<p>The objective of personalised news recommendation is to assist users in discovering news items that align with their own interests. The majority of current methodologies acquire single-user embeddings by analyzing the user’s behavioral history to reflect their interests. However, during the process of making actual decisions, user behavior is often the result of multiple factors entangled together, including users’ diverse interests and the multidimensional nature of news content. This leads to the limitation that single-user embeddings may not adequately model users’ interests. To remedy this deficiency, in this paper, we propose a news intent disentanglement method to extract different intents from news, on top of which we model users long- and short-term interests in different intents of news. Meanwhile, due to the lack of manually labels for users interests, users’ long- and short-term interests are also often entangled. To address this, we propose a self-supervised method to disentangle them by providing explicit supervisory signals. Experiments on two real-world datasets demonstrate HDNR’s effectiveness in news recommendation.</p>