Danmaku, a commenting system that displays users’ synchronous or asynchronous comments within the videostream, is widely used in Asian countries, especially in China and Japan (Wu & Ito, 2014). In a Chinese online talk show Roast! that features mock impoliteness speech events, Danmaku comments provide easy access to a vast amount of third-party participants’ evaluations of mock impoliteness. Particularly in Roast!, the Danmaku system also allows users to vote for the comments by clicking on a thumbs-up button at the end of each comment and automatically displays the count of such votes. This possibly unique system renders the phenomenology of the Danmaku a two-stage process: i) the third-party participants make qualitative evaluations about the mock impoliteness speech events; and then ii) the third-party participants vote for the evaluations made via the form of Danmaku, leading to the number of votes. Thus, the aim of this paper is to investigate: i) what factors contribute to the qualitative evaluations; and ii) what factors contribute to the number of votes that each comment gets. By creating an effective coding scheme which captures intertextual references, sequence of speech events, emotive stances, count of votes, etc. a large number of raw data was annotated for further analysis. Then, a conditional inference tree model (Hothorn et all 2006; Tagliamonte and Baayen 2012; Tantucci and Wang 2018), which is “a method for regression and classification based on binary recursive partitioning” (Levshina 2015:291), was fitted to answer these two questions. This method generated clear data visualizations by displaying the ranking of contributing factors to i) qualitative evaluation; and ii) the number of votes. The quantitative results are then interpreted in terms of qualitative analysis of typical examples from the data.