This paper presents a novel approach to enhance multi-hop knowledge graph reasoning (KG-R) through reward shaping techniques. The research addresses the challenges of incomplete knowledge graphs, which often lead to incorrect inferences. By partitioning the UMLS benchmark dataset into rich and sparse subsets, the study investigates the effectiveness of pre-trained BERT embeddings and prompt learning methods to refine the reward shaping process. This approach improves the accuracy and robustness of knowledge inference in complex KG frameworks. The paper introduces a framework that uses reinforcement learning (RL) with the REINFORCE algorithm to navigate KGs for multi-hop query resolution. However, the on-policy nature of REINFORCE can lead to suboptimal policies due to the incompleteness of practical KGs. To address this, the study proposes a reward shaping strategy that leverages existing KG embedding models for KG completion. These models map entities and relations to a vector space and estimate the likelihood of facts using a composition function. The research also explores the use of transfer learning within the UMLS dataset. By pretraining a "Reward Shaper" module on a rich KG and applying it to a sparse KG, the study aims to enhance the generalization capabilities of the reward-shaping mechanism. The study compares different methods, including BERT contextualization and prompt learning, and finds that prompt learning-based reward shaping outperforms traditional methods in most metrics. The experiments show that the proposed reward shaping approach significantly improves the performance of RL agents in multi-hop reasoning tasks. The study also highlights the importance of contextual understanding in KGs and the potential of transfer learning in enhancing reward shaping. The results demonstrate that training the Reward Shaper on a sparse KG can yield better performance than on a dense KG, suggesting a potential overfitting issue when using a rich Reward Shaper. The study's findings contribute to the field of KG reasoning by offering a new perspective and methodological advancement.This paper presents a novel approach to enhance multi-hop knowledge graph reasoning (KG-R) through reward shaping techniques. The research addresses the challenges of incomplete knowledge graphs, which often lead to incorrect inferences. By partitioning the UMLS benchmark dataset into rich and sparse subsets, the study investigates the effectiveness of pre-trained BERT embeddings and prompt learning methods to refine the reward shaping process. This approach improves the accuracy and robustness of knowledge inference in complex KG frameworks. The paper introduces a framework that uses reinforcement learning (RL) with the REINFORCE algorithm to navigate KGs for multi-hop query resolution. However, the on-policy nature of REINFORCE can lead to suboptimal policies due to the incompleteness of practical KGs. To address this, the study proposes a reward shaping strategy that leverages existing KG embedding models for KG completion. These models map entities and relations to a vector space and estimate the likelihood of facts using a composition function. The research also explores the use of transfer learning within the UMLS dataset. By pretraining a "Reward Shaper" module on a rich KG and applying it to a sparse KG, the study aims to enhance the generalization capabilities of the reward-shaping mechanism. The study compares different methods, including BERT contextualization and prompt learning, and finds that prompt learning-based reward shaping outperforms traditional methods in most metrics. The experiments show that the proposed reward shaping approach significantly improves the performance of RL agents in multi-hop reasoning tasks. The study also highlights the importance of contextual understanding in KGs and the potential of transfer learning in enhancing reward shaping. The results demonstrate that training the Reward Shaper on a sparse KG can yield better performance than on a dense KG, suggesting a potential overfitting issue when using a rich Reward Shaper. The study's findings contribute to the field of KG reasoning by offering a new perspective and methodological advancement.