What are the building blocks of language stored in memory and how are they utilized in linguistic tasks? It has been proposed that meaningful strings of all lengths—morphemes, words, and sequences of multiple words—can be stored, with the last kind playing a crucial role in language processing. This dissertation investigates the existence of stored multiword representations and their usage in Chinese reading. Stored multiword representations are operationalized by using two words that frequently co-occur and comparing them with those that do not. Morphosyntactic structure is also manipulated, with the main comparison between noun-noun and verb-object sequences. Two tasks are used to study visual recognition of multiword sequences: (1) a rapid masked visual presentation without sentence context probes how many words in a string can be simultaneously recognized and whether this limit is modulated by the co-occurrence frequency of the two words in the string; (2) a naturalistic sentence reading task with a gaze-contingent boundary change paradigm probes how far/deep Chinese ix readers process downstream text not yet directly fixated (i.e., parafoveal processing) and whether this limit is modulated by the co-occurrence frequency of the two words in the downstream string. The results show that co-occurrence frequency facilitates rapid visual recognition without sentence context, making parallel recognition of the two embedded words possible. This is the case for both noun-noun and verb-object sequences. In sentence reading, however, co-occurrence frequency influences online processing differently for strings of different structures. It facilitates processing extremely early on for noun-noun sequences: readers process Characters n+3 and n+4 in the parafovea beyond the visuo-orthographical level, while no such evidence is found for verb-object sequences. However, relatively late foveal processing does appear to be facilitated by co-occurrence frequency, for both kinds. Based on the findings, I argue that while language users are highly sensitive to statistical regularities of word sequences of various structures, this possibly yields only familiarity with the surface multiword forms and ease of on-the-fly composition of the two embedded words for verb-object sequences. Compound nouns on the other hand may be lexically stored to have direct form-meaning mapping via frequent exposure, hence the additional early facilitation observed. Future models of Chinese reading must incorporate mechanisms to explain the current data: (1) extremely fast access to frequently co-occurring strings’ orthography, which suggests that a single decomposition route alone is likely insufficient; (2) distinctive processing patterns throughout the time course between noun-noun and verb-object sequences, which points toward structural/semantic composition in addition to utilization of word statistics and contextual probability.