News, with a twist: Using contrastive learning to improve user embeddings for diverse news recommendations

News recommender systems (NRS) play a key role in delivering personalised content in fast-paced, high-volume environments. However, models optimised solely for accuracy often overlook important societal objectives such as fairness and diversity, leading to over-personalisation, biased exposure, and narrow content consumption. In this paper, we propose a contrastive learning framework for improving user representations in neural news recommendation. We build upon a bi‑encoder architecture and introduce self-supervised objectives that group semantically related news items by theme, encouraging the model to bring similar items closer in the embedding space while pushing dissimilar ones apart. This strategy mitigates embedding collapse and guides the model toward producing recommendations with broader topical coverage. We evaluate our approach on the MIND dataset, comparing against state-of-the-art neural models, including LSTUR and NAML. Our results show that the proposed method achieves competitive accuracy and yields measurable improvements in beyond-accuracy objectives, particularly in content diversity and exposure fairness. Our results demonstrate the potential of contrastive learning to support more balanced and responsible news recommendations.