In the current digital era, where online information is increasingly abundant and heterogeneous, providing users with relevant and trustworthy content has become more challenging. In news recommendation, this challenge is intensified by the time-sensitive nature of the content and the susceptibility of users to social issues such as biases, filter bubbles, and misinformation. Personalized recommendation systems, such as collaborative filtering and content-based recommenders, aim to deliver relevant news recommendations aligned with user interests. However, many of these systems require additional user information, such as behavior patterns or search history, which poses limitations in privacy-sensitive environments or when access to such data is not available due to anonymous content access. In such cases, the absence of detailed user information can sometimes be compensated by the use of factual information and knowledge hidden in the news itself. In this project, we develop several news recommendation models that predict the suitability of one news article as a recommendation for another. All these models are restricted to information contained within news articles, which eliminates the need for additional user data. Our approach leverages Knowledge Graphs and Knowledge Graph Embedding Models to create embedding representations of named entities found within the titles and texts of news articles. These embeddings provide the foundation for extracting features that describe the relationships between pairs of articles, which we then use to train a classifier that predicts the quality of recommendation. The proposed recommendation models are evaluated using a human-annotated dataset that contains information about the quality of recommendations and the similarity of news pairings. These models are compared with a traditional recommendation approach that relies on text similarity rather than semantic relations between named entities in the news. Our findings indicate that Knowledge Graph-based recommendation models are effective in capturing semantic relations and providing recommendations of equal quality to text-only approaches. Furthermore, these models outperform text-based models when using only the entities from the news title, offering a robust framework with reasonable performance in contexts that lack additional user information.