A multi-stage framework for emotionally intelligent dialogue systems using topological sentiment encoding and causal policy optimization
7 affiliations
- 1Computer Science & Engineering Department, MIT School of Computing, MIT Art, Design and Technology University, Pune, India.
- 2School of Computer Science and Engineering, Ramdeobaba University, Nagpur, India.
- 3Department of Artificial Intelligence, G H Raisoni College of Engineering and Management, Nagpur, Maharashtra, India.
- 4Department of CSE, New Horizon College of Engineering, Bengaluru, Karnataka, India.
- 5Department of Mechanical Engineering, Koneru Lakshmaiah Education Foundation, Vaddeswaram, Guntur, India.
- 6Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, India.
- 7Department of AI&ML, Gokaraju Lailavathi Engineering College, Hyderabad, India.
Abstract
Emotionally intelligent dialogue systems should not only express sentence-level sentiment but also sense shifts in emotion from sentence to sentence, from speaker to speaker, and over time, and be interpretable down to the decision level. Existing approaches consider only sentiment as a single and static feature, and they either do not consider multi-turn context, temporal variation, speaker-specific trends, or the correlations between emotion and action. To fill these gaps, we introduce a unified 5-stage framework for systematically encoding, transforming, and operationalising emotion in multi-turn conversations. First, Context-Aware Multi-Level Sentiment Encoding (CMLSE) uses RoBERTa embeddings, self-attention, and Dynamic Time Warping (DTW), while preserving the speaker identity, turn order and sentiment continuity. Second, Emotion Clustering Topological Signal Mapping (ECTSM) uses an emotion vector sentiment and persistent homology to obtain robust emotion clusters based on topology and shape. Third, Emotion-Memory Reinforced Graph Aggregation Network (EMR-GAN) constructs a dynamic emotion graph and a long-term memory network, learning the best manner for optimizing emotional coherence through Reinforcement Learning (RL). The Adaptive Causal Decision-Graph Transformer (ACDGT) incorporates causal inference into graph attention to discover hidden emotion–behaviour causality and imitate counterfactual decisions. Finally, the Emotion-Driven Reinforcement Policy Synthesizer (EDRPS) combines emotional memory and causal representations with dual-reward Proximal Policy Optimization (PPO) to generate policies in real-time, emotion-aware mode. Our model shows an average accuracy of 78.8% (+6.4% above the best baseline) on MELD, 0.61 Adjusted Rand Index (ARI) for emotion clustering, and 92.5% accuracy in aligning the policy on DailyDialog. These assessments show scores of empathy, relevance, and naturalness to be higher, marking a new state-of-the-art for affective dialogue systems.
Related
- Sentiment classification optimization based on improved Senti-BERT and BiLSTM-Attention
- Novel hole quality indices and intelligent optimization of abrasive waterjet drilling of Al 7075-T6 using grey wolf optimizer
- An embedded deep learning–fuzzy logic framework for intelligent and automated irrigation in smart farming
- GEO-BERT: A Retrieval augmented LLM Framework for Intelligent Dataset Retrieval from the Gene Expression Omnibus
- Intelligent immunotherapy: AI-driven optimization of CAR-T therapy in hematologic malignancies
BibTeX
@article{khan2026multistage,
title = {{A multi-stage framework for emotionally intelligent dialogue systems using topological sentiment encoding and causal policy optimization}},
author = {Amreen Habibullah Khan and Amit Pimpalkar and Tabassum H Khan and Sonali Mahure and Sireesha Koneru and Niraj Kumar Dewangan and Harika Vanam},
journal = {PLOS One},
year = {2026},
month = oct,
doi = {10.1371/journal.pone.0361365},
url = {https://doi.org/10.1371/journal.pone.0361365},
note = {CC BY},
}
RIS
TY - JOUR
TI - A multi-stage framework for emotionally intelligent dialogue systems using topological sentiment encoding and causal policy optimization
AU - Amreen Habibullah Khan
AU - Amit Pimpalkar
AU - Tabassum H Khan
AU - Sonali Mahure
AU - Sireesha Koneru
AU - Niraj Kumar Dewangan
AU - Harika Vanam
PY - 2026
DA - 2026/10/28
JO - PLOS One
DO - 10.1371/journal.pone.0361365
UR - https://doi.org/10.1371/journal.pone.0361365
AB - Emotionally intelligent dialogue systems should not only express sentence-level sentiment but also sense shifts in emotion from sentence to sentence, from speaker to speaker, and over time, and be interpretable down to the decision level. Existing approaches consider only sentiment as a single and static feature, and they either do not consider multi-turn context, temporal variation, speaker-specific trends, or the correlations between emotion and action. To fill these gaps, we introduce a unified 5-stage framework for systematically encoding, transforming, and operationalising emotion in multi-turn conversations. First, Context-Aware Multi-Level Sentiment Encoding (CMLSE) uses RoBERTa embeddings, self-attention, and Dynamic Time Warping (DTW), while preserving the speaker identity, turn order and sentiment continuity. Second, Emotion Clustering Topological Signal Mapping (ECTSM) uses an emotion vector sentiment and persistent homology to obtain robust emotion clusters based on topology and shape. Third, Emotion-Memory Reinforced Graph Aggregation Network (EMR-GAN) constructs a dynamic emotion graph and a long-term memory network, learning the best manner for optimizing emotional coherence through Reinforcement Learning (RL). The Adaptive Causal Decision-Graph Transformer (ACDGT) incorporates causal inference into graph attention to discover hidden emotion–behaviour causality and imitate counterfactual decisions. Finally, the Emotion-Driven Reinforcement Policy Synthesizer (EDRPS) combines emotional memory and causal representations with dual-reward Proximal Policy Optimization (PPO) to generate policies in real-time, emotion-aware mode. Our model shows an average accuracy of 78.8% (+6.4% above the best baseline) on MELD, 0.61 Adjusted Rand Index (ARI) for emotion clustering, and 92.5% accuracy in aligning the policy on DailyDialog. These assessments show scores of empathy, relevance, and naturalness to be higher, marking a new state-of-the-art for affective dialogue systems.
SN - 1932-6203
ER -