对于竞赛众包，如何快速有效地对数量较多的备选创意进行筛选是亟须解决的问题。文章基于创意关系视角，从新颖性和可行性两个方面衡量创意质量，并采取去除低质量创意和保留高质量创意的筛选策略。首先针对以文本形式描述的创意，采用隐含狄利克雷分布主题模型，按照创意之间的差异化程度（新颖性）将创意划分为多个不同的主题；其次采用杰卡德系数计算方法构建每个主题及其所包含的全部创意的语义网络，并转化为对应的累积分布，进而通过Kolmogorov-Smirnov统计量度量网络之间的距离；最后对每个主题中的创意按照可行性的高低（网络距离的大小）进行排序，从而为创意筛选提供依据。进一步地，将此方法用于筛选3个竞赛众包中的创意，结果表明该方法在识别低质量创意中的精确率较高，在识别高质量创意中的召回率较高。相较于常见的基于文本长度、大众评价、SBERT模型的筛选方法，该方法在筛选的精确率和召回率上均表现更好。

In crowdsourcing contests, a big challenge is to screen good ideas from a large number of alternative ideas. To address this challenge, this study develops a relation-based intelligent method for idea screening (Rela-IdeaS). This method adopts the strategies of retaining ideas with high quality and eliminating ideas with low quality based on idea novelty and feasibility. Specifically, this study employed the latent dirichlet allocation method to classify the ideas into various themes based on their novelty. Furthermore, the Jaccard coefficient method was used to construct semantic networks of each idea theme and its included ideas. With the help of cumulative distribution functions, these semantic networks were transformed into cumulative distributions Moreover, we used Kolmogorov-Smirnov statistics to measure the distance between semantic networks of ideas and idea themes. Finally, the ideas in each theme were ranked for further screening based on their values of feasibility (i.e., network distance). We applied our method to screen ideas from three crowdsourcing contests. The results indicated that using the proposed method to screen the ideas with low quality has a good precision rate. For screening the ideas with high quality, the method has a good recall rate. Additionally, compared to the commonly used methods, i.e., methods based on idea text length, crowd voting, and the SBERT model, our method performed better in precision and recall rates.