Knowledge Graphs (KGs) are fundamental to digital ecosystems, facilitating data integration, annotation, and interoperability across diverse domains. However, their construction and update depend on the merging of heterogeneous data sources that can introduce ambiguities and contradictions, particularly when introduced conflicting perspectives. Existing contradiction detection methods primarily resolve conflicts by discarding one of the contradictory statements, often overlooking the possibility of a nuanced, composite truth. When contradictions explicitly violate KG semantics, they can be identified through logical reasoning. However, in the absence of domain-specific semantics, external contextual information becomes crucial for accurate detection. While advancements in KG representation learning (KGRL) methods, such as negation-aware KG embeddings, have improved semantic expressiveness, they are still limited in their ability to capture and reason about contradictions. Among neural-based KGRL are Language Models (LM), that still struggle with reasoning over conflicting information. This work explores how a neuro-symbolic approach – integrating symbolic and LM neural representations – can enhance contradiction detection and improve KGRL robustness, reliability, and predictive performance, paving the way for more trustworthy machine learning applications over KGs.