Next-basket recommendation (NBR) aims to infer the next basket given a basket sequence. The existing NBR methods rely solely on item IDs and ignore the semantic information of items. What’s more, these methods only consider binary item relationships which are often in higher order in the NBR scenario. In this paper, we propose HEKP, which addresses these challenges by pretrained language model (PLM) and hypergraph. Specifically, we use PLM to encode the basket sequence by masked user prompt (MUP). However, PLM-based recommendation will degrade when encountering Out-Of-Vocabulary (OOV) items. To tackle this issue, we construct another knowledge tree prompt (KTP) as the explanation of those OOV items. Additionally, we design a multi-item relation encoder to model the high order correlations among items by building a hypergraph based on item similarities. Lastly, we design a frequency based gating module to recommend the next basket. Extensive experiments are conducted on HEKP on three datasets, and the results validate its effectiveness against state-of-the-art methods.