Understanding spoken or written language in real time is essential to our daily life. The ubiquitous presence of long distance linguistic dependencies (e.g., subject—verb dependencies across a relative clause as, for example, in “The director who embarrassed the actor apologized”) indicates that some type of memory representation is needed for successful integration of the dependent items. Although, there has been a long history of investigation into the role of working memory (WM) in sentence comprehension, controversy remains regarding the kind of memory system that is critical for online sentence parsing.

Most studies of the role of WM in sentence processing have focused on capacity demands involved in maintaining constituents prior to integration or maintaining predictions of upcoming syntactic structure (Daneman and Carpenter, 1980; King and Just, 1991; Just and Carpenter, 1992; Gibson, 1998, 2000; Gordon et al., 2001, 2002, 2004; Warren and Gibson, 2002; Fedorenko et al., 2006, 2007; Daneman and Hannon, 2007). For example, capacity-based accounts attribute the standard advantage in speed and accuracy for processing subject relative clauses (SRCs, as in 1a) compared to object relative clauses (ORCs, as in 1b) to increased WM demands imposed by ORC constructions (Gibson, 1998, 2000; Warren and Gibson, 2002).

Sentences like (1a) could be interpreted in a largely incremental fashion. The attachment of the gap in subject position of the RC to “reporter” happens immediately as the verb “attacked” is parsed. However, according to the WM capacity account, comprehenders have to hold “reporter” in sentence (1b) across some new discourse referents (e.g., the object referent “senator”) before attaching it to “attacked” as an object. Thus, the activation of “reporter” decays more in the ORC structure due to more discourse referents being processed and fewer resources being available for maintaining syntactic representations, as there is assumed to be a trade-off between processing and maintenance in most capacity models.

A more recent body of work has focused on interference as an explanation for these effects. These studies emphasize the content of memory representations, rather than the quantity of information that can be actively maintained in memory (Gordon et al., 2001, 2004, 2006; Van Dyke and Lewis, 2003; Lewis et al., 2006; Van Dyke and McElree, 2006, 2011; Van Dyke, 2007). For instance, Gordon et al. (2001, 2004) demonstrated that the standard disadvantage for ORCs compared to SRCs was substantially reduced when the two noun phrases (NPs) prior to the verb had dissimilar referential properties—that is, when the embedded clause NP (e.g., “the senator” in 1a and 1b) was replaced by a pronoun (e.g., “you”) or proper name (e.g., “Joe”). They concluded that their results favored a similarity-based interference account in which memory retrieval was hampered by similarity in the referential properties of the constituent NPs. These findings were not compatible with a pure memory capacity-based account, since the WM loads were the same across noun distinctiveness conditions. Importantly, however, Van Dyke and Lewis (2003) and Van Dyke (2007) have demonstrated that the same nouns may cause more or less interference in sentence comprehension depending on their syntactic roles in the sentence. For example, for sentences (2a and 2b) below, longer processing times at the main verb “was complaining” and greater comprehension errors were observed for sentences like (2b), where the noun phrase (i.e., “warehouse”) was a syntactic subject than when it was the object of a prepositional phrase (as in 2a), although the distances between “was complaining” and its subject “resident” were the same.

Van Dyke and colleagues (Van Dyke and Lewis, 2003; Van Dyke, 2007) attributed this effect to the fact that comprehenders need to retrieve the main clause subject in order to integrate it with the main verb (“was complaining”). The fact that “warehouse” is a syntactic subject in (2b) causes more interference in locating the appropriate subject (“worker”) than when “warehouse” is a prepositional object as in (2a). Relatedly, greater difficulty is observed when the intervening noun phrase has semantic properties that make it more plausible as the subject of the main clause verb. For example, longer reading times and more errors to comprehension questions are observed for sentences like (2c) than (2b), due to “neighbor” being a more plausible subject of “was complaining” than “warehouse.” Thus, these studies demonstrated both syntactic and semantic interference effects during sentence processing.

(2c) The worker was surprised that the resident who said that the neighbor was dangerous was complaining about the investigation.

To explain these and other interference effects, Lewis et al. have advocated a cue-based parsing approach to sentence comprehension (McElree, 2000; McElree et al., 2003; Van Dyke and Lewis, 2003; Lewis et al., 2006; Van Dyke and McElree, 2006, 2011; Van Dyke, 2007). According to this approach, sentence parsing is accomplished through a series of efficient cue-based memory retrievals. The retrieval cues are a subset of the features of the item to be retrieved, and they are derived from the incoming word, context, and grammatical knowledge. Using evidence from both empirical studies and computational modeling, these researchers have suggested that sentence processing is constrained by the degree of interference from non-target constituents, instead of storage demands. Importantly, Lewis and colleagues found that the degree of interference as predicted by the cue-based parsing theory could successfully account for effects observed in previous studies that had been attributed to storage demands, such as longer reading times (RTs) in ORC vs. SRC structures (Van Dyke and Lewis, 2003; Lewis and Vasishth, 2005; Lewis et al., 2006). In addition to the assumption of an extra retrieval in linking the object gap to the relative pronoun in ORCs, this approach suggests difficulty for ORCs due to interference from “reporter” while binding “senator” as the subject of the embedded verb in sentence 1. That is, in the ORC, when processing “attacked,” comprehenders need to retrieve a syntactic subject that is semantically plausible as its agent (Van Dyke, 2007). “Reporter” fits both of these requirements. In contrast, in the SRC, there is no other preceding noun phrase to provide interference.

As the cue-based retrieval approach focuses on interference as the source of difficulty in sentence processing, there is less need to resort to memory capacity as an explanatory factor per se. The move away from capacity accounts is consistent with recent embedded processes accounts of WM, which define WM as the activated portion of long-term memory, together with a very small number of these activated items (e.g., from one to four) within the focus of attention (Lewis, 1996; Cowan, 1999, 2000; Oberauer, 2002; McElree, 2006; Öztekin et al., 2010). In the embedded processes view, information outside the focus is accessed via cue-based retrieval. Consonant with this WM approach, the cue-based parsing model of sentence processing assumes that skilled sentence processing depends on the maintenance of as few as 1–2 chunks of information in WM. Information outside the focus is retrieved through a match of retrieval cues to stored representations. These hypotheses have been supported through a computational implementation (Lewis and Vasishth, 2005; Lewis et al., 2006), and behavioral studies, using a range of methodologies, including self-paced reading, eye-tracking, and speed-accuracy tradeoff techniques (McElree and Dosher, 1989; McElree, 2000, 2006; Van Dyke and Lewis, 2003; Van Dyke and McElree, 2006, 2011; Van Dyke, 2007; Van Dyke and Johns, 2012). The claims from the cue-based parsing approach challenge research pointing to individual differences in WM capacity as a source of sentence processing difficulty (Daneman and Carpenter, 1980; Fedorenko et al., 2006, 2007; Daneman and Hannon, 2007; Unsworth and Engle, 2007). If skilled sentence processing requires at most 1–2 chunks of active memory, this renders moot the claim that poor sentence comprehension is due to low WM capacity.

However, despite the substantial evidence in favor of such a highly limited WM capacity, some criticisms have been raised (Cowan, 2011; Caplan and Waters, 2013). Even among memory researchers who endorse an embedded processes account of WM, there is an ongoing debate about the storage limits of the focus of attention. For example, while McElree and colleagues (e.g., McElree and Dosher, 1989; Öztekin et al., 2010) have estimated the capacity of the focus of attention to be only one chunk of information based on behavioral and neuroimaging results, Cowan (2011) has critiqued these claims and argued for a multi-item focus of attention with a limit of about 3 or 4 chunks. With a larger capacity for the focus of attention, the possibility remains of meaningful individual differences in this capacity. In fact, Unsworth et al. (2014) and Shipstead et al. (2014) have argued, based on a confirmatory factor analytic approach, that a contributing factor to variation in working memory capacity is the capacity of the focus of attention. Therefore, it remains important to examine whether measures of working memory capacity predict comprehension performance, once other relevant factors have been controlled for.

Historically, a substantial body of research has shown a relation between WM capacity measures and the ability to process complex sentences (Daneman and Carpenter, 1980; King and Just, 1991; Just and Carpenter, 1992; Gibson, 1998; Fedorenko et al., 2006, 2007; Daneman and Hannon, 2007; see Long et al., 2006, for a review). Most commonly, these studies have used complex span measures, such as reading span and operation span to index WM capacity (Daneman and Carpenter, 1980; Turner and Engle, 1989). These measures involve both processing and storage components, in that individuals carry out some processing task (e.g., sentence verification in the reading span task or arithmetic computations in the operation span task) while simultaneously maintaining a secondary verbal load (e.g., words or letters). The claim has thus been that these measures reflect a single capacity that can be flexibly allocated to either processing or storage (Just and Carpenter, 1992). In the sentence comprehension domain, storage could involve maintenance of, for instance, lexical items or conceptual representations, and processing could involve, for instance, accessing these representations or assigning thematic roles.

The importance of using WM measures that combine both processing and storage has been emphasized, as other indices of WM capacity such as standard digit or word span, which mainly reflect storage of phonological representations (Baddeley et al., 1998) have typically shown little relation with the ability to process syntactically complex sentences (Waters et al., 1991; Martin and Romani, 1994; Daneman and Merikle, 1996; Caplan and Waters, 1999; Hanten and Martin, 2000; Friedmann and Gvion, 2003; Daneman and Hannon, 2007; Caplan et al., 2013; Kush et al., 2015). Some recent studies, however, have provided some support for a role for phonological storage in complex sentence comprehension (Acheson and MacDonald, 2011; Pettigrew and Hillis, 2014), suggesting that the issue may warrant further attention.

Even though complex span measures have more consistently shown a relation to sentence processing, the source of the WM-language relationship remains unclear, as WM capacity might relate to various aspects of comprehension. For example, Caplan and Waters (1999) showed that many of the findings relating complex span measures to online comprehension ability failed to replicate or did not support the conclusions that had been drawn. In their own work, they found that neither simple span (e.g., digit span) nor complex span measures related to online measures of syntactic processing ability—that is, measures related to the processing of each word as the sentence unfolds. However, these measures did relate to offline processing ability, which involves using the products of comprehension to carry out some task such as matching a sentence to a picture. The implication is that the WM tapped by span tasks is involved in reviewing or checking the results of comprehension rather than the initial interpretation of a sentence. Thus, Caplan and Waters (1999) concluded that for online sentence processing, a WM system specialized for language interpretation is involved. More recent work by Caplan and colleagues (Caplan et al., 2013; Evans et al., 2015) has supported these conclusions.

In contrast to the inconsistent results for complex span measures and simple span measures tapping phonological storage, one measure which has been consistently related to argument integration during sentence processing is the category probe task, an index of semantic short-term memory (STM) in which participants are presented with a word list and asked to judge whether a probe word is in the same semantic category as any list word (Martin and Romani, 1994; Martin et al., 1994; Hanten and Martin, 2000; Martin and He, 2004; Martin, 2005; Harris et al., 2013). Martin et al. (1994); Martin and He (2004) reported a double dissociation between aphasic patients with semantic STM deficits and patients with phonological STM deficits, with the two types of STM deficits having different consequences for sentence comprehension. Aphasic patients with impaired semantic STM but relatively spared phonological retention had difficulty in understanding sentences in which the integration of semantic information of words was delayed rather than immediate. For example, when detecting the anomaly in sentences in which one to three nouns appeared before a verb (e.g., “Rugs cracked during the move”; “Rugs, vases, and mirrors cracked during the move”) relative to sentences in which the nouns followed the verb (e.g., “The movers cracked the rugs”; “The movers cracked the mirrors, vases, and rugs”), performance was equivalent and at a high level when there was only one noun preceding or following the verb, but declined substantially with increasing numbers of nouns before the verb, but remained at the same high level with increasing numbers of nouns after the verb. Similar results were obtained for sentences with varying numbers of adjectives before a noun (“The rusty old red swimsuit”…) vs. after a noun (e.g., “The swimsuit was old, red, and rusty…). In contrast, patients with a phonological STM deficit showed a normal pattern of effects of the delayed vs. immediate integration conditions but had difficulty with sentence repetition (Hanten and Martin, 2000; Martin and He, 2004). Interestingly, the patients with semantic STM deficits performed at a high level and showed no effect of distance on a grammaticality judgment task that varied the distance between words signaling a grammatical error in ten different types of sentence structures (e.g., for verb phrase deletion: “The hopeful young contestants didn't win and neither ^*was their rather aggressive competitor” vs. “Susan didn't leave despite many hints from her tired hosts and neither ^*was Mary”). In contrast, one patient who had little deficit in either semantic or phonological STM demonstrated a detrimental effect of distance in this grammaticality judgment task (Martin and Romani, 1994; Martin and He, 2004). Thus, Martin and colleagues put forward a multiple-component model within the language processing domain, with separate capacities for the retention of phonological, semantic, and syntactic information (Martin and Romani, 1994; Martin and Saffran, 1997; Martin and Freedman, 2001; Martin et al., 2003; Martin and He, 2004; Hamilton et al., 2009). According to this model, semantic and syntactic STM capacities, but not phonological STM, are critical for maintaining unintegrated word meanings and syntactic structures during sentence comprehension, respectively.

However, as discussed earlier, the cue-based parsing approach challenges the long-standing assumption that individual differences in WM capacity are a source of the variation in sentence processing ability, and provides an alternative explanation of the prior neuropsychological results. According to the cue-based parsing approach, comprehension difficulty may arise either from variation in the quality of to-be-retrieved representations, or variation in the ability to efficiently use retrieval cues to activate target information and inhibit irrelevant information (Van Dyke et al., 2014). These two accounts point to language experience and executive control ability as playing important roles for determining comprehension ability. Thus, the relation between semantic STM and sentence comprehension might actually reflect underlying deficits in semantic knowledge representations, which resulted in less rich encoding of semantic features during sentence processing. This assumption is partially supported by the finding that even though the patients with semantic STM deficits in Martin et al.'s studies performed well in terms of accuracy on single word semantic tasks, they did show some deficits in latencies on certain timed semantic tasks (Martin and Romani, 1994; Martin and He, 2004). Thus, these patients may have had some mild degree of semantic deficit per se that affected their comprehension. Another possibility is that these patients have a deficit in the mechanism employed for interference resolution. There are findings suggesting that the left inferior frontal region damaged in the patients with semantic STM deficits is crucial for aspects of executive function (Hamilton and Martin, 2005, 2007), i.e., “semantic control” according to Lambon Ralph and colleagues (Jefferies et al., 2007; Whitney et al., 2011) and “selection from competitors” according to Thompson-Schill and colleagues (Novick et al., 2009; Barde et al., 2010), both of which might be involved in interference resolution during sentence comprehension. Thus, to evaluate the multiple capacities hypothesis, it would be important to show that semantic capacity predicts semantic interference resolution, even after taking into account variations in semantic knowledge and executive control.

Thus, far there has been only a single study to investigate individual differences in sensitivity to interference during sentence processing which takes into account variation in language knowledge. Van Dyke et al. (2014) utilized a dual-task paradigm to assess participants' ability to suppress proactive interference from distractors that appeared in a 3-word memory list (e.g., TABLE, SINK, TRUCK) prior to reading the critical sentence. The critical contrast was between the sentences where the verb was manipulated as follows: It was boat that the guy who drank coffee FIXED/SAILED for 2 sunny days. A previous study (Van Dyke and McElree, 2006) with university-level participants demonstrated that when the verb appeared as fixed, participants experienced retrieval interference from the items in the memory list (which are all fixable items). The Van Dyke et al. (2014) study sampled from a broader range of ability levels and administered a comprehensive battery of 24 individual differences measures. After partialling out the variance that each measure shared with a composite measure of IQ (combining the vocabulary and matrix reasoning subtests of the Weschler Abbreviated Scale of Intelligence; Psychological Corp.; Wechsler, 1994/1999), they observed that WM capacity no longer interacted with individuals' sensitivity to interference whereas a receptive vocabulary measure (Peabody Picture Vocabulary Test-Revised; Dunn and Dunn, 1997) did, such that the comprehension for individuals with low vocabulary was more affected by interference. Van Dyke et al. interpreted this result as most consistent with the view that the quality of to-be-retrieved representations (assumed to be reflected in the receptive vocabulary measure) is a critical determinant of sensitivity to interference. In addition, Van Dyke et al. also observed a significant interaction of IQ with sensitivity to interference, which mirrored the effect found with vocabulary: individuals with lower IQ were more affected by interference. This interaction with IQ is difficult to fully interpret in light of the findings suggesting that IQ shares significant variance with WMC and that this shared variance is itself multi-faceted (Engle et al., 1999a,b; Kane and Engle, 2002; Hambrick et al., 2005; Kane et al., 2007; Shipstead et al., 2014; Harrison et al., 2015). Due to the collinearity of fluid intelligence and WM, we included the WAIS vocabulary measure (Wechsler, 1997; WAIS-III, 2002) as a control variable, because this task is generally viewed as a measure of crystallized intelligence, which has less shared variance with WM capacity (Kane et al., 2007). The inclusion of this task provides a means of assessing the role of WM capacity independent of lexical processing ability.

For the executive control hypothesis, several prior studies have supported a role for general executive control (e.g., as measured by the verbal Stroop task) in comprehending garden-path sentences (Novick et al., 2005, 2014; Vuong and Martin, 2014; Hsu and Novick, 2016). Nevertheless, a potential problem with these findings is that the use of garden path constructions, where correct comprehension requires overriding preferred interpretations of words or syntactic structures, may engage resolution processes differently than in unambiguous sentences (i.e., they may be consciously engaged.) Thus, an important contribution of the current study is to provide data on how executive function becomes involved when parsing unambiguous sentences more like those routinely encountered in everyday conversation.

The current study is the first to examine individual differences in resolving syntactic and semantic interference from distractors embedded within a sentence during online processing. Motivated by the studies summarized above, we examined whether interference resolution depends on general WM capacity, STM capacity (phonological or semantic), executive control abilities, and/or aspects of representational quality. Given the ongoing debates about the nature of the WM-sentence comprehension relation, we aimed to test specific hypotheses about the relation between these various tasks and language processing as predicted by different theories. We summarize these predictions in Table 1 with reference to the specific tasks we use to represent each cognitive construct (see Section Methods for task descriptions).

While the above accounts entail a range of predictions, some differences between them are critical for the present study. The general WM account implies that WM measures will be related to both semantic and syntactic interference, potentially both in online and offline measures. Thus, numerous interactions between WM and sentence effects are predicted. In contrast, the multiple capacities approach predicts that only specific relations will be obtained—for instance, between a measure of semantic capacity and semantic interference resolution but not syntactic interference resolution. Thus, fewer interactions are predicted which follow specific patterns. The language-specific WM approach predicts no relations between WM measures and sentence processing measures, at least in online processing. The cue-based parsing approach predicts that language knowledge and executive function should interact with interference effects. Interactions with capacity measures are generally not expected. However, based on the findings of Van Dyke et al. (2014)—assuming they hold for cases where distractors are embedded within the sentence—interactions with the portion of WM capacity variance related to IQ may occur. This could be predicted for semantic interference, which is the only type of interference examined in the Van Dyke et al. study.

Overall, these results provided further evidence demonstrating both syntactic and semantic interference effects during sentence processing (Van Dyke and Lewis, 2003; Van Dyke and McElree, 2006, 2011; Van Dyke, 2007; Glaser et al., 2013; Van Dyke et al., 2014). Participants were slower to read phrases and less accurate and slower to answer comprehension questions with high semantic or syntactic interference.

Of particular interest for the current study are the implications of the interactions between individual differences measures and sensitivity to interference for theories of the relation between WM and sentence processing (see Table 1). Table 8 indicates the significant interactions that were obtained. The discovery of an interaction of syntactic interference with the WM composite in reading times is an important finding, as no study has yet examined individual differences with respect to this type of interference. This finding is inconsistent with both the general WM and language-specific WM approaches. That is, the general WM approach predicted interactions between the composite WM measure and both syntactic and semantic interference during online processing and question answering, whereas the language-specific approach would not have predicted WM to be related to either type of interference in online measures.

With respect to the multiple capacities approach, the results were mixed. Consistent with this approach, the category probe measure of semantic STM interacted significantly with semantic but not syntactic interference in question answering RTs, even when a measure of verbal knowledge (i.e., WAIS vocabulary) was included as a control variable. Thus, semantic capacity per se beyond semantic knowledge related to semantic interference. Moreover, the measure of phonological retention (digit span) did not interact in the predicted direction with any measure. The specific relation of the composite WM measure to syntactic but not semantic interference sensitivity in self-paced reading might also be seen as consistent with the multiple capacities view, to the extent that the complex span measures tap a retention capacity that is more relevant to syntactic than semantic processing. However, the interaction of the complex span measure, but not the semantic STM measure, with semantic interference in accuracy is inconsistent with this view.

The results with respect to the cue-based parsing approach are also mixed. The significant interaction between syntactic interference and the WM composite in the critical and spillover regions, together with the interaction of complex span with semantic interference in question answering, appear to contradict the assertion that WM capacity per se is not involved in sentence processing (e.g., McElree et al., 2003; Van Dyke et al., 2014). The Van Dyke et al. study, which is the only other study to examine individual differences in sensitivity to interference, albeit semantic interference from distractors outside the sentence, did not find an interaction with IQ-partialled complex span tasks, however they did observe an interaction with IQ, which shares significant variance with complex span measures. Thus, it is possible that the effects observed here are tapping the variance shared between WM and IQ. Further research is needed to determine the nature of this variance.

The other primary result in the Van Dyke et al. (2014) study was to emphasize the quality of linguistic representations, as measured by an assessment of receptive vocabulary, as a factor in determining retrieval success. On the basis of this, we hypothesized that verbal knowledge (as assessed by WAIS vocabulary) may be related to either syntactic or semantic interference or both. We found some evidence for such a claim, as we found that vocabulary did interact with semantic interference, though only in questioning answering and with an effect in the opposite direction to that expected (i.e., higher vocabulary subjects showed greater interference). We nevertheless interpret the significant result as supporting the suggestion that qualitative aspects of the to-be-retrieved representations contribute to the size of interference effects. Additionally, a role for another general ability measure related to interference resolution ability (i.e., the Stroop effect) might have been expected on the cue-based parsing approach, but this failed to interact with interference in any dependent measure. This is discussed further below.

An issue for all models is that, as shown in Table 8, the interactions between syntactic interference and capacity measures only appeared during sentence reading, whereas those for semantic interference only appeared during question answering. The failure to find an interaction between semantic interference and any individual differences measure during online processing is somewhat surprising given that a significant 47 ms semantic interference effect was obtained in the spillover region; however, this null interaction with individual differences measures was also reported in an earlier study (Van Dyke et al., 2014). The significant interaction in question answering RT between category probe and semantic interference might be taken to imply that semantic STM or WM capacity only comes into play in offline semantic processing (e.g., in reviewing the sentence interpretation before answering a question). However, it is possible that the question answering effects reflect online processes that were begun earlier, but were not complete until past the end of the sentence. This is plausible given that the integration of semantic information appears to be slower than that for syntactic information (McElree and Griffith, 1995, 1998; Boland and Blodgett, 2001; Friederici, 2002; Hagoort, 2003). The difference in time course could be due to the finite set of grammatical features to be considered vs. the more complex considerations involved in determining semantic fit (e.g., “the play arrives” may be plausible even though “play” is inanimate). In addition, an important point to note for the present paradigm is that the correct resolution of semantic interference depends on using discriminative syntactic cues. That is, further consideration of semantic features will not resolve the semantic interference between two nouns if both are equally plausible as the subject of the verb. Therefore, perhaps the later time course for semantic interaction effects occurs because semantic conflict attracts attention to semantic features, whereas resolution of the conflict involves a shift of attention to syntactic features⁶.

In line with prior studies, our study demonstrated semantic and syntactic interference from unavailable constituents during sentence processing, consistent with the cue-based parsing approach. The novel aspect of the present study was the investigation of the role of individual differences in WM in modulating these interference effects. We found that general WM capacity derived from complex span tasks showed a relation to syntactic interference during online sentence processing, and to semantic interference during question answering. In addition, a measure of semantic STM capacity predicted the size of semantic but not syntactic interference effects in question answering, while phonological capacity did not predict the size of any interference effects. These interactions with WM were observed in both online and offline processing, even when controlling for vocabulary differences.

The pattern of results argues against claims that a specialized WM is involved in sentence parsing that is different from the capacities tapped by standard simple or complex span measures (Caplan and Waters, 1999). We speculate that the relations to general WM capacity reflect the role controlled attention and potentially secondary memory retrieval involved in both complex span measures and in resolving interference during sentence comprehension. In addition, we consider the specific relation between semantic STM and semantic interference as an indication that code-specific retention capacities mediate resolution and that the rate of loss of semantic or syntactic features (if such could be measured) may differ separately across individuals. This latter assumption is consistent with the multiple capacities approach to WM in which there are separable syntactic and semantic capacities, with the rate of loss of featural information replacing the notion of capacity limits.

There are limitations to the current study, which could be addressed by future research. Most of these were mentioned earlier, but will be summarized here. First, the multiple capacities approach assumes a separable syntactic capacity, but no specific measure of that capacity was included here. In future work, it would be important to include a measure of the ability to retain syntactic information per se, which would have to be demonstrated to be separable from semantic retention and general working memory capacity. Second, with respect to the cue-based parsing approach, despite the proposal that WM capacity may be so limited as to be irrelevant for parsing, we did nevertheless observe significant relationships between our measure of general WM capacity and comprehension. In the face of data pointing to a highly limited WM capacity for sentence processing (e.g., McElree et al., 2003; Johns et al., 2015), future research will need to address the question of what these measures represent, if not capacity. We have noted that recent research with WM capacity as measured by complex span tasks has suggested that separate mechanisms of maintenance in primary memory (which seems equivalent to the focus of attention), attentional control, and retrieval from long-term memory (e.g., Shipstead et al., 2014; Unsworth et al., 2014) underlie this construct. Future work will need to sort out which of these is the source of the relation between general WM and sentence processing observed here. Here we suggested that attentional control may underlie our findings, but did not measure this capacity directly. This would be important to do. In particular, it will be important to test whether effects of WM capacity would disappear when a measure of the ability to resolve proactive interference (which is distinct from the ability to resolve response/distractor interference as in Stroop; Friedman and Miyake, 2004) is modeled. Finally, given the differences between some of the results obtained here and those reported elsewhere (such as the influence of vocabulary on the direction of interference effects), future work will be needed to understand the extent to which the relationship between WM, vocabulary, and interference resolution depends on the type of task (dual-task vs. standard reading) or the location of distractors (sentence-internal vs. sentence-external).

This study was carried out in accordance with the recommendations of the Rice University Institutional Review Board with written informed consent from all subjects. All subjects gave written informed consent in accordance with the Declaration of Helsinki. The protocol was approved by the Rice University Institutional Review Board.

Conception or design of the work: YT, RM, and JV. Data collection: YT. Data analysis and interpretation: YT, RM, and JV. Drafting the article: YT and RM. Critical revision of the article: YT, RM, and JV. Final approval of the version to be published: YT, RM, and JV.