The text is informal in social networks. One of the common phenomena is that there are a lot of morphs in social networks. People are keen on creating morphs to replace their real targets to avoid censorship and express strong sentiment. In this paper we aim to solve the problem of finding real targets corresponding to their entity morphs. We exploit the temporal and semantic and POS constraints to collect target candidates. Then we propose a method based on joint character-word training to sort the target candidates. Our method does not need any additional annotation corpora. Experimental results demonstrate that our approach achieved some improvement over state-of-the-art method. The results also show that the performance is better when morphs share the same character as targets.