Knowledge Graph completion (KGC), which predicts missing triples to enhance knowledge graphs, holds significant application value in semantic search and intelligent question answering. However, existing methods typically rely solely on explicit neighborhood structures, neglecting deep semantic information, and are constrained by the input length limitations of language models in fully integrating context. To address these challenges, we introduce the CERS (Contextually-Enriched Relational Semantics Model) approach, which leverages Large Language Models (LLMs) to enrich entity descriptions and relation semantic representations: at the entity level, CERS retrieves neighboring triples of target entities and performs weighted sampling based on relation rarity, selecting the most representative triples to generate semantically enriched entity descriptions; at the relation level, through diverse sampling of associated entities, it constructs forward and reverse relation texts to capture bidirectional semantic features of relations; by integrating neighborhood context and relation characteristics into enhanced semantic descriptions, the model provides more effective information for training. Experiments demonstrate that CERS outperforms mainstream baseline models on the WN18RR and FB15k-237 benchmark datasets across key metrics, including MRR and Hits@1/3/10, with ablation studies further validating the crucial role of multi-level neighborhood information and bidirectional relation descriptions in performance enhancement.