Introduction

Language is not a disparate and haphazard collection of idiosyncratic templates for how to formulate sentences. Instead, sentence templates form intricate systems of partial overlaps and alignments, an observation that has fueled inquiry since the Indian scholar Pāṇini wrote the first explicit grammar of a language over 2,500 years ago. For example, sentences like “The gardener planted a tree,” “The gardener crouched,” “The gardener worked hard,” or “A tree was planted by the gardener” align with each other in an abstract way by employing templates that begin with the same structure, here, with a noun phrase (the subject: “the gardener” and “a tree”) followed by a verb (“plants,” “crouches,” “works,” and “was”), while they differ in the remainder. Such alignments differ between languages in striking ways.

Some languages align noun phrases like “the gardener” in a sentence with 1 noun phrase (intransitives, e.g., “the gardener crouched”) with the agent noun phrase in sentences with 2 or more noun phrases (transitives, e.g., “the gardener planted trees,” “the gardener poured water into the trough,” etc.). Other languages do not align the 2 sentence types and instead add a special marker in sentences with more than 1 noun phrase, keeping them formally distinct from sentences with only 1 noun phrase. These languages contrast a plain, unmarked noun phrase in intransitives (“the gardener crouched”) with a marked agent noun phrase in transitives (“the gardener+x planted trees,” with an additional marker represented here by “+x”; cf. Fig 1). This phenomenon, known as ergativity [1,2], is found in about one-third of the world’s languages [3].

Download:

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

Fig 1. Illustration of the contrast between aligned and nonaligned agent expressions in Hindi, which allows both types of expressions (for detailed examples see S1 Table).

In aligned expressions (used with imperfective aspect), unmarked noun phrases are compatible with both intransitive (“crouches”) and transitive sentences (“plants trees”). In nonaligned expressions (used with perfective aspect), unmarked noun phrases are only compatible with intransitive sentences (“crouched”) because in transitive sentences (“planted trees”) noun phrases are marked by an additional element “+x” (“ne” in Hindi).

https://doi.org/10.1371/journal.pbio.3001038.g001

Alignments and other kinds of partial overlaps are recognized as central properties of the human language faculty in all theories of grammar, albeit using widely different representational formats, e.g., derivations or inheritance schemata [4]. Despite their prominence in linguistic theory and the striking nature of variation, however, alignments have received remarkably little attention in neuroscience and psychology. The difference between aligned and nonaligned expressions poses 2 unresolved questions for neural processing and its relation to the variation between languages:

First, are there differences in the neural planning processes for aligned and nonaligned expressions? More concretely, is the way speakers plan sentences shaped by whether initial noun phrases are unmarked (“the gardener…,” aligned expression) or whether there is an opposition between unmarked and marked noun phrases (“the gardener” versus “the gardener+x,” nonaligned expression; Fig 1)? Does this difference affect sentence planning already in its early phases or only later, when speakers encode the phonological form of and articulatory motor plans for words with and without an additional agent noun phrase marker “+x”) [5,6]? Specifically, sentence structures with aligned expressions might allow speakers to delay commitment to the choice between an intransitive and a transitive sentence plan because their beginnings overlap (“the gardener… crouches/plants trees”) [7,8]. At the same time, this intermittent compatibility with multiple sentence plans might also require that 2 possible options need to be kept distinct while speakers construct a syntactic plan [9,10].

Second, if neural information processing is sensitive to how structures are aligned, does the neural processing mirror the observation that aligned structures are structurally simpler (with no additional marking) and more common in the world’s languages [11]? Such a mirroring has been motivated by evidence from event-related potentials in sentence comprehension: Across languages, both with and without nonaligned expressions, comprehenders initially interpret unmarked initial noun phrases as referring to agents, i.e., following an aligned pattern [3,12–16]. A bias for aligned structures is also found in phylogenetic models of linguistic evolution [3]: After controlling for contingencies of history (such as language contact and language shifts), languages are universally more likely to develop and maintain aligned expressions than the opposite. Such a correlation of neural processing and evolutionary developments might stem from a tight interlacing of comprehension and production processes [17] or from shared neural underpinnings [18,19], and would be consistent with other findings on how the brain’s processing constraints shape the form of languages [20–24].

Here, we seek to resolve these questions by exploiting a contrast between aligned and nonaligned syntax in Hindi for an experiment on sentence planning. Hindi aligns noun phrases in what is known as the imperfective aspect system (“the gardener crouches/was crouching” and “the gardener plants/was planting trees”) and keeps them nonaligned in the perfective aspect system (“the gardener crouched/has crouched” versus “the gardener+x planted/has planted trees,” Fig 1).

Speakers of Hindi described pictures of 1-participant (intransitive) and 2-participant (transitive) events by using sentences with either aligned or nonaligned syntax, with alignment condition split between groups of participants. While they prepared their responses, i.e., planned their sentences, we measured neural oscillatory processing using electroencephalography (EEG) simultaneously with overt visual attention as reflected in eye gaze. Visual attention allocation during planning is related to the syntactic structures being planned [25–28] and thus indicates how contrasts in structures shape the preparation of speakers’ sentence plans. With respect to neural processes, we focus on changes in total oscillatory power (event-related desynchronization [ERD] and event-related synchronization [ERS], [29]) in the theta and alpha frequency bands, typically ranging from 3 to 7 Hz and 8 to 12 Hz, respectively. While subserving a wide range of processes [30], these frequency bands are implicated in the processing of syntactic (and possibly semantic) dependencies during the comprehension of sentences (theta: [31–39]; alpha: [34,35,40–42]). ERD and ERS are currently among the most suitable means to study the neural processes underlying early stages of sentence planning because oscillatory power can capture both evoked (phase-locked) and induced (not phase-locked) responses [43,44]. This allows the effective study of neural events involved in sentence planning that either begin at the moment when the to-be-described picture is presented or emerge during the process of sentence planning (and are therefore not phase-locked).

We target relational and structural planning processes in the first 800 ms after the picture onset, in line with crosslinguistic results on eye tracking during sentence planning [45,46]. Under the hypothesis that different processes underlie the preparation of aligned and nonaligned structures, we expected to find behavioral and neural dissociations in these early stages of planning. Specifically, we expected differences between sentence types in gazes toward agents as well as in neural oscillatory activity in the theta and alpha bands. These effects are taken to reflect speakers’ differential engagement of working memory and attentional processes [36,47–51] when committing to a specific structure earlier or later during the planning of aligned and nonaligned sentences.

Results

In the first 800 ms after picture onset, speakers fixated on the agent characters in the pictures more when planning aligned than when planning nonaligned sentences (Fig 2A). This is reflected in an interaction between the cubic time term and sentence type in a logistic mixed-effects growth curve regression [52] ( = 1.13, SE = 0.26, 95% CI = [0.63, 1.63], p < 0.001, S4 Table). EEG oscillatory power changes were analyzed with mixed-effects regression trees [53], which partition the data set by sentence type and regions of interest (ROIs; averaging across electrode positions within these regions) with respect to the time course of power changes, analyzed as polynomial growth curves (see Materials and methods). Our model identifies broadly distributed ERS in the theta band (Fig 2 B–D, S5 Table). Both intransitive and transitive aligned sentences exhibited stronger theta ERS than nonaligned sentences, starting from around 200 ms, especially over mid-frontal electrode sites, as well as subsisting after 600 ms in right-central and posterior electrode sites (Fig 2C and 2D).

Download:

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

Fig 2.

(A) Smoothed grand average of proportions of eye fixations to agent and patient characters in the picture stimuli during planning of transitive sentences (“The gardener plants trees” and “The gardener+x planted trees”). The solid vertical line represents speech onset (grand mean of transitive sentence responses). Overall, the planning of aligned sentences reveals more visual attention to agents (p < 0.001 in linear mixed-effects regression). (B) Smoothed grand average of event-related power changes (dB relative to a baseline period of −600 to −200 ms) in individually defined theta frequency bands. Grid cells in this panel represent ROIs (see S7 Table for details). Dashed vertical lines (in A and B) indicate analysis time windows, ribbons indicate standard errors, and t = 0 is the onset of stimulus picture presentation. (C) Regression model tree for power changes in individually defined theta frequency bands between 0 and 800 ms. Scalp maps show electrode positions of ROIs included in the respective grouping. All splits are statistically significant at p < 0.001. Tree tips show model fits for the respective grouping; ribbons represent 95% CIs, and colors code the same distinctions as in A and B. The model also identified lower-order splits which are subsidiary to the main difference between aligned and nonaligned sentences (see S4 Fig, S5 Table). (D) Topographic maps of power differences of aligned minus nonaligned sentences in t-values for a fixed theta frequency band (3–7 Hz). (Underlying data and scripts are available from https://osf.io/uhtcn/ and in the Supporting information file S1 Data.) Overall, the planning of aligned sentences reveals ERS (more positive dB) in the theta band. ERS, event-related synchronization; ROI, region of interest.

https://doi.org/10.1371/journal.pbio.3001038.g002

In the alpha frequency band, we found a related but different effect (Fig 3, S6 Table). Here, the main distinction is between all marked and all unmarked noun phrases (Fig 3A), rather than between aligned and nonaligned syntax (Fig 1). Specifically, we found broadly distributed ERD starting after around 400 ms and peaking around 700 ms (Fig 3B), which was larger in sentences with unmarked noun phrases (“The gardener…”) than in sentences with marked noun phrases (“The gardener+x…”). The alpha ERD effect was most pronounced over central and posterior regions (Fig 3C and 3D).

Download:

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

Fig 3.

(A) Illustration of the contrast between unmarked and marked noun phrases (which crosses over aligned and nonaligned expressions) in Hindi: Aligned expressions are always unmarked (“The gardener…”), as are nonaligned noun phrases in intransitive sentences (“The gardener crouched”), while nonaligned noun phrases in transitive sentences are marked (“The gardener+x planted trees”). (B) Smoothed grand averages of event-related power changes (dB relative to a baseline period of −600 to −200 ms) in individually defined alpha frequency bands. (C) Regression model tree for power changes in individually defined alpha frequency bands between 0 and 800 ms. The model also identified lower-order splits which are subsidiary to the main difference between marked and unmarked expression (see S5 Fig, S6 Table). (D) Topographic maps of power differences (in t-values) between sentences with marked unmarked expressions for a fixed alpha frequency band (8–12 Hz). (A–D) All other conventions are identical to the conventions in Fig 2. (Underlying data and scripts are available from https://osf.io/uhtcn/ and in the Supporting information file S1 Data.) Overall, the planning of sentences with unmarked expression (intransitive and transitive aligned sentences and intransitive nonaligned sentences) reveals ERD (more negative dB) in the alpha band. ERD, event-related desynchronization.

https://doi.org/10.1371/journal.pbio.3001038.g003

All models controlled for a series of potential confounds and group-level variables (see Materials and methods). We specifically also controlled for (cumulative) syntactic priming, the phenomenon that the repetition of sentence structures alters and eases their subsequent processing [54,55]. We included the position of each utterance in the experiment as predictor so that effects persisted above and beyond any priming effects that may accumulate over an experimental session.

There were no differences in reaction times, i.e., speech onset latencies, between the production of aligned and nonaligned sentences (see S2 Fig, S3 Table). This suggests that the task demands were indistinguishable between describing the same picture as representing an ongoing (imperfective) or a recently completed (perfective) situation. Even though we cannot rule it out categorically, the eye tracking and EEG results are therefore unlikely to stem from the fact that the planning of aligned sentences necessitated a completed conceptualization of the depicted events, while the planning of nonaligned sentences necessitated an ongoing conceptualization of the same event.

Our choice of the analysis time window makes it unlikely that we measured the encoding of phonological word forms rather than differences in syntax. It is generally assumed that the generation of grammatical structure precedes lexical encoding and the planning of phonological word forms [26,56,57], although the exact timing of the switch from grammar to phonology is not well understood [58,59]. In line with this, phonological encoding is unlikely to have influenced grammatical encoding in our measurements because variation in phonological lengths of noun phrases (in syllables) has only small effects in our models, magnitudes smaller than the effects of the sentence type differences (S4–S6 Tables). Furthermore, our analysis time window and frequency bands largely guard against possible confounds from the articulatory encoding of phonological forms in terms of speech-related muscle movement plans since these processes set in only close to actual speech onset [57,60], which occurred at least 700 ms after our analysis window (see Materials and methods), and dominantly affect frequencies that are higher (>20 Hz [61]) than the alpha and theta bands we target. Finally, in single word production, theta band ERS and alpha band ERD specifically index the retrieval of conceptual and lexical information from memory [60,62] also when the phonological form of words is held constant.

Discussion

Our results suggest that the planning of different types of sentences is neurally implemented in distinct ways. This is reflected by differences in the temporal dynamics of visual attention allocation and neural oscillatory power changes in the theta and alpha frequency bands. We propose that the observed pattern of results stems from 2 separate processes.

First, speakers directed more visual attention to agents and exhibited greater theta ERS when planning aligned than when planning nonaligned sentences (Fig 2). This suggests that they used different planning strategies for the preparation of aligned and nonaligned sentences. Speakers need to commit to a sentence plan earlier in the nonaligned condition because intransitive and transitive sentences already differ in the expression of the first noun phrase (with or without the additional agent marker “ne”). This requires speakers to prioritize early relational (and possibly structural) planning processes to encode who does what to whom [25,26]. This early (structural-)relational planning is reflected in looks that are more distributed over the whole picture and thus less focused on agents. In aligned sentences, by contrast, speakers do not need to decide on the full sentence plan as early [7]. The formal overlap between intransitive and transitive agent expressions allows speakers to delay their decision for 1 structure or the other. Speakers were thus able to primarily fixate on the agent as the action initiator [63] and as the referent of the first noun phrase in the sentence. In this way, speakers delayed the completion of relational processing [46] and held available all (aligned) structures that were compatible with a sentence starting with an unmarked agent noun phrase [10]. This scenario yields a natural interpretation of the increased theta ERS in the aligned condition as reflecting the simultaneous and noncommittal engagement of multiple compatible structures.

Intriguingly, sentence comprehension research has revealed that apart from being modulated by semantic and syntactic processing demands [32,33,37,38,64], theta oscillations play a role in maintaining working memory representations [34–37] and are linked to information retrieval from working memory [50,65]. Specifically, theta ERS is sensitive to the number of alternatives to be retrieved [50,66,67], in line with our proposed scenario of how aligned sentences are planned. The timing and frontal–parietal topography of theta ERS in our data also matches with frontal midline theta effects observed during the retrieval of syntactic structures in sentence comprehension [36].

The difference we detect in speakers’ commitment to sentence plans is consistent with behavioral findings on the importance of syntactic dependencies for the time course of planning: Like in the Hindi aligned condition, speakers of English and Japanese (where agent noun phrases are always aligned) plan agent noun phrases without an initial commitment to a verb and a sentence structure [68–70]. This contrasts with the earlier commitment that is required by sentences where the first noun phrase opens a strong dependency with the remaining sentence. In Hindi, we found this effect with marked, nonaligned noun phrases, and it is paralleled by the early commitment speakers need to make when planning patient noun phrases in English and Japanese which structurally depend on the verb [68–70]. More generally, the contrast we find between structures in Hindi expands previous findings that different dependencies afford different patterns in sentence planning [25,27,45,46,71–74].

Second, the planning of sentences with unmarked noun phrases induced larger ERD in the alpha band (Fig 3). We interpret this effect to reflect speakers’ need to keep distinct structures that share the same form (namely an unmarked noun phrase) at the beginning of sentences. This can be achieved by increased active neural information processing [29,75,76], e.g., in the form of increased engagement of cortical networks that are involved in processing syntactic information [40,42]. Converging evidence for our interpretation of the alpha ERD effect comes from reaction time findings that suggest higher processing loads when speakers need to separate alternative plans during sentence planning [9,10]. The need to keep distinct overlapping unmarked structures might also have contributed to the increased fixations to unmarked agents (Fig 2A) as longer fixation durations can indicate increased processing [49,77].

Our interpretation of the alpha ERD effect as reflecting the processing of syntactic alternatives is further supported by its central–posterior topography and its latency, with the largest differences setting in around 600 ms. The topography and latency are similar to alpha ERD effects associated with syntactic processes in sentence comprehension [34,35,40–42]. The current topographies are at the same time consistent with neuromodulatory evidence on the role of theta ERS and alpha ERD in controlling working memory–related processes [47]. Moreover, in conversation, alpha ERD is associated with shifting from comprehending interlocutors’ turns to preparing one’s own production [78,79]. Our results show that the observed alpha ERD effects go beyond mere attention shifts to production and are also sensitive to differences in syntactic planning processes between sentence types.

Our finding that theta power increases are associated with alpha power decreases fits with previous reports on memory processes outside language [80]. However, our results also suggest a small timing difference: While alpha ERD was most pronounced toward the end of our analysis time window, around 700 ms, theta ERS was most pronounced around 300 ms. This could reflect that speakers decided on which structure to produce in aligned contexts probably between 300 and 700 ms after the start of sentence planning. Thus, while eye tracking allows the detection of distinct steps in the planning process, our concurrent multimodal approach demonstrates that the time course of sentence planning can be characterized in greater detail with additional evidence from highly time-resolved electrophysiological recordings. The combination of eye gaze and neural oscillatory power changes opens up the possibility of additional insights into the fine-grained structure of sentence planning in future studies.

The origin of oscillatory activity can be exogenous or endogenous for the comprehension and perception of language [81–83]. For language planning and production, however, there is no exogenous source that would obviously give rise to power changes in the theta and alpha bands. Participants were “externally stimulated” by pictures that were to be described—the same ones for aligned and nonaligned sentences. Theta ERS and alpha ERD in the planning of different sentence types therefore most likely reflect the internal processing of sentence structures, building on memory-related control processes [62].

The increased neural activity in the planning of aligned and unmarked syntax (in terms of theta ERS and alpha ERD) contrasts with the apparent simplicity and frequency of systems with aligned expressions among the world’s languages. This challenges theories that propose sentence production processes as the key drivers of language form and the distribution of variants in language use [84,85]. These theories are based on the observation that speakers shape language use through choosing sentence structures that cause the least production difficulty, specifically, by placing easier to retrieve words first [86–88] and by reducing interference from similar words and structures [84,89,90]. Other structures would then be less likely to be produced and could only be used in restricted contexts or disappear from a language over time [84].

Our findings suggest either that (i) the relevant notion of production ease is independent of the neural activity we found and is instead grounded in more global constraints on sentence planning or that (ii) the distribution of aligned expressions in the world’s languages is not driven by sentence planning processes.

Under the first possibility, what matters most for speakers is fluency in sentence planning. Delaying the commitment to 1 grammatical structure is arguably beneficial for speakers’ fluency as it allows more flexibility when creating and implementing sentence plans incrementally [88,91]. Aligned expression facilitates this. The resulting benefit, e.g., for rapid response planning in conversation [78,92,93], would then be more important for language evolution than the increase in neural activity we found under aligned expression and the interference effects that this might generate.

The second possibility derives the prevalence for aligned expression from comprehension, specifically from a general preference of the sentence comprehension system to expect sentences to start with an unmarked noun phrase in the agent role [12–16]. Intriguingly, this expectation extends to languages with nonaligned syntax. For example, the nonaligned syntax of Hindi leads to unmarked noun phrases that regularly denote patients instead of agents in transitive sentences (e.g., the patient “paudhe” “trees” in Fig 1 is unmarked). When these noun phrases occur initially in a sentence (as they often do), the comprehension parser nevertheless assigns them an agent role transiently and then shows an electrophysiologically detectable effect of reanalysis and prediction failure at the position of the verb [3]. This incurs an additional processing demand [94,95] that might explain the bias against nonaligned syntax in the world’s languages.

Under this view, differences in planning processes between types of sentences have much less of an effect on linguistic distributions than processes during sentence comprehension. This is in line with theories of sound change that locate such effects chiefly in perception rather than production [96]. Prediction failure and reanalysis effects in comprehension would thus exert a vastly greater pressure on languages to abandon these sentence structures over time than the increased processing demands during the planning of aligned structures. This view is consistent with proposals that processing differences related to working memory engagement are key constraints of language evolution [20,21,97,98], to the extent that these differences stem from comprehension.

Alternatively, comprehension and production might form a trade-off: While the comprehension system favors unmarked, aligned syntax, production favors marked, nonaligned syntax, although to a much weaker extent. This would explain why nonaligned systems exist at all, despite their lower probability to emerge and persist during the evolution of languages [3].

More research is now needed to resolve these questions and to probe contrasts in alignment across diverse syntactic structures and languages. Our results demonstrate how new insights on the relationship between language production and the simplicity and prevalence of specific language structures critically rely on evidence from neural processes that underlie different syntactic variants in the earliest stages of sentence planning. This opens new avenues for research on the neural processes in language planning and speech production that go beyond the level of individual words [62,99,100] and beyond the small set of languages that have dominated the field so far [22,71,74,101–103].

Materials and methods

Supporting information

Acknowledgments

We thank Cyril Bosshard, Samina Fayyaz, and André Müller for help with processing the audio recordings; Anna Jancso and Lei He for programming help; and Prabir Sarkar for providing access to the eye tracker.