Conversational Entity Retrieval from Knowledge Graphs: Benchmarking and Fine-Grained Relevance Aggregation

Conversational access to knowledge graphs requires identifying the entity that best answers a free-form query while accounting for graph structure and information introduced in earlier dialogue turns. We study this problem through \emph{Conversational Entity Retrieval from a Knowledge Graph} (CER-KG), an information-retrieval task formulated as a sequence of interdependent candidate-entity ranking problems. We introduce QBLink-KG, a benchmark adapted from manually written conversational reading-comprehension dialogues and grounded in DBpedia. Its verbose queries and multi-turn dependencies expose challenges that are not adequately addressed by conventional entity retrieval, multi-hop question answering, answer generation, or zero-shot prompting of large language models. We develop two learning-to-rank architectures that progressively address these challenges. NACER represents each candidate using lexical and semantic matching signals between the current query, preceding answers, and components of the candidate's immediate knowledge-graph neighborhood, including entities, predicates, categories, and literals. Building on the finding that both dialogue context and local lexical and semantic evidence are important, we propose DRAGON, a graph-aware model that avoids prematurely compressing all neighborhood evidence into a single candidate vector. DRAGON first constructs and prunes a local subgraph for each candidate response. Each node is assigned a sparse 12-dimensional feature vector measuring its lexical and semantic similarity to the current query and preceding responses. Graph-convolution layers propagate and filter these relevance signals according to the subgraph topology; multi-head self-attention then emphasizes the most informative neighborhood evidence before fully connected layers produce the candidate's ranking score. The model is trained with a listwise, rank-sensitive objective. On the QBLink-KG test set, DRAGON achieves Recall@1 of 0.7497, Recall@10 of 0.9744, and mean reciprocal rank of 0.8021. Relative to NACER, it improves Hits@1 by 12.5\% and mean reciprocal rank by 5.9\%. DRAGON also performs comparably to a fine-tuned cross-encoder while using only 7,537 trainable parameters rather than approximately 109 million and providing fourfold lower per-candidate inference latency in the reported setting. Ablations show substantial degradation when graph convolution or self-attention is removed and confirm that lexical and semantic evidence are complementary: either feature family alone produces markedly lower retrieval accuracy. These results establish an effective and computationally efficient framework for graph- and dialogue-aware retrieval over structured knowledge sources.