A number of regions in the human brain robustly respond to linguistic input. These regions include most prominently regions on the lateral surface of the left frontal, temporal and parietal cortices, but also a number of other cortical, subcortical and cerebellar regions (Figure 1A). Originally discovered in patients with brain damage (e.g., Broca, 1861; Dax, 1863; Wernicke, 1874/1969; Geschwind, 1970), these regions have been observed in PET and fMRI since the earliest days of brain imaging research (e.g., Petersen et al., 1988; Petersen and Fiez, 1993; Binder et al., 1997). These regions are consistent (albeit variable in their exact topography; e.g., Fedorenko et al., 2010) across individuals (e.g., Frost et al., 1999; Allendorfer et al., 2012), languages (e.g., Chee et al., 1999a,b; Illes et al., 1999; Klein et al., 1999; Hernandez et al., 2001; Pu et al., 2001; Hasegawa et al., 2002; Chee et al., 2003; Mahendra et al., 2003; Briellmann et al., 2004; see e.g., van Heuven and Dijkstra, 2010 and Sebastian et al., 2011 for reviews), modality of presentation (e.g., Chee et al., 1999c; Pinel et al., 2007; Buchweitz et al., 2009; Fedorenko et al., 2010; Braze et al., 2011) and developmental experiences, including complete sensory deprivation in the auditory or visual modality (Neville et al., 1998; Newman et al., 2010; Bedny et al., 2011). Furthermore, these regions can be quickly (in ~10–15 min) and reliably identified in individual participants (Fedorenko et al., 2010), and they are stable within an individual over time (Figure 1B; Fedorenko et al., 2010; Mahowald and Fedorenko, in preparation), as well as being robust to changes in the materials, modality of presentation, and task (Figure 1C), and language for bilingual speakers (Figure 1D).

This set of brain regions can be identified with a variety of contrasts that compare a more language-like stimulus with a less language-like stimulus¹ (e.g., words vs. fixation or tones—Binder et al., 1997; Diaz and McCarthy, 2009; words vs. pseudowords—Petersen et al., 1990; sentences vs. fixation—Kuperberg et al., 2003; sentences vs. false font or consonant strings—Bavelier et al., 1998; Robertson et al., 2000; Noppeney and Price, 2004; sentences vs. lists of words or pseudowords—Snijders et al., 2009; Fedorenko et al., 2010; Fedorenko and Kanwisher, 2011). Language-like-ness can be operationalized in terms of the amount of overlap between the stimulus and natural language. For example, phonotactically legal pseudowords and words match the sound-level properties of natural language, real words further match the lexical representations, and phrases or sentences match both lexical representations as well as larger structural / meaning units. And the process of language comprehension can be thought of, at least in part, as finding matches between the input and the stored language knowledge representations, with more/better matches leading to greater responses. In the remainder of the paper I will refer to this set of brain regions as the “language system”^2,3.

The stability of language activations within individuals across time and their robustness to variation in many properties of the defining contrast suggest that the language system may constitute a “natural kind,” i.e., a meaningful and stable subset of the brain. Two further lines of evidence suggest that these regions constitute an integrated functional system⁴. The first comes from studies of resting-state functional correlations, often referred to as “functional connectivity” (e.g., Fox and Raichle, 2007). In particular, the entire language system discussed above consistently emerges in the analyses of low-frequency oscillations across the brain during rest (e.g., Turken and Dronkers, 2011; Newman et al., 2013; Blank et al., submitted; see e.g., Catani et al., 2005, for DTI data consistent with the idea that these regions form a network). Although the interpretation of resting-state correlation patterns is still debated, these correlations appear to capture stable aspects of the functional organization of the human brain that persists across different mental states including sleep (e.g., Horovitz et al., 2008) and anesthesia (e.g., Vincent et al., 2007), and in some cases goes beyond known anatomical connections (e.g., Honey et al., 2009).

The second line of evidence comes from investigations of cortical thinning patterns in primary progressive aphasia, a neurodegenerative condition that disproportionately, and perhaps selectively, affects language processing (e.g., Mesulam, 2001; Grossman and Ash, 2003; Gorno-Tempini et al., 2004). The pattern of cortical thinning in this disorder—especially in the semantic variant—is strikingly similar to the functional activations for the contrasts, like e.g., sentences > pseudoword lists (e.g., Listerud et al., 2009; Rohrer et al., 2009; Dickerson, 2011; Gorno-Tempini et al., 2011; Rogalski et al., 2011). The precise mechanisms of degeneration constitute an area of active research, but one influential proposal that has been put forward argues for propagation along transsynaptic connections (Seeley et al., 2009).

In summary, a set of regions in the human brain (a) robustly respond to language input (with responses decreasing as the stimulus becomes less language-like, or when attention is drawn away from the linguistic properties of the stimulus and toward its perceptual features); (b) show strong correlations in their time courses during rest; and (c) are jointly susceptible to neurodegeneration in primary progressive aphasia. Together, these sets of findings suggest that these regions constitute a functional system. Given that these brain regions get activated by linguistic input and given that damage to these regions in mature brains leads to language deficits (e.g., Damasio, 1992; Bates et al., 2003), it is natural to assume that they play an important (and causal) role in interpreting the linguistic signal, although some components of this system have been argued to not be exclusively engaged by language but to instead support more abstract semantic processing (e.g., Hagoort et al., 2004; Patterson et al., 2007; Binder et al., 2009).

A number of regions in the human brain have been implicated in a broad range of goal-directed behaviors (e.g., Posner and Petersen, 1990; Cabeza and Nyberg, 2000; Corbetta and Shulman, 2002; Cole and Schneider, 2007; Duncan, 2010). These regions include parts of the dorsolateral prefrontal cortex (along the inferior frontal sulcus/middle frontal gyrus), parts of the insular cortex, regions along the precentral gyrus (going inferiorly to the posterior aspects of the inferior frontal gyrus, IFG), pre-supplementary and supplementary motor area, parts of the anterior cingulate, and regions in and around the intraparietal sulcus (Figure 2). This set of regions—with sometimes slightly differing inclusion criteria and/or subdivisions—is referred to in the literature by many names, including “task-positive network,” “cognitive control network,” “fronto-parietal attention network,” and “multiple-demand (MD) system.” Following Duncan (2006, 2010), I will use the term “multiple-demand (MD) system” throughout the paper, but none of the arguments in section Narrowing Down the Hypothesis Space for the Relationship between Language Processing Mechanisms and Cognitive Control Mechanisms hinge on the details of any particular proposal about these regions.

In brain imaging investigations, difficulty contrasts across many manipulations have been shown to activate the MD system. For example, Duncan and Owen (2000; also Duncan, 2006) performed a meta-analysis of activation peaks from neuroimaging studies that manipulated (i) the number of items held in memory (more vs. fewer items), (ii) the duration of holding information in memory (long vs. short), (iii) inhibitory demands (high vs. low), (iv) task novelty (new vs. practiced tasks), and (v) perceptual difficulty (difficult to perceive vs. easy to perceive). Across all of these manipulations, activations were observed in the frontal and parietal MD regions. More recently, Fedorenko et al. (2012, 2013; see also Wojciulik and Kanwisher, 1999; Stiers et al., 2010) provided evidence for overlap among diverse demanding tasks at the single-subject level, ruling out the possibility that the overlapping regions that emerged in the earlier meta-analyses were simply an artifact of spatial averaging across studies (e.g., Nieto-Castañon and Fedorenko, 2012). Furthermore, a set of brain regions that very much resembles the MD system also emerges in the resting-state correlation data (e.g., Power et al., 2011).

Even stronger evidence of domain-generality comes from single-cell recording studies in non-human primates, which have shown that many neurons in the frontal lobes exhibit substantial flexibility, varying their response properties according to task demands (e.g., Freedman et al., 2001; Miller and Cohen, 2001; Duncan, 2001; Cromer et al., 2010). For example, Freedman et al. (2001) trained macaques to categorize visual stimuli according to one dimension (cats vs. dogs). Following training, a substantial proportion of frontal neurons responded categorically to the relevant dimension. However, after training on a new task that used the same stimuli but required attention to a different dimension of the stimuli, the same neurons that previously categorized stimuli into cats and dogs now showed categorical responses to the new task-relevant dimension (see also Roy et al., 2010). These and other results suggest that these frontal neurons adapt the information they code to fit current goals. Similar “adaptive coding” has been reported in the parietal cortex (Freedman and Assad, 2006).

How do MD regions support complex behaviors? As of now, this remains an open question. Some notions that have been prominent in the literature in the context of this system include attention (e.g., Posner and Petersen, 1990; Desimone and Duncan, 1995; Petersen and Posner, 2012), working memory (e.g., Goldman-Rakic, 1995), cognitive control (e.g., Miller and Cohen, 2001; Koechlin et al., 2003; Badre and D'Esposito, 2009), structure building/unification (e.g., Hagoort, 2005), timing and/or sequencing (e.g., Luria, 1966; Janata and Grafton, 2003; Fuster, 2008), attentional episodes in goal-directed behavior (Duncan, 2010), and conscious awareness (e.g., Dehaene and Changeux, 2011), among others. However, most existing proposals are generic enough to be compatible with a wide range of data patterns. Nonetheless, whatever the precise computations conducted by the MD regions turn out to be (see e.g., Rigotti et al., 2010, for a proposal), this system is clearly of fundamental importance to humans, having been causally linked to fluid intelligence (Woolgar et al., 2010).

Given the spatial extent of the MD system and the cytoarchitectonic and connectomic diversity of its regions, many researchers have attempted to divide the MD system into sub-systems (e.g., Botvinick et al., 2001, 2004; Koechlin et al., 2003; Dosenbach et al., 2007; Badre and D'Esposito, 2009), and/or to map specific components of the system onto particular mental functions [e.g., Aron et al. (2004) argued that the right IFG plays a critical role in cognitive control; Novick et al. (2005) made a similar argument for the left IFG]. Claims of dissociations among different executive functions have also been made based on behavioral work in healthy participants (individual differences and dual-task paradigms; e.g., Engle et al., 1999; Miyake et al., 2000) and in brain-damaged individuals (e.g., Vallar and Shallice, 1990; Hamilton and Martin, 2005). Correlations across regions in resting functional data have also been taken to argue for a fractionation of this system (Power et al., 2011). However, the broad similarity in functional responses among the MD brain regions is striking. As a result, for the purposes of this paper I consider the MD system as a whole, while allowing for the possibility that only a subset of this system may end up being important for language (see section The Hypotheses Space for the Relationship Between Language Processing Mechanisms and Cognitive Control Mechanisms).

In recent work we investigated the relationship between high-level language processing brain regions and domain-general cognitive control brain regions (Fedorenko et al., 2011): we defined the regions of the language system using the sentences > pseudoword lists contrast and then examined the responses of those functionally-defined regions of interest (fROIs) to a number of non-linguistic cognitive tasks that have been previously argued to share machinery with language processing. With the exception of one region (the LMFG fROI), language-responsive regions showed no response to arithmetic processing (see also Monti et al., 2012), working memory, or cognitive control tasks, like Stroop. However, the latter set of tasks robustly activated the MD system, whose subsets are located in close proximity to language-responsive regions, especially in the left frontal lobe (Fedorenko et al., 2012). Based on these results, we argued that language-responsive regions are functionally specialized for linguistic processing and require language input to drive them.

There are at least three possible objections to these results and their interpretation, and I attempt to briefly address them here. The first objection is as follows: perhaps some regions of the MD system do respond to sentences more than pseudoword lists—just like the language regions do—but the response is weaker and/or more variable across individuals. This would lead us to miss some language-responsive regions and thus to potentially miss overlap between responses to language and demanding cognitive tasks. In Fedorenko et al. (2011) we tried to ameliorate this concern by performing a whole-brain search—at liberal thresholds—for overlap between the responses to the language contrast (sentences > pseudoword lists) and the hard > easy contrast in each of the non-linguistic demanding tasks. This search did not reveal much beyond what the basic analysis of the language-responsive fROIs had already shown: (i) language and verbal working memory showed overlap in the LMFG fROI; and (ii) there was a small region of overlap between language and two of the tasks (verbal working memory and Stroop) in posterior and dorsal-most part of the LIFG fROI, possibly due to the fact that MD regions abut the language-responsive parts of LIFG dorsally and posteriorly (see also Fedorenko et al., 2012), with spatial smoothing leading to the appearance of overlap (this possibility remains to be tested empirically using high-resolution scanning of the frontal cortex; cf. Schwarzlose et al., 2005). In Fedorenko et al. (2013; see also Supplementary Material), we report an analysis that shows that MD regions in fact respond to sentences and pseudowords in the opposite way from the language regions: they respond more strongly to linguistically degraded stimuli (pseudoword lists) than to linguistically meaningful and structured stimuli (sentences), suggesting that the language and the MD systems are spatially and functionally distinct.

The second objection is that perhaps the sentences we use in our “language localizer” task (Fedorenko et al., 2010) are too simple and don't contain a sufficient number of features that have been shown to cause comprehension difficulty, such as non-local dependencies (e.g., Gibson, 1998; Grodner and Gibson, 2005), lexical and/or structural ambiguity (e.g., Frazier, 1987; MacDonald et al., 1994), or low-frequency words or constructions (Preston, 1935; Forster and Chambers, 1973; Jurafsky, 1996; Levy, 2008). Maybe if more such features were present in the sentences, we would observe greater overlap between language and MD activations, which would manifest as (a) a greater response to MD tasks in the language regions, and/or (b) a greater response to sentences than pseudoword lists in the MD regions. In Supplementary Material, we show that even when we use naturalistic language materials (from the Brown corpus; Kucera and Francis, 1967) that are representative of the kind of input that our language comprehension system receives, we find similar non-overlap between the activations for the language localizer contrast and demanding cognitive tasks.

Finally, the third objection is as follows: given that (a) we typically use a memory probe task in our language localizer (where after each sentence or sequence of pseudowords participants have to decide whether a probe word/pseudoword appeared in the preceding stimulus; Fedorenko et al., 2010), and (b) the memory probe task is more difficult in the control condition (pseudoword lists) than in the sentences condition, we may be biasing ourselves against finding overlap with demanding tasks because, by design, we are excluding regions that respond to general cognitive effort. To investigate this possibility, we compared responses to a demanding task (spatial working memory) in fROIs defined by two different versions of the language localizer (with vs. without the memory probe task; see Supplementary Material). Across both versions of the localizer, we found little or no response to the conditions of the spatial working memory task in the language fROIs, similar to what we originally reported in Fedorenko et al. (2011). This result is not surprising given the similarity in the topographies of activations for different versions of the language localizer (see Figure 1).

In summary, regions of the language system are spatially and functionally distinct from the domain-general MD system. In contrast to the language regions, the MD regions respond at least as much, or more, during the processing of unconnected meaningless elements (pseudowords) as during the processing of sentences, including naturally occurring ones. The most important implication of the spatial segregation between the language system and the MD system is that we need to distinguish between the two in characterizing their roles in language comprehension/production, because the computations they perform are likely to be different given their different response profiles. This is especially important in the left frontal lobe, where subsets of each system reside side-by-side within Broca's area (Fedorenko et al., 2012).

Before proceeding to the next section, two conceptual issues that sometimes get conflated in the literature are important to clarify. First, functionally specialized circuits (e.g., brain regions that selectively respond to linguistic input) need not be encapsulated (see e.g., Coltheart, 1999; Barrett and Kurzban, 2006 for discussion). Our brain is highly interconnected, although some brain regions have been argued to be more globally connected than others, serving as “hubs” (e.g., Achard et al., 2006; Sporns et al., 2007; Hagmann et al., 2008; Heuvel et al., 2008; Buckner et al., 2009; Cole et al., 2010). Given this interconnectivity, the notion of encapsulation is a priori not plausible as applied to language-responsive or any other brain regions. Moreover, apart from perhaps quite obvious interactions between high-level language processing regions and sensory (visual and auditory) regions as well as motor regions that support articulation or control eye-movements during reading, abundant evidence shows that the language system interacts with many higher-level cognitive systems, including the visual system (e.g., Myachykov and Posner, 2005; Ferreira and Tanenhaus, 2007), the system that supports social cognition (e.g., Brennan et al., 2010; Fitch et al., 2010), and the domain-general working memory/cognitive control mechanisms (as will be discussed in section Narrowing Down the Hypothesis Space for the Relationship between Language Processing Mechanisms and Cognitive Control Mechanisms). So, although lack of encapsulation of the language system has sometimes been offered as an argument against domain-specificity of language (e.g., Blumstein and Amso, 2013), this argument does not hold because functional specialization is perfectly compatible with interactions between specialized mechanisms and the rest of the mind and brain (see also Fedorenko and Thompson-Schill, 2014).

And second, specialized circuits need not be innate (see e.g., Karmiloff-Smith, 1992; Elman et al., 1996 for discussion). Functional specialization can develop via extensive experience with particular stimuli. One notable example is the visual word-form area, vWFA, a visual region that responds selectively to letters in one's native script (e.g., Baker et al., 2007). Recent work with non-human primates also suggests that specialized circuits can develop via an experiential route (e.g., Srihasam et al., 2011). Given that language is one of the most frequent and salient stimuli in our environment from birth (or even before) and throughout our lifetimes, it is computationally efficient to develop machinery that is specialized for processing linguistic stimuli.

Given that most evidence for the engagement of cognitive control mechanisms in language comprehension comes from cases where language processing is effortful, let us consider how often comprehension difficulties arise in naturalistic linguistic exchanges. For example, how frequently do we encounter ambiguous words whose meaning is not fully constrained by the preceding context? What about non-local dependencies between words? Or cases where we have to rely on a single cue (e.g., word order) to interpret the propositional content of an utterance? Although it is difficult to quantify the proportion of such phenomena in typical linguistic exchanges, corpus analyses suggest that linguistic phenomena that many studies in the field of sentence processing have focused on may not be very common. For example, Piantadosi et al. (2012) demonstrated that ambiguous words are typically used in contexts that strongly favor the relevant meaning. Collins (1996; also Temperley, 2007) has shown that most linguistic dependencies are between adjacent elements (see also Frank et al., 2012). And Roland et al. (2007) showed that object-extracted relative clauses with two full animate noun phrases (e.g., “The senator that the reporter attacked was tall”)—perhaps the most frequently investigated construction in the study of syntactic processing—rarely occur.

Indeed, typical linguistic input abounds with cues to meaning, including lexical information, syntactic information, plausibility/world knowledge information, linguistic and non-linguistic (e.g., visual, social) context, and prosodic/punctuation cues. During the last 30 years, research in the field of sentence processing has established that comprehenders rationally use all the information sources available in the input to derive an interpretation of an input string (e.g., Trueswell and Tanenhaus, 1994; Gibson and Pearlmutter, 1998). It is therefore possible that the focus on controlled manipulations that alter the statistics of the human language has led us to overestimate the importance of domain-general working memory and cognitive control mechanisms in understanding language.

However, comprehension difficulty is of course not categorical. Instead, it varies continuously as we perceive linguistic input and is determined by some combination of (i) how expected the input is from the preceding context, and (ii) how much memory is required for integrating the incoming element into the evolving structure/meaning representation (e.g., Demberg and Keller, 2008; Levy, 2008; Gibson et al., 2014; Levy et al., 2014). The question then becomes: What does it take for the domain-general mechanisms to “kick in” during language understanding? In particular, is the MD system (or some subset of it) always active when we perceive language? Does this system get engaged only when we pay attention to the linguistic input, i.e., when the information in the linguistic signal is somehow relevant to us, or maybe only when some threshold of comprehension difficulty has been reached? Or perhaps we turn to the domain-general mechanisms only in rare cases, as a last resort, when the language system “gives up” on interpretation? Now that we have ample evidence that domain-general mechanisms do sometimes get engaged during language comprehension tasks, we can perhaps focus on understanding the precise conditions under which these mechanisms are recruited, to narrow down these various possibilities.

What does the engagement of cognitive control brain regions reflect? Is this engagement functionally important, so that without a properly functioning MD system (or some subset thereof) language interpretation would be severely hampered or impossible? Or is the activation of the MD system simply an “echo” of the effort experienced by the “core” language interpretation system?

To tackle these questions, we need methods that would allow us to examine the effects on language comprehension of inaccessibility of domain-general cognitive control mechanisms. As a result, most evidence from brain imaging investigations does not directly inform these questions. This is also true of much of behavioral evidence, although super-additive processing difficulty observed in dual-task paradigms discussed above does afford some degree of causal interpretation and suggests that taxing domain-general working memory resources can interfere with the processing of (at least syntactically complex) sentences (e.g., Gordon et al., 2002; Fedorenko et al., 2006, 2007). Most direct evidence, however, comes from investigations of individuals with impaired cognitive control abilities. Below I review some of this evidence in light of two alternative positions: cognitive control mechanisms are vs. are not necessary for language comprehension.

The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.