Knowledge graphs (KGs) offer a structured framework for representing biomedical entities and their interrelations, supporting advanced clinical decision-making and research. However, biomedical KGs often suffer from noise, incompleteness, and lack of contextual specificity, especially in oncology. This paper presents a novel approach for adaptive refinement of oncology-focused KGs using distant learning from heterogeneous textual sources, including PubMed abstracts and clinical trial summaries. We employ a combination of semantic parsing, relation extraction, and graph embedding techniques to identify and incorporate high-confidence relations while suppressing spurious connections. A reinforcement learning-based module adaptively filters candidate triples using contextual cues and domain-specific priors. Experimental results on curated cancer-related datasets demonstrate improved precision in entity linkage, relation quality, and downstream tasks such as question answering and evidence retrieval. Our framework enhances the utility of biomedical KGs for precision oncology by dynamically evolving the graph structure with minimal manual supervision.