Learning to Refer: How Scene Complexity Affects Emergent Communication in Neural Agents

Dominik Künkele, Simon Dobnik


Abstract
We explore how neural network-based agents learn to map continuous sensory input to discrete linguistic symbols through interactive language games. One agent describes objects in 3D scenes using invented vocabulary; the other interprets references based on attributes like shape, color, and size. Learning is guided by feedback from successful interactions. We extend the CLEVR dataset with more complex scenes to study how increased referential complexity impacts language acquisition and symbol grounding in artificial agents.
Anthology ID:
2025.iwcs-1.26
Volume:
Proceedings of the 16th International Conference on Computational Semantics
Month:
September
Year:
2025
Address:
Düsseldorf, Germany
Editors:
Kilian Evang, Laura Kallmeyer, Sylvain Pogodalla
Venues:
IWCS | WS
SIG:
Publisher:
Association for Computational Linguistics
Note:
Pages:
299–307
Language:
URL:
https://preview.aclanthology.org/iwcs-25-ingestion/2025.iwcs-1.26/
DOI:
Bibkey:
Cite (ACL):
Dominik Künkele and Simon Dobnik. 2025. Learning to Refer: How Scene Complexity Affects Emergent Communication in Neural Agents. In Proceedings of the 16th International Conference on Computational Semantics, pages 299–307, Düsseldorf, Germany. Association for Computational Linguistics.
Cite (Informal):
Learning to Refer: How Scene Complexity Affects Emergent Communication in Neural Agents (Künkele & Dobnik, IWCS 2025)
Copy Citation:
PDF:
https://preview.aclanthology.org/iwcs-25-ingestion/2025.iwcs-1.26.pdf