This paper investigates the so-called bi-directional verbs in Mandarin Chinese such as zu ‘rent’ and jie ‘lend’ with a special focus on how the applied argument (extra argument) in ditransitive constructions is syntactically and semantically licensed in applicative constructions. In previous research (e.g., Tang (1979); Huang (2007); Tsai (2009)), Mandarin bi-directional verbs were categorized into two types: the give-type and the rob-type. On the one hand, according to Paul and Whitman (2010) the Mandarin give-type double object constructions could be captured by the low applicative construction. On the other hand, Tsai (2009) related the Mandarin rob-type double object constructions to the middle applicative construction, hence similar to Pylkkänen’s (2008) high applicative construction. In this study, the three diagnostics proposed by Pylkkänen (2008): verb semantics, the transitivity restrictions, and the passivization test, are applied to bi-directional verbs zu ‘rent’ and jie ‘lend’. Also, both the Xiang and Ba constructions are employed and prove that bi-directional verbs in double object constructions with inward reading can not be classified into rob-type verbs. This study then examines the semantic relationship between the applied arguments and the other sentence arguments, and whether the semantic property is captured by the syntactic structure. I conclude that the ambiguity of bi-directional verbs in Mandarin Chinese should be clearly attributed to the underspecified directionality of the applicatives as the case with Spanish alquila ‘rent’ as argued by Cuervo (2003). The extra argument of the bi-directional verbs is licensed by the low applicative construction, which expresses the meaning of “transfer”. The bi-directional meaning takes place since the low applicative head licenses both Source and Recipient. By contrast, the extra argument of the rob-type verbs is licensed by the high applicative construction, which encodes “affectedness”, not “transfer”. Thereafter, I argue that the directional reading of the rob-type verbs is not structurally encoded, but inferred conceptually.