Aims: Relations in language and mechanistic processes

Our first empirical research goal is to document γ-range activity in top-down structure building in the domain of anaphora enabled by wh-filler re-representations in intermediate and thematic gap positions. Assessing the role of γ in the implementation of basic operations in language interpretation and generation can be achieved by documenting specific, structurally dependent oscillations as basic referential chain objects are created in the implementation of anaphora resolution, as a hierarchical dependency between expressions [6,13,24,27]. A second research aim is to characterize the creation of referential objects by implementational cortical circuitry within the γ range as different types of anaphoric relations are computed in ongoing interpretation. Understanding γ-range processes more precisely can also be achieved by establishing the neurocognitive mechanisms enabling syntax-level referential objects for Comps and discourse-level referential objects for Mods. A third empirical goal is to document γ oscillatory effects across languages and constructions—with possible variations as different languages and constructions offer different orders of grammatical information in processing—to fully establish the nature of these effects.

First, the Mod vs. Comp difference constitutes a basic grammatical distinction with syn-sem effects. In Comps, such as message about Frank, the preposition about is lexically required as a feature of the noun message (see message *of/√about). Semantically, the structure message about constitutes a function linking entities to related communications. The noun licensing a complement is viewed as relational. Such structures minimally involve a basic combination (e.g., {message, about Frank}) labelled by a head (e.g., the underlined nominal element in {message, {message, about Frank}}), with an interpretation as function application which blends the two elements [57]. Grammatical extraction is possible out of Comps, as in Who did Frank say that Diana received a message about? even if long-distance dependencies are costly and stranded prepositions are criticized as improper speech by purists. In Mods, such as message regarding Frank, the expression regarding Frank offers additional rather than relational information. Chomsky [57] argued that Mods involve a distinct operation: pair-Merge. Pair-Merge conjoins the two expressions into the ordered pair <message, regarding Frank > , which can be defined as {{message}, {message, regarding Frank}} [61], so that the Mod expression characterizes the relevant communications both as messages and as relating to Frank. As a result of this structure, extraction out of a modifier is ungrammatical: *Who did Frank say that Diana received a message regarding? Hence, if message about Frank and message regarding Frank provide the same overall information, they do so with different grammatical processes: in a lexically encoded relational fashion for Comps and as additional post-lexical information for Mods.

Second, pronouns, as grammatical elements valued contextually in syntax or in discourse, also have top-down constraints on their interpretation. Pronouns like him must be free within a local binding domain that contains them, typically a clause containing a subject, unlike anaphors like himself that must be bound (i.e., syntactically dependent on an antecedent) within this domain [62]. Thus, in John said that Bill saw him/himself, the pronoun him cannot refer to Bill, whereas the anaphor himself must refer to Bill. Indeed, Bill resides within the local binding domain for him/himself, the embedded clause. The pronoun him—but not the anaphor himself—can, however, refer to John in discourse or be bound to John in syntax, since John resides outside the embedded-clause binding domain. A nominal can constitute a binding domain, but only if it has a nominal subject. Hence, in John repeated Bill’s story about himself, the anaphor himself must be syntactically dependent on the nominal subject Bill. In contrast, a pronoun as in John repeated Bill’s story about him is valued outside the nominal. John repeated a story about him is also possible with John as binder for the pronoun him. However, in this case, the story must originate with someone other than John. Chomsky [62] argued that in such cases, a silent nominal subject allows a nominal binding domain of last resort, when a pronoun inside a Comp has an antecedent within the clause. End-of-sentence processing evidence also aligns with this claim [63].

Therefore, the basic Mod vs. Comp difference interacts with referential processes in long-distance dependencies in strings like which message regarding/about him did Frank say that Diana received unexpectedly, allowing the processing to be observed experimentally [27,64]. Long-distance wh-dependencies have been argued to involve silent wh-filler re-representations, described as silent copies (or as movement traces, indicated as t) in recursive steps through the embedded-clause edge into the thematic position as in (1a, b) [62,65,66].

1. (1a) [[which <message, regarding him>] [did [Frank say [t [that Diana received t unexpectedly]]]]]
2. (1b) [[which {message, about him}] [did [Frank say [t [that Diana received t unexpectedly]]]]]

Psycholinguistic evidence (priming [35]; pupillometry [34]; self-paced reading [36–38]) has been found to be consistent with the computation of wh-filler re-representations in intermediate clause-edge and thematic gap positions, with clause-edge re-representations facilitating thematic integration relative to constructions without such re-representation. The Mod vs. Comp distinction in wh-dependencies influences referential relations, with Mods as conjoined expressions residing in a separate dimension from the computation based on lexical selection [57]. Mods may, therefore, be interpreted only once during the dependency, whereas lexically required Comps are obligatorily linked to every step of computations implicating the head noun (see [56–60], for extensive discussion). For instance, the string which message regarding Frank did the idiot say that Diana received unexpectedly allows the expression the idiot to anaphorically refer to Frank. However, in the string which message about Frank did the idiot say that Diana received unexpectedly, the expression the idiot must refer to someone other than Frank, similarly to the fact that the idiot cannot refer to Frank in The idiot said that Diana received a message about Frank.

Likewise for pronouns, even when the information conveyed is similar across Mod and Comp structures, as in the string which message regarding/about him did Frank say that Diana received unexpectedly, distinct computations are expected, given that Mods, unlike Comps, need only be interpreted once. In regarding him, the pronoun can be assigned a discourse referent in the sentence-initial position, with the referent later identified as Frank. In about him, however, the evaluation of him requires an embedded-clause binding domain with the wh-filler re-represented at the clause edge (and at the thematic gap inside it). This enables a structural referential binding relation in which the grammatical elements for him and Frank form a referential chain object. As noted earlier, an embedded-clause binding domain also allows for a discourse-level coreference relation for the pronoun in about him. Thus, as a wh-filler is re-represented as in the string which message about him did Frank say that Diana received unexpectedly, an embedded-clause structural configuration in slow rhythms [3,13] induces a referential chain object computation involving the antecedent Frank and pronoun him implemented in the cortical γ-implemented operational workspace. Hence, in a processing loop involved in the implementation of relations in syntax as well as in discourse for “hierarchically organized elements” [24, p. 723] the wh-movement configuration with a Comp structure licenses the creation of a referential chain object for the elements Frank and him, implemented by γ oscillations at specific computational moments.

Using French, which grammatically marks verbal Mods (le concernant, ‘concerning him’) vs. prepositional Comps (à propos de lui ‘about him’), Dekydtspotter et al. [27] argued that such structures provide a way to examine a γ-implemented syn-sem operational workspace functioning across first and second language backgrounds. Hence, wh-fillers qualified with Mods vs. Comps predict specific ERPD patterns reflecting syntax-based anaphora for Comps vs. discourse-based anaphora for Mods at different computational moments. In addition, since β rhythm activity tracks the maintenance of sentential information enabling prediction, Dekydtspotter et al. [64] also reported β-range ERPDs in induced power that signaled differences in processing load for Mod (le concernant, ‘regarding him’) vs. Comp (à propos de lui, ‘about him’) structures in anaphora in French, in both first- and second-language speakers. β-range sentential processing load effects were associated with the extraction of bridge verb information and confirmation at the subordinator of a tensed embedded clause, with evidence of greater processing load among second-language speakers. These effects were timed with clause-edge computations for syntactic binding vs. coreference in anaphora resolution. They revealed greater power for lexically encoded Comps (vs. Mods) in matrix-clause antecedent match in anaphora resolution, and greater power for Mods (vs. Comps) in matrix-clause antecedent mismatch as anaphora resolution is delayed. These effects are consistent with binding vs. coreference processes of anaphora resolution for Comps and Mods and with the need to maintain an unidentified discourse referent across the clause edge in coreference vs. binding as anaphora resolution is delayed.

Addressing the role of the γ rhythm in the computation of such dependencies at the clause edge, Dekydtspotter et al. [27] examined γ-range ERPDs in evoked power at the clause edge as cell assemblies formed in referential processing. γ power differences matching those asymmetries reported in β occurred at the subordinator in low γ, followed by broadband-γ effects, consistent with the retrieval of information in low γ and integration in high/broadband γ. Ultimately, however, establishing γ oscillations as a cortical operational workspace executing syntactic vs. discourse referential relations in anaphora requires finding effects across languages and constructions. In the current study, we examine Mod vs. Comp differences in English, using bi-clausal interrogative structures with two important differences from the French structures tested by Dekydtspotter et al. [27]: First, the morphosyntax of the pronoun remains the same across Mod regarding him and Comp about him, unlike in French. Second, lexico-grammatical elements provide a different timing of information as did marks interrogative and past-tense features and unambiguously requires a main verb, which limits the need for reanalysis, relative to French, in which past tense auxiliary avoir ‘have’ in a dit ‘said’ can also have a present-tense main verb alternative value.

A fuller integration of Minimalist approaches to linguistics with neuroscience minimally assumes that features, basic elements, operations, as well as phrasal representations are implemented by simple neuronal assemblies [66] and that a neurolinguistic model for language must account for parsing [6]. The biocomputational basis of Minimalist approaches is well established in parsing research [50,67]. In left-corner parsing, sentence processing involves bottom-up and top-down processes implemented in any order consistent with incremental processing [48,54]. In incremental parsing of a wh-filler-gap dependency, a wh-determiner phrase must be structured within a D(eterminer) > n(ominal)> root √N(oun) concept category hierarchy for nominal expressions (Table 1) and must be integrated within a C(omplemetizer) > T(ense) > v(erbal)> root √V(erb) concept hierarchy for clauses (Table 2). Crucially, a binary γ-implemented cortical operational workspace for basic operations (e.g., Merge, together with labeling and feature valuation) interacting with a general resource workspace enables left-corner parsing options [68], as shown in Tables 1 and 2. In incremental processing of a wh-filler within the γ-implemented binary cortical operational workspace, neurocognitive processes implementing basic operations involve the coactivation of cortical sites across distinct levels of representations in information processing. Hence, for instance, a wh-determiner element for which (i.e., D: which) can be combined with a precomputed nominal-category-level combination (i.e., n: Number + √N[0])—including a bare nominal-concept category √N, with an empty feature space [0] as a placeholder—into a set {D: which, n: Number + √N[0]} and labeled into a new partially specified combination {D: which, {D: which, n: Number + √N[0]}} for externalization and interpretation [69]. In further processing, a second phase of computations can then enable the identification of root N-level contents with the conceptual root features for message added in valuation.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Parsing of a wh-DP filler item within the D(eterminer) > n(ominal)> root √N(oun) category hierarchy in recursive deployment of γ operational workspace.

https://doi.org/10.1371/journal.pone.0333820.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Parsing of wh-DP dependencies within the C(omplemetizer) > T(ense) > v(erbal)> root √Verb hierarchy for clauses in the γ operational workspace.

https://doi.org/10.1371/journal.pone.0333820.t002

Computational efficiency is expected to guide parsing. The derivation in Table 1 starts by merging an n: Number category element, containing an unspecified Number value, with an empty root nominal-concept placeholder √N[0] for the bare category (column 2). The D: which element is then merged as the stored combination is retrieved into the workspace and is normalized to a basic element (column 3). The next computational step involves the computation of the root N-level representation in additive valuation (column 4). An alternative derivation in top-down D > n > √N processing would start by merging the D-category with a noun-category placeholder n[0]. Its valuation requires merging a n: Number element with a nominal-concept √N[0] placeholder. Its valuation then requires additional cycles of operations as the root-level information is added. Under Minimal Yield, “the fewest possible new items accessible to further operations” should be constructed [70, p. 19]. The derivation in Table 1 diminishes the number of combinations by selecting as much information as possible.

The wh-filler element encoded in item memory can be recursively reinstated to the γ-implemented cortical operational workspace, to be merged with various sentential category objects in additional phases of computation as shown in Table 2 columns 2–4. Hence, in classical feature-based approaches to wh-movement, wh-fillers and gaps reflect strong wh*-features triggering movement to the interrogative clause-edge category. Additionally, the extended projection principle (EPP) requires strong uninterpretable category features [uD*] for movement between argument positions [71]. Therefore, in the top-down parsing of wh-filler-gap dependencies, the wh-filler phrase (i.e., wh-DP) is re-loaded into the γ-implemented cortical operational workspace together with a normalized interrogative category object C[Q, wh*]+T[0] to be merged in set formation and receive a label in a bare phrase structure design (see Table 2, column 2). This clause-edge category object enables the wh-filler to restrict the question operator (Q) [72–74]. It also contains a T-category placeholder T[0] to be valued in the structural integration of the next phase of computations with the hierarchy. The next phase is, therefore, triggered by the need for T-category content. Under the EPP, a tense-marked category requires that a determiner phrase be merged with a tense category object, T[uD*]+v[0]. In incremental processing, the wh-DP element is therefore reloaded from item working memory into the binary operational workspace together with a normalized tense category object in further Merge {wh-DP, T[uD*]+v[0]} inducing the active filler strategy [75,76]. Upon labeling, the new T-category phrasal object {T, {wh-DP, T[uD*]+v[0]}} normalized to a basic element can provide the value for the tense category placeholder T[0]. A verbal category placeholder v[0] to be identified allows for next updates. During the next phases of computations, the v category for verbal features with the verbal root concept information allows for the thematic integration of the wh-filler: { √Verb, {wh-DP, √Verb[uD]}}, completing the process (column 4).

Table 2 illustrates a classical feature-based account of movement that echoes complementizer agreement in wh-movement as in Irish [77], together with the expression of clause-edge copies [30–32]. It reflects an automatic Merge-Label system in coactivation between sites across levels of representations. However, since transfer to the interfaces requires a label, movement dependencies have been argued to reflect the timing of the labeling process [78,79]. A labeling-based wh-dependency computed in phases [_CP wh-DP... {wh-DP, C + I[0]}] in left-corner parsing would reflect a Minimal Search-induced labeling process within the loop architecture [50,80].

Basic processes can also allow the establishment of referential chain relations in anaphora resolution during syntactic computations [27]. The position that anaphoric relations involve Merge-like processes is well established in Minimalist grammar. Kayne [81] proposed that referential relations as in John said that Mary saw him implicate a ‘doubling constituent’ [John him] together with the movement of John as in [John said that Mary saw [<John> him]] leaving a silent copy <John > . Fong and Ginsburg [52] implemented Kayne’s approach in proposing a Minimalist computational parsing mechanism. However, cases such as The dog John petted bit him would require a complex parallel/sideways movement [82], which has been criticized as biologically problematic [83]. Furthermore, such movement is ruled out by a dual-workspace design [84]. The ‘doubling constituent’ [John him] approach does not match constraints on movement computations together with the involvement of pragmatics in coreference relations but could work for anaphors like himself [85]. A dual-workspace design of the type considered in this paper between general working memory for vocabulary items and for structures and a cortical γ operational workspace for operations can capture aspects of Kayne’s intuitions of a Merge-like operation in referential relations for pronouns within a more traditional framework. These referential relations involve similar processes in the creation of referential chain objects: first combining the syn-sem elements Frank and him as in {Frank, him} and mapping this combination of referential grammatical elements to a single discourse referent x for an entity named Frank in the mental model. This mapping creates a referential chain object {x, {Frank, him}} for storage in working memory and for integration of an anaphoric relation into the ongoing sentential representation. Again, pronouns can always receive a discourse referent. However, a grammatical binding process in which multiple elements receive a discourse referent in one fell swoop eliminates the need for multiple discourse referents in anaphora, as argued by Reuland [86] and experimentally tested by Koornneef and Reuland [87].

In a string such as which message regarding her did Frank say that Diana received unexpectedly, the Mod regarding her need only be part of the interpretive cycle for the wh-filler noun phrase once. A variable discourse referent must be assigned early on to the pronoun (i.e., y: her), to be later identified with a referent for an antecedent (i.e., x: Diana) in discourse anaphora resolution. In cortical processing, the identification (x = y) of the discourse referents for Diana and her in coreference can be achieved by mapping a combination of discourse referents {x: Diana, y: her} to an entity (d) in a cognitive model for a possible world with someone named Diana [55]. This mapping creates a discourse representation object {d, {x: Diana, y: her}} for integration into the ongoing discourse representation structure. In the string which message about her did Frank say that Diana received unexpectedly with a Comp structure about her reliant on an embedded-clause domain, a nominal syntactic category antecedent placeholder can be anticipated in advance of encountering the expression Diana in the anticipation of a referential chain object for binding: {n: Animate: Gender: Feminine, her}. The creation of syntactic and discourse referential chain objects, to be stored in memory and integrated into the ongoing development of grammatical and discourse structures, predicts certain patterns of effects in implementational circuitry as a wh-dependency allows grammatical or discourse referential elements to be returned in low γ to the γ-implemented operational workspace, and then as syntax-based or discourse-based referential chain objects are created in high γ for integration into the ongoing sentential interpretation.

Crucially, the processes defining basic operations in terms of the nature of the manipulated elements involve information-preserving (Markovian) processing. Hence, processing for syntax seems compatible with computational Bayesian inference approaches to information processing by cortical circuitry. George and Hawkins discussed a generative hierarchical neural Markovian model of information processing in which “higher levels of the hierarchy represent larger amounts of space and longer durations of time” [88, p. 2]. The hierarchical γ-band processing of the type needed for basic language operations might therefore be characterizable in terms of coincident high probabilities and sequences of coincident high probabilities across coactivated processing nodes for syntactic and discourse-semantic elements, combinations, and objects, enabling information conservation in processing. The characterization of coactivation as probability patterns between nodes across the hierarchy does not negate specialized domains and feature-specific circuits, let alone internally driven computations. George and Hawkins [88], therefore, offer a potential account that highlights the fact that symbolic processing within a network may involve activations of random subsets of the circuits encoding features across domains. The elementary processes linked to the creation of linguistic objects of various natures in high γ suggest that implementational circuits for a hierarchical workspace can account for various operations as a reflex of the representations involved.

Hypotheses, research questions, and predictions

We examined critical items like (2a-d), manipulating the contents of the wh-filler, with either a Mod (2a, c) or Comp (2b, d) qualification of the noun. The pronoun inside the wh-filler matched the gender of an antecedent either in the matrix clause (N1 antecedent-pronoun match; 2a, b) or in the embedded clause (N1 antecedent-pronoun mismatch/N2 antecedent-pronoun match; 2c, d).

1. (2a) Which message regarding him did Frank say that Diana received unexpectedly?
2. (2b) Which message about him did Frank say that Diana received unexpectedly?
3. (2c) Which message regarding him did Diana say that Frank received unexpectedly?
4. (2d) Which message about him did Diana say that Frank received unexpectedly?

In top-down re-representation of wh-fillers in the dependencies in (2a-d), the main verb say, as a main verb is expected after the helping verb did, triggers a clause-edge category linking the wh-filler to the embedded clause. The embedded verb provides the eventual object with the thematic role of the wh-filler, indicating that a message under questioning was received. The Mod structure regarding him in (2a, c) constitutes a binding domain, enabling a variable discourse referent to be immediately assigned to him, and identified as Frank in coreference. In matrix-clause anaphora computations for Comps (2b, d), a referential chain object can only be created in syntactic binding as a structural binding domain is computed, upon the integration of the wh-filler with the subordinate clause. With a matching matrix-clause antecedent, Comp-enabled anaphora depends on an embedded-clause binding domain, with anaphora through binding (in addition to coreference) predicting additional costs. With a mismatching matrix-clause antecedent, an unresolved variable discourse referent in Mods (2c) must be actively maintained into the embedded clause, whereas a binding chain with an anticipated structural binder {n: Gender: [-Feminine], him} avoids these costs in Comps (2d).

Table 3 provides a timing overview of anaphoric processes for Mods vs. Comps in N1 antecedent-pronoun match vs. N1 antecedent-pronoun mismatch/N2 antecedent-pronoun match involved in the formation of the referential chain objects constructed in an operational workspace, along the lines considered in Adger [84], all implemented by cortical γ oscillations. Thus, in processing strings such as which message regarding/about him did Frank say that Diana received unexpectedly, the construction of the referential chain object {x, {Frank, him}} in binding is enabled by the Comp about him (2b), as per line 3 in Table 3. The process of constructing the referential chain object {x, {Frank, him}} should induce greater power, as the elements for Frank, him are retrieved from working memory into the γ operational workspace and as a discourse referent is assigned to the combination {Frank, him}. These effects are expected as an embedded-clause domain enabling syntactic binding is computed in filler-gap dependencies. The discourse object {f, {x: Frank, y: him}}, regarding him (2a), as per line 2 in Table 3, is enabled upon processing Frank, and therefore need only be refreshed as the wh-filler is re-represented. A pre-established discourse referential chain object should therefore implicate less power within the specific window for the structural computation. In contrast, in late-matching antecedent strings such as which message regarding/about him did Diana say that Frank received unexpectedly, the processes for regarding him (2c), as per line 4 in Table 3, should induce greater power for the active maintenance of the unidentified discourse referent y for him into the embedded clause as well as for multiple referents mapping to the individual f, for someone named Frank, in the model’s entities. All of this creates the discourse referential chain object {f, {x: Frank, y: him}}. As shown in line 5 in Table 3, for (2d), Comps enable a partial referential chain object for syntactic binding {x: ¬ Feminine(x), {n: Animate: Gender: -Feminine, him}} pre-established into the embedded clause. This avoids mapping a discourse referent for the pronoun. This syntactic referential object can be later updated as {x, {Frank, him}} as a nominal binding domain is computed. As argued by Reuland [86] and by Koornneef and Reuland [87], avoiding discourse referents eliminates referential processing costs at the discourse level. This necessary syn-sem computation of a sentential structure requires fewer resources within the embedded clause than multiple referents in coreference.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Timing of referential chain objects construction in syntactic binding and discourse coreference during the filler-gap dependency enabling a binding domain for the embedded clause.

https://doi.org/10.1371/journal.pone.0333820.t003

Mod- vs. Comp-modulated anaphora-linked ERPD patterns in N1 antecedent-pronoun match vs. N1 antecedent-pronoun mismatch/N2 antecedent-pronoun match, specifically timed as a binding domain is established in embedded-clause wh-dependency processing, should reflect the availability of syntactic binding in Comps vs. Mods. The availability of binding in (2b) vs. discourse coreference in (2a) in N1 antecedent-pronoun match, along with the cost of maintaining an unidentified referent in coreference in N1 antecedent-pronoun mismatch/N2 antecedent-pronoun match (2c vs. 2d), predict specific Mod vs. Comp ERPD patterns [27]. As Coopmans and Nieuwland [29] and Nieuwland and Martin [22] demonstrated, low-γ oscillations can be linked to the retrieval of referential elements from working memory and high-γ oscillations to semantic unification. In terms of the timing of these ERPD effects, the retrieval of referential elements from working memory and their combination should take place as soon as a wh-filler containing a pronoun is linked to the embedded-clause edge in re-representation. Mod- vs. Comp-modulated anaphora-linked ERPD effects in γ should first arise as an embedded-clause binding domain induces the low-γ retrieval of discourse- vs. syntax-level referential elements and their integration into new objects in high-γ computations. Hence, low-γ ERPDs due to retrieval should be followed by high-γ effects as elements are combined and as these combinations receive referential values in anaphoric interpretation enabling the consolidation of a sentential interpretation when the wh-filler is thematically integrated.

Therefore, turning to the role of γ oscillations, RQ1 asks: Will Mod- vs. Comp-modulated anaphora-linked ERPDs arise as the bridge verb requires an intermediate gap position, consistent with the low-γ reinstatement of referential elements to a γ-implemented syn-sem operational workspace? RQ2 asks: Will Mod vs. Comp structure-dependent anaphora-linked ERPDs then arise as the wh-filler is integrated within the thematic domain in high γ as referential chain objects are computed to terminate the interpretation of the structure? In response to these research questions, we expect that, due to syntactic binding allowed by Comps vs. discourse coreference processes in Mods, anaphora-linked Mod vs. Comp ERPDs should arise first in low γ as referential elements are reinstated from working memory and then in high γ as referential chain objects are computed and an anaphoric relation resolved enabling the finalization of the on-going interpretation of the sentential structure upon the thematic integration of the wh-fillers.

Crucially, only in specific top-down language structuring [6,13,28] should Mod- vs. Comp-modulated anaphora-linked γ ERPDs arise without being accompanied by any observable differences in the synchronous processing of phrasal arrays in Mod vs. Comp conditions. Only in specific top-down language structuring will Mod- vs. Comp-modulated anaphora-linked γ ERPDs reflect the timing of silently computed gaps in re-representation. Only in specific top-down language structuring will Mod- vs. Comp-modulated anaphora-linked γ ERPDs reflect the retrieval of referential elements in low γ as well as the computations of referential chain combinations in high γ—everything else being equal in terms of the semantic fitness of Mod vs. Comp structures. This is because in bottom-up approaches to language processing [3,24], in which higher order processes in γ oscillations respond to the tracking of phrasal arrays [18,25], such effects are not expected and would be purely coincidental. Computations that do not involve the re-representation of wh-fillers should also not result in such a timing of effects [89].

Participants and testing procedures

A total of 27 participants with English as their first language were recruited for this study. One participant aborted the study, leaving the total participant count at 26 (14 F, 12 M; 23 RH, 3 LH). We report results from 23 right-handed participants (13 F, 10 M). These participants were undergraduate students with no history of dyslexia. Accuracy rates of 76% on factual comprehension checks show the task to be challenging. On comprehension checks related to anaphoric interpretation, respondents interpreted the pronoun as referring to the gender-matched noun phrase 75% of the time. However, comprehension check accuracy was not used as a filter for analysis: Real-time brain processing as participants compute a bi-clausal filler-gap dependency is expected to be independent of their behavior on comprehension checks following individual sentences.

EEG preprocessing and data extraction

EEG was recorded at a 1000 Hz sampling rate via a 64-electrode EGI system (Electrical Geodesics Inc., Eugene, OR; as displayed in Fig 1) referenced to Cz (vertex) online. The signal was collected using a Net Amps 300 amplifier with a gain of 5000 and acquisition software Netstation (version 4.5.4). Impedances were verified to be below 50 kΩ before each of the four blocks in the task. All preprocessing and data cleaning procedures were performed using the EEGLAB toolbox based on MATLAB (version 9.5) [91]. An 8ms latency shift due to the amplifier was corrected before preprocessing. Line noise was removed using the CleanLine plugin for EEGLAB [92].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. The 64-electrode EGI system.

https://doi.org/10.1371/journal.pone.0333820.g001

The continuous data were then divided into 3.60-second epochs starting with did and running to the end of the sentence. Following segmentation, we visually inspected each epoch for bad channels and, if a channel was bad in more than 10% of epochs, we removed the whole channel. It has been demonstrated that muscle activity can create noise in high-frequency EEG measures, and γ-band results should thus be reported and interpreted with caution [93]. Hipp and Siegel [94] showed that removing such artifacts from the EEG recording through rejecting data sections affected by artifactual signals or Independent Component Analysis (ICA) can allow for more confident analysis of high-frequency EEG. Therefore, we visually inspected each epoch and systematically removed any epoch including unexpected EMG activity (i.e., furrowing of the brow, face and neck movements, but not blinks). We used the runica algorithm for ICA to extract and then manually check the 32 most impactful components generated by principal component analysis (PCA) so that we could effectively remove the remaining ocular and cardiac activity along with any remaining artifacts. An average of 89% of trials were retained across subjects (SD = 3.120), with an average rejection number per subject of 19.01. The average number of trials retained was similar across conditions, (2a), M = 27.00, SD = 3.37; (2b), M = 26.61, SD = 3.60; (2c), M = 25.83, SD = 3.08; (2d), M = 27.26, SD = 2.26. The data were average referenced and missing channels were interpolated for the time-frequency analysis.

Time-frequency analysis

The preprocessed EEG data were loaded into the FieldTrip toolbox [95] as four datasets for the four structural conditions (2a - d). The time-frequency analysis was run in the time window of 3600ms from the onset of the word did to the end of the sentence to achieve stable frequency decomposition with a long baseline before the first analysis at the bridge verb. One might wonder if coreference computations could be entertained at the matrix subject in N1 antecedent-pronoun match conditions (2a, b) and at the embedded subject in N1 antecedent-pronoun mismatch/N2 antecedent-pronoun match conditions (2c, d). Crucially, we examined Mod vs. Comp comparisons in early vs. later antecedent-pronoun match conditions [(2a - 2b) – (2c - 2d)]. Mod- vs. Comp-modulated ERPDs in N1 vs. N2 antecedent-pronoun match are completely dependent on structural binding domain constraints on referential chain object creation. Across binding and coreference referential relations, the interpretation of pronouns in Comps depends on wh-filler integration within an embedded clause defining a default binding domain. In contrast, pronouns inside a clausal Mod structure defining the binding domain allow immediate coreference. This makes highly specific predictions about the timing of the processes that implement the computation of referential chain objects resulting in Mod- vs. Comp-modulated anaphora-linked ERPDs between N1 antecedent-pronoun match and N1 antecedent-pronoun mismatch/N2 antecedent-pronoun match. N1 antecedent-pronoun match vs. N1 antecedent-pronoun mismatch/N2 antecedent-pronoun match across Mod and Comp structures therefore simply could not yield Mod- vs. Comp-modulated ERPDs in early and late anaphoric matches until wh-filler re-representations at gap sites enable the relevant processes. The processing of subject positions alone cannot carry these ERPDs induced by the timing of referential chain objects for Comps in syntactic binding supported by wh-filler computations. Indeed, no Mod- vs. Comp-modulated N1 antecedent-pronoun match vs. N1 antecedent-pronoun mismatch/N2 antecedent-pronoun match ERPD effects arose in the γ range at the matrix-clause and embedded-clause subjects.

Following Dekydtspotter et al. [27], we examined evoked γ power that reflects both time-locked and phase-locked oscillatory responses. We first computed the ERPs of each condition for each subject. Then, we extracted power information by convolving a family of Morlet wavelets of 7 cycles in .5 Hz steps within the selected time window of each EEG trial, which yielded the time-frequency information of the neural activity. The length of the wavelets was set as 3 standard deviations of the Gaussian kernel. At 60 Hz, the wavelet duration was 0.037 seconds. The spectral bandwidth was 17.143 Hz. We log-transformed (10*log10) the derived power in Fieldtrip to standardize the unit as decibels at each of the frequencies between 20 Hz and 100 Hz for each condition of each subject. The statistical analysis time window lasted 500ms starting at 50ms before the presentation of each verb, as we expect information prior to 200ms into a word to reflect predictive processing and information starting around 200ms into a word to reflect computations dependent on lexical information for the specific word [96]. These time windows capture processing moments where the bridge verb enables an embedded-clause dependency through an intermediate gap position, and when the embedded-clause verb enables a thematic gap position enabling the thematic integration of the wh-filler.

Statistical analysis

We calculated Mod vs. Comp power differences between the N1 antecedent-pronoun match [(2a) – (2b)] and N1 antecedent-pronoun mismatch/N2 antecedent-pronoun match [(2c) – (2d)] conditions to address distinct allocations of resources in discourse-based vs. syntax-based anaphoric processes for Mods vs. Comps as in Table 3, reflecting different binding domains for prepositional Comps vs. clausal Mods. We then examined differences between Mod and Comp structures in N1 antecedent-pronoun match vs. N1 antecedent-pronoun mismatch/N2 antecedent-pronoun match. Data were analyzed with cluster-based nonparametric permutation tests to avoid the multiple comparison problem for our medium-density electrodes, on the assumption that the spatially adjacent channels exhibit similar spectral-temporal features [97–99]. We performed cluster-based permutation tests using Monte Carlo simulations with 10,000 random samplings for each channel-frequency-time triplet [99]. A reviewer wondered why we did not use an ANOVA given a 2x2 design and instead used paired-samples t-tests examining differences of differences. We follow the guidance of FieldTrip (https://www.fieldtriptoolbox.org/faq/stats/clusterstats_teststatistic/). As pointed out on the FieldTrip website, for a two-sample setting, t and F distributions are “interchangeable,” since F = t². However, in cluster-based permutations, t and F value sums (cluster mass) for each iteration provide radically different null distributions, so that only the t version of the cluster-based permutation test “maintain[s] the statistical properties of the test (in terms of power).” Cluster-based permutation testing based on F values, therefore, is not recommended. Finally, we note that permutation tests provide the time window and electrodes within which a significant effect arises. They, however, lack precision as to the exact timing and location of effects [100]. Our discussion of the timing of ERPD effects is therefore limited to the window in which the effects are found.

At each critical segment (intermediate non-thematic and thematic gap positions), we conducted paired-samples t-tests within two bins (low γ 30–50 Hz and high γ 50–90 Hz). The two bins reflect frequency bands that have crucially been shown to be engaged in the processing of pronouns, with low γ frequencies first implicated in information retrieval followed by high γ frequencies in integration [22,29]. Non-thematic intermediate clause-edge gaps and thematic gaps reflect distinct theoretically dependent steps in filler-gap dependencies, with an intermediate non-thematic clause-edge gap dependent on an internal Merge approach, not required for effects obtained in thematic integration alone. The different positions therefore reflect different hypotheses about the nature of language computations [89]. The non-thematic intermediate gap position enabling the computation of a binding domain for the embedded clause is predicted to induce low-γ ERPDs in the retrieval of referential elements into an γ-implemented operational workspace inducing high-γ ERPDs as referential chain objects are generated. In contrast, the thematic gap position is predicted to induce high-γ ERPDs in the finalization of anaphora resolution. In view of the specificity of these positions associated with different processes supporting different research questions, motivated by a range of psycholinguistic research findings supporting a parse implicating the embedded clause edge [27,35–38,64,101] together with specific roles for low and high γ, we adopted a conservative Bonferroni correction of α = .025 for two frequency bands as most appropriate. A lower threshold of α = .0125 despite specific predictions for the two segments could mean that strongly suggestive ERPD effects with clear gap-position timing in the retrieval of referential elements in low γ and the computation of referential chain objects within high γ might be artificially dismissed, with their potential theoretical significance wrongly ignored.