Similarity-Based Action Retrievalin Intent-Based Systems : The Case of Dialogue-StateRepresentations
Year: 2026
Type: dissertation
Abstract: Autonomous network management requires systems that can retrieve relevant actionsfrom structured operational knowledge under changing and partially observed conditions.Exact symbolic matching provides traceability, but can be brittle when new states, incomplete information, or unseen combinations appear. This thesis investigates whetherVector-Symbolic Architecture (VSA) can serve as an efcient retrieval layer for structuredknowledge representations in autonomous network management.The proposed approach encodes Resource Description Framework-style subject-predicateobject triples into high-dimensional hypervectors using role-fller binding and bundling.These vectors are stored in associative memory and retrieved through similarity search, allowing structured states to be compared without relying solely on exact symbolic matches.The architecture is evaluated on MultiWOZ 2.4, used as a proxy for multi-domain structured state and action retrieval rather than as a direct telecommunications benchmark.The results show that the VSA-based method performs competitively in the in-domainsetting, achieving a Hit@1 of 0.668, compared with 0.677 for RDF2Vec and 0.617 forTF-IDF nearest-neighbor retrieval. More importantly, the ablation study shows thatstructural pattern fallback is a central part of the method: removing it reduced Hit@1from 0.6653 to 0.5631, indicating that abstract structural matching contributed substantially to retrieval performance and adaptation to unseen state combinations. In anexploratory comparison with a quantized Large Language Model, the VSA-based systemachieved approximately 11.2× lower mean latency while maintaining higher constrainedaction-label accuracy.The fndings suggest that VSA-based representations can provide an efcient and interpretable retrieval layer for vectorized knowledge graphs, especially when fast retrieval,structural matching, and direct memory insertion are important. However, the methoddoes not eliminate domain-shift challenges and should be understood as a candidate retrieval mechanism rather than a complete autonomous network reasoning system. Overall, role-fller binding, structural pattern fallback, and low-latency associative retrievalmake VSA a promising direction for further study in dynamic knowledge-driven systems.
Author: Eriksson, Elliot
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7164454974