executive Summary.pdf
accessibile in internet per tutti
Descrizione: Executive Summary
Dimensione
1.01 MB
Formato
Adobe PDF
|
1.01 MB | Adobe PDF | Visualizza/Apri |
2024_12_Dettori_Tesi.pdf
accessibile in internet per tutti
Descrizione: Master Thesis
Dimensione
5.01 MB
Formato
Adobe PDF
|
5.01 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/230464