Tag recommendation for new resources is one of the most important issues discussed recently. Many existing approaches ignore text semantics and can not recommend new tags which are not in the training dataset (e.g., FolkRank). Some exceptional semantic approaches use a probabilistic latent semantic method to recommend tags in terms of topic knowledge (e.g., ACT model). However, they do not perform well because many entities in these models result in much noise. In this paper, we propose hybrid approaches in folksonomy to challenge these problems. Hybrid approaches are combination of Language Model (LM) for keyword based approach and Latent Dirichlet Allocation (LDA), Tag-Topic (TT) model and User-Tag-Topic (UTT) model for topic based approaches. Our approaches can recommend meaningful tags and can be used to discover resource implicit correlations. Experimental results on Bibsonomy dataset show that LM performs better than all other hybrid and non-hybrid approaches. Also the hybrid approaches with less number of entities (e.g., LDA with only one entity) perform better than those approaches having more entities (e.g., UTT with three entities) for tag recommendation task.