Studies of Mandarin relative clause processing have yielded mixed results. Subject relatives (SRCs) are found more frequent than object relatives (ORCs) in corpora; however, the word order of ORCs resembles canonical SVO simple sentences. The present study examined the experienced-based effect of structural frequency and regularity by conducting a simulation in a simple recurrent network, trained on 10000 simple and RC sentences, in the proportions found in Chinese Treebank 7.0. The model contained 18 input and output units and 36 hidden and context units. Network performance was assessed on the level of grammatical prediction error of the next word. Following 10000 training epochs, ORCs performance showed garden-pathing upon encountering disambiguating relativizer due to resemblance to the SVO order in the beginning. For SRCs, which have an irregular verb-first structure, error was high but dropped later at the relativizer, showing sensitivity to language statistics. Results reflect both frequency and regularity effects.