Knowledge graph completion(KGC) aims to infer missing facts from observed relational triples, which demands jointly modeling textual semantics and relational structure. Existing structure-based methods capture graph topology but often struggle with sparse entities and long-tail relations, while text-based methods exploit entity and relation descriptions but weakly model relational neighborhoods. In this paper, we propose LQ-KGC, a latent-query refinement framework that enables iterative interaction between semantic inference and structural reasoning for KGC. Instead of treating all missing triples as latent variables, LQ-KGC constructs a restricted set of semantically plausible latent queries and alternates between two refinement steps: the language model predicts latent tail distributions from textual descriptions, while the graph model incorporates high-confidence pseudo tails as pseudo edges for relation-aware message passing. Meanwhile, structure-aware graph predictions are distilled back into the language model to refine semantic representations. Experiments on WN18RR and FB15k-237 demonstrate that LQ-KGC consistently improves over text-based and graph-based baselines, with additional analyses highlighting the effectiveness of pseudo-edge augmentation and latent-query construction.