<![CDATA[Retrieval-augmented approaches have become central in knowledge-grounded dialogue systems, yet incorporating topical preferences remains difficult due to privacy constraints on user interaction data. This study introduces a lightweight federated topic-preference (FedTP) mechanism that models session-level preferences without centralizing raw data and uses client-level differential privacy (DP). Using the Topical-Chat dataset (8,628 conversations), each conversation is treated as a client, and evidence routing is framed as selecting relevant knowledge snippets based on dialogue context. The proposed method augments a TF-IDF relevance score with a small preference-based component derived from both local session distributions and a DP-aggregated global prior. Experimental results on 9,553 grounded test turns show a consistent but limited improvement in evidence hit rate, from 0.6167 to 0.6194. The small optimal preference weight (λ = 0.005) indicates that the preference signal mainly influences decisions when competing candidates have similar relevance scores, rather than substantially altering routing behavior. A privacy–utility analysis under Gaussian DP (ε ranging from 9.69 to 0.606, δ = 1e−5) shows negligible changes in performance, which is expected given the large number of clients in a one-shot aggregation setting. Additional metrics remain largely stable, suggesting that the method affects selection margins rather than overall alignment. These findings suggest that federated preference aggregation can provide a modest, privacy-preserving bias for evidence routing, but its practical impact remains incremental and context-dependent.]]>
Copyrights © 2025
