Similarity join is widely used in data cleaning, data integration and the detection of near duplicate Web pages. Existing similarity join algorithms fall into two categories:Threshold-based similarity join and Top-k similarity join. Top-k similarity join is suitable for applications in which the threshold is unknown in advance. The most efficient Top-k similarity join algorithm is Top-k-join, which is proposed by Xiao et al. In order to resolve the performance problemsof Topk-join, a novel Top-k similarity join algorithm Opt-join is proposed in this paper. By integrating the token batch processing technique into the existing event-driven framework, Opt-join reduces the cost of processing the prefix events. In addition, Opt-joinreduces the cost in hash lookup by switching the positions of the hash lookup and filtering operations. The correctness of the new algorithm is proved. Experimental results show that 1.28x-3.09xspeed-up is achieved by Opt-join compared with Topk-join. More importantly, with the increase of the record length or the k value, Opt-join surpasses Topk-join by a larger margin.