In everyday life, humans rely on working memory (WM) processes to make sense of relationships between linguistic elements that are not linearly adjacent. For example, to understand the sentence The dog that the cat chased is cute, we encode the referent the dog into WM, maintain and retrieve it after reading the verb chased in order to interpret the dog as the object of chased. Increased grammatical complexity of the sentence has been shown to increase the effort needed to retrieve the target representation (Gibson, 1998; Lewis and Vasishth, 2005). In this dissertation, I focus on how changing the properties of the target representation - its complexity and internal coherence - affects the effort needed for the three WM processes, especially the under-studied maintenance process. Understanding this interaction can help inform our understanding of how humans package and store linguistic units in WM and better characterize WM architecture during sentence comprehension. To give a multi-dimensional view of the impact on WM processes, I approached the question from both behavioral and neural perspectives. Furthermore, through EEG experiments, I examined the role of the Sustained Anterior Negativity (SAN) in maintenance during sentence comprehension. Chapters 2 and 3 investigate how encoding, maintenance and retrieval of complex noun phrase (NP) representations differ from those of their simple counterparts through both a behavioral and neural lens, respectively. Through self-paced reading experiments in Chapter 2, I demonstrate that while the effects of representational complexity on encoding and retrieval were inconsistent, having a complex target NP always facilitated maintenance. This facilitation results from two factors: increased distinctiveness of the complex representations leading to decreased vulnerability to interference, and increased activation level due to increased encoding effort. Chapter 3 describes an EEG experiment that followed up thebehavioral experiments in Chapter 2. Previous works (Fiebach et al., 2002; King and Kutas, 1995; Phillips et al., 2005) found the SAN when there was increased processing cost during maintenance. In this EEG experiment, the sentences with simple NPs elicited a sustained global negativity relative to sentences with complex NPs during maintenance, giving evidence that increased representational complexity of the target NP decreases maintenance effort. Chapter 4 investigates how the encoding, maintenance and retrieval of internally coherent NP representations differ from those of their incoherent counterparts using both behavioral and EEG experiments. Additionally, these experiments further explore the role of representational complexity and how it interacts with internal coherence. Two self-paced reading experiments revealed that having a coherent target NP facilitates encoding, early maintenance and to a lesser extent, retrieval. They also provided evidence for less effort during early maintenance of complex target NPs. Accordingly, in the EEG experiment, the conditions with simple or incoherent target NPs triggered a SAN during early maintenance relative tothe conditions with complex or coherent target NPs. All these experiments together suggest that having coherent features eases both encoding and early maintenance processes. Overall, the findings demonstrate that increasing complexity or coherence at the representation level consistently facilitates WM maintenance. They also show that the SAN provides a viable ERP index for processing effort during maintenance though its behavior might be modulated by predictive processes.