Mock impoliteness, a term encompassing a wide array of phenomena (e.g., banter, teasing, mocking, jocular mockery, jocular abuse/insults, humour, etc.), has long been grounded in the framework of (im)politeness. It has also been studied under terms such as “anti-normative politeness” (Zimmerman, 2003), “sociable rudeness” (Kienpointner, 1997) and “ritual abuse” (Parkin, 1980). Having attracted a plethora of scholarly attention for several decades (Leech, 1983; Culpeper, 2005, 2011; Culpeper et al., 2017; Mills, 2003; Grainger, 2004; Terkourafi, 2008; Haugh, 2010; Haugh & Bousfield, 2012), the heated debates of mock impoliteness center around (1) its theoretical grounding, (2) its definition, and (3) its relationship with genuine impoliteness, mock politeness and politeness. This thesis contributes to such debates by investigating mock impoliteness in the context of a Chinese game show featuring “roast”, which is of particular relevance to mock impoliteness, focusing on (1) How is mock impoliteness constructed?; and (2) How is mock impoliteness evaluated by the third-party participants?. In investigating the construction of mock impoliteness, this thesis adopts Culpeper (2011) and Culpeper et al. (2017)’s mixed messages and Spencer-Oatey (2002, 2005)’s rapport management as its theoretical frameworks (modification was made when necessary), following a general integrative pragmatics approach (Culpeper and Haugh, 2014; Haugh and Culpeper, 2018), which also takes multimodality and metalanguage into consideration. Evidence shows that mock impoliteness is constructed dynamically, and different types of mock impoliteness show a strong preference for targeting at hearers’ quality face. In investigating the evaluation of mock impoliteness, a specific feature of this data, that is, Danmaku, an online commenting system imbedded in the video frame, allows the access of a large amount of metapragmatic evaluations of mock impoliteness. An effective coding scheme that captures many dimensions of Danmaku data was created for analysis. Then an unusual approach to the data (at least in the field of pragmatics), a machine learning technique –– conditional inference tree model (Hothorn et al., 2006; Tagliamonte and Baayen, 2012; Tantucci and Wang, 2018) was adopted to answer the research question. This method generates clear data visualization based on statistical significance. The results demonstrate that Funniness and Impoliteness are the two most statistically significant factors of evaluations of mock impoliteness. With modification of the theoretical framework and investigation of a rather new type of data, the Danmaku data, this thesis makes both theoretical and methodological contribution to the field of (mock) (im)politeness while redressing the possible Anglocentric bias by offering solid empirical evidence in Chinese data.