EAA 2026 Abstracts

Area 1 - EAA

Full Papers

Paper Nr:	5
Title:	Graph-Based Pixel Representation Using GCN for Semantic Face Segmentation
Authors:	Agnieszka Polowczyk, Alicja Polowczyk, Marcin Woźniak and Michał Wieczorek
Abstract:	Image segmentation is widely used in different fields, marking regions of interest accordingly. The most popular architectures used in this problem are convolutional neural networks, which respectively first encoder features, then reconstruct them into a labeled image. However, such methods can have problems capturing the relationships, details and relationships of objects present in an image. In this work we propose our multilayer Graph Convolutional Network in which input images are mapped into a graph structure, which provided an opportunity to use the mechanism of aggregating information from neighbors. We conduct experiments on FASSEG Instances dataset and show that our model outperforms the classic U-Net in terms of accuracy and efficiency, achieving a higher Dice score for more categories and getting mDice = 74.17%. In addition, our proposed architecture achieved a higher Accuracy = 91.24%. One of the key strengths is the significant reduction in the number of parameters required for training, from 31 032 265 (U-Net) to 1 058 313 (our model), representing a reduction in complexity of 96.59%. So our solution opens up new possibilities for creating lightweight and efficient models in image segmentation problems, surpassing U-Net, considered the benchmark in this field.

Paper Nr:	6
Title:	A Multi-Agent Architecture for Emotion Generation and Regulation in Conversational Agents: An Inside-Out Approach
Authors:	Aaron Pico, Joaquin Taverner, Emilio Vivancos, Ana Garcia-Fornes and Vicente Botti
Abstract:	This paper proposes a multi-agent systems based on large language models (LLMs) to simulate emotional processes. We present EMAD, an architecture designed to model both the generation and regulation of emotions, explicitly grounded in the theories of Paul Ekman and James J. Gross. The proposal uses a set of heterogeneous and specialized agents that collaborate to simulate different components of emotional processing. The core of the system consists of a set of emotional agents, each one specialized in interpreting information from the evaluative framework characteristic of one of the basic emotions described by Ekman. These agents are responsible for the attention and evaluation phases and cooperate to interpret information through a weighted voting mechanism. This mechanism reproduces Ekman’s refractory period: the relative influence of each agent is dynamically adjusted according to the current emotional state, so that the dominant emotion biases perception and cognitive evaluation as in Ekman’s theory . The process also includes a supervisory meta-agent, responsible for monitoring the overall state of the system. This meta-agent can activate regulatory agents capable of intervening in the attention process or in cognitive evaluation, thus modulating emotional dynamics. Through this approach, the system implements a form of emotional self-regulation inspired by Gross’s framework.

Paper Nr:	7
Title:	Toward Relationship-Aware Architectures for Large Language Models
Authors:	Alba Baeza-Bosca, Alma Sena, Ricardo Peris and Joaquin Taverner
Abstract:	Large language models can generate fluid and apparently empathetic dialogues, but they often confuse the user’s momentary feeling with the quality of the relationship. This limitation causes them to react shorttermed to negative emotions and prevents them from modeling the dynamics of bonds over time. In this position paper, we propose a structural paradigm shift and propose a relationship-aware architecture that explicitly separates the emotional axis, denoting immediate affection, from the relational axis, representing the evolving emotional bond. The system maintains a dynamic affective link between the agent and the user, stored in a GraphRAG-based knowledge graph, which serves as a compact representation of the relationship history, current status, and affetive bound. During real-time interaction, the agent retrieves an affective context, comprising its personality, and relational memories, to generate responses by jointly selecting an emotion and a communicative intention based on SPAFF and Turning-Toward theory. After each dialogue, a post-interaction module labels emotions, evaluates message pairs using an intention-emotion weighting matrix, and updates the affective link. A theoretical overview of the design illustrates how this framework allows agents to maintain functional affective links despite temporary negative emotions.

Paper Nr:	9
Title:	``This Is Me'': Restoring Vocal Identity and Regional Dialect for ALS Patients from Limited Legacy Audio
Authors:	Isabel Ferri-Molla, Jordi Linares-Pellicer, Carlos Aliaga-Torro and Juan Izquierdo-Domenech
Abstract:	Degenerative diseases often silence speech, stripping individuals of their unique identity and emotional expression. Standard assistive technologies can support functional communication, but they rarely replicate the user's distinctive acoustic character or regional accent. Futhermore, these type of technologies are unable to transmit emotions with the same breadth and fidelity as the person's natural voice. In this work, we address this limitation by reconstructing a personalised voice from a limited set of spontaneous voice notes recorded by the user before they lost their voice. While voice notes do not have the quality of a recording studio, they have the advantage of capturing the natural rhythm and timbre more effectively than scripted recordings. Our architecture integrates large language models to adapt text to specific dialectal nuances and to inject expressive and emotional signals. This processing allows learning from imperfect recordings and supports recovery of the speaker's original vocal tone. A longitudinal case study indicates that the system restores the characteristic energetic signature of the user's voice. These results suggest that emotional authenticity and regional identity can be brought back into communication even when only a limited set of data is available.

Paper Nr:	10
Title:	TransExplain-LD with Neuro-Symbolic Embeddings
Authors:	Nour El Houda Ben Chaabene, Zied Selmi and Hamza Hammami
Abstract:	Standard machine translation systems fail to preserve emotional nuance and cultural context. We introduce TransExplain-LD, a neuro-symbolic Transformer that dynamically adjusts depth based on emotional complexity and integrates hybrid embeddings fusing affective lexicons, contextual emotion vectors, and knowledge graphs. Template-guided decoding ensures syntactic coherence and interpretability. The system is optimized with a multi-objective loss balancing translation quality, emotional fidelity, and explanation clarity.