This doctoral dissertation reports on an empirical study, which takes a feature-based approach and probes the L2 acquisition and processing of Chinese bei passives by adult English native speakers. In Chinese, an individual passive marker bei is used to mark passive constructions. Whilst historically used as a lexical verb, bei is in the process of being grammaticalised (i.e. semi-lexical) and hence contains a semantic component (Liu, 2012a). Three forms of bei passives and their semantic properties have been investigated: basic long bei passives (i.e. with an external argument), basic short bei passives (i.e. without an external argument), and the retained-object construction of bei. In total, 75 English native speakers with intermediate and advanced Chinese proficiency, and 33 native Mandarin Chinese speakers (serving as a control group) were tested by a series of on-line methods (a self-paced reading task and a reaction-time picture elicited word rearrangement task) and off-line methods (an untimed acceptability judgement task and a fill-in-the-blank task). The current study finds that the reconfiguration of target semantic features of bei is a gradual process and occurs feature-by-feature, depending on consistent and ample input-based evidence. This lends support to the Feature Reassembly Hypothesis (Lardiere, 2005, 2008, 2009). It is also found that morphosyntax-semantics mismatches lead to acquisitional difficulties, as predicted by the Bottleneck Hypothesis (Slabakova, 2008, 2009b), which shares a similar view to the Feature Reassembly Hypothesis. In addition, L1 English L2 Chinese learners are found to be subject to the formation strategy of English short passives, in line with Montrul (2001). A disjunction in L2 performance between off-line and on-line tasks has been found in the advanced learners, who show target-like on-line sensitivity to violations of semantic constraints on bei but fail to converge on the target grammar in off-line judgements. These findings are compatible with Ullman’s (2001, 2005) declarative-procedural model and suggest that the increase in convergence on real-time comprehension and production in the advanced learners is a result of the more involved procedural system. The general findings of the current study lend support to the view (Sorace, 2009; White, 2011) that representational and processing difficulties must be teased apart in L2 acquisition.