Does entropy modulate the prediction of German long-distance verb particles?

Kate Stone; Shravan Vasishth; Titus von der Malsburg

doi:10.1371/journal.pone.0267813

Abstract

In this paper we examine the effect of uncertainty on readers’ predictions about meaning. In particular, we were interested in how uncertainty might influence the likelihood of committing to a specific sentence meaning. We conducted two event-related potential (ERP) experiments using particle verbs such as turn down and manipulated uncertainty by constraining the context such that readers could be either highly certain about the identity of a distant verb particle, such as turn the bed […] down, or less certain due to competing particles, such as turn the music […] up/down. The study was conducted in German, where verb particles appear clause-finally and may be separated from the verb by a large amount of material. We hypothesised that this separation would encourage readers to predict the particle, and that high certainty would make prediction of a specific particle more likely than lower certainty. If a specific particle was predicted, this would reflect a strong commitment to sentence meaning that should incur a higher processing cost if the prediction is wrong. If a specific particle was less likely to be predicted, commitment should be weaker and the processing cost of a wrong prediction lower. If true, this could suggest that uncertainty discourages predictions via an unacceptable cost-benefit ratio. However, given the clear predictions made by the literature, it was surprisingly unclear whether the uncertainty manipulation affected the two ERP components studied, the N400 and the PNP. Bayes factor analyses showed that evidence for our a priori hypothesised effect sizes was inconclusive, although there was decisive evidence against a priori hypothesised effect sizes larger than 1μV for the N400 and larger than 3μV for the PNP. We attribute the inconclusive finding to the properties of verb-particle dependencies that differ from the verb-noun dependencies in which the N400 and PNP are often studied.

Citation: Stone K, Vasishth S, von der Malsburg T (2022) Does entropy modulate the prediction of German long-distance verb particles? PLoS ONE 17(8): e0267813. https://doi.org/10.1371/journal.pone.0267813

Editor: Nicola Molinaro, Basque Center on Cognition Brain and Language, SPAIN

Received: October 5, 2021; Accepted: April 18, 2022; Published: August 4, 2022

Copyright: © 2022 Stone et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Data Availability: All data, code, and materials are openly available at: https://osf.io/h75jm/ DOI 10.17605/OSF.IO/H75JM.

Funding: Kate Stone was supported by a doctoral scholarship from the Potsdam Graduate School. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Competing interests: The authors have declared that no competing interests exist.

Introduction

Contextual cues in a sentence preactivate words in memory, such that processing of those words begins before they are seen [1, 2]. If contextual cues are sufficiently constraining, preactivation may crystallise into prediction of a specific word. Such lexical predictions can benefit readers in processing long-distance dependencies, where key information to interpreting an event is delayed [3–5]. Take for example the German particle verb construction in (1), where a large amount of material separates the verb “fuhr” carried from its particle “fort” on:

(1). Der Professor fuhr mit seinem Vortrag trotz regelmäßiger
The professor carried with his lecture despite regular
Störungen fort.
interruptions on.

Rather than waiting for the particle to appear, and thus waiting to know the exact meaning of the sentence, readers may predict “fort” in order to facilitate processing of the intervening material. However, other plausible particles could include “ab” (the professor drove off), “zusammen” (the professor was startled), or “zurück” (the professor reversed [his position] or drove back). This creates uncertainty about the intended meaning of the sentence. In this paper, we use German particle verbs to test how uncertainty may affect predictive processing difficulty in long-distance dependencies. Although we did not find a conclusive answer to this question, we report our attempt as a resource for future researchers who may wish to revisit the issue. Furthermore, by reporting the inconclusive results, we avoid contributing to publication bias via the file drawer problem [6].

Prediction in long-distance dependencies

Evidence that readers make long-distance predictions is most often observed for syntactic information, including in verb-particle constructions. Piai et al. (2013) found evidence that Dutch verbs that could take a particle elicited a larger left anterior negativity (LAN) than those that never took a particle. The LAN has been associated with maintaining information in working memory [7], suggesting that the verb triggered a prediction for a particle. However, there was no evidence that the number of different particles licensed by the verb affected LAN amplitude, suggesting that the prediction was for the syntactic features of the verb-particle dependency and not for its lexical properties. The LAN effect was still present at the direct object, suggesting that readers maintained their syntactic prediction while processing the intervening sentence.

Further evidence for long-distance syntactic predictions comes from filler-gap dependencies such as I know what_i you hit and broke the mirror with ___i [8]. Encountering a transitive verb within a wh-dependency is argued to induce active prediction of a gap site, inducing longer reading times in the region of hit than at the same point in the non-wh-dependency sentence I know that you hit and broke the mirror with the ball [8–11]. This predicted gap is assumed to be maintained until it is resolved. Evidence from event-related potentials (ERPs) supports predicted gap maintenance by showing that a sustained anterior negativity (SAN) is larger for wh-questions that are expected to be long vs. short; this has been attributed to a higher burden placed on working memory when a longer structure is expected [12–14].

Maintenance of a syntactic prediction should facilitate downstream processing, and indeed such facilitation has been observed in reading times. For example, encountering the word either facilitates processing of a co-ordinate or structure in sentences like “The team took either the train or the subway…” compared to “The team took the train or the subway…” [15, 16]. This suggests that either triggers a syntactic prediction for the co-ordinate structure, which is sustained to facilitate reading at its resolution. Staub and colleagues also observed that syntactic predictions facilitated processing by protecting against misanalysis of S-coordination as NP-coordination. When either was present in sentences such as “Either Linda bought the red car or her husband leased the green one”, the NP her husband was never misanalysed as the direct object of Linda bought. However, when either was absent, a strong preference for NP-coordination appeared to trigger misanalysis, as suggested by a larger number of regressions in the spillover region following her husband. Long-distance syntactic predictions may therefore facilitate processing through both preactivation and protection against misanalysis.

In contrast to syntactic prediction, lexical prediction may represent commitment to a more specific sentence interpretation. Specific predictions can increase the facilitatory advantage of a correct prediction, but also run an increased risk of costly misprediction. The increased risk may explain why Piai and colleagues (2013) did not find ERP evidence of lexical prediction in their particle verb stimuli, as discussed above.

Commitment to a specific sentence interpretation refers to one account of the predictive mechanism in sentence processing where a specific interpretation might be stored in working memory [1, 2, 17, 18]. However, the mechanisms underlying predictive processing are by no means resolved. An alternative account is that the level of preactivation alone is sufficient to explain predictability effects [19]. This is supported by the success of metrics such as surprisal that measure the probability of a word given its context—which could be assumed to be linked to word preactivation—in accounting for variability in predictability-related empirical effects [20–25]. On the other hand, it has been shown that suprisal alone is not sufficient to explain all differences in reading times or the N400 [26, 27]. In the case of reading times, this was taken to suggest the need of an additional explanatory factors such as reanalysis or revision [26]. Assuming that reanalysis or revision can only be triggered if a commitment has been made, such findings could be taken as evidence that predictions can involve commitments.

On the other hand, reading time studies do provide some support for long-distance lexical predictions. A study of particle verb-like constructions in Persian found that increased lexical predictability of a head-final verb facilitated reading times in both short- and long-distance dependencies [4]. One interpretation of these results is that readers predicted the verb early in the sentence and sustained this prediction until it was resolved at the verb site. However, the pre-verbal information in both short- and long-distance conditions may have been informative about the identity of the verb, and so it is possible that the verb in this study was only predicted immediately prior to its being seen.

In similar constructions in Hindi, increased lexical predictability was also found to facilitate reading times at both short and long distance [3]. The pre-verbal region in this study was not informative about the identity of the verb, and so it could reasonably be concluded that the predictability effect in the long-distance condition was evidence of long-distance lexical predictions.

In support of the Hindi findings, a study of German particle verbs found that higher certainty about the identity of a particle facilitated eye fixation times at the particle in both short- and long-distance conditions [5]. In this study again, the pre-particle region was not informative about the identity of the particle, supporting the conclusion that long-distance predictions may have been made. However, the uncertainty effect was not replicated in self-paced reading times, suggesting that the effect of lexical certainty on eye movements may have been quite subtle.

Taken together, the above studies suggest that long-distance predictions are generated and maintained to facilitate the processing of long-distance dependencies. However, whether lexical content is included in these predictions appears to be more difficult to observe. One possibility is that higher variability among lexical versus syntactic continuations of a sentence creates a higher degree of uncertainty for the parser. Combined with the likelihood that intervening material may require the long-distance lexical prediction to be reanalysed, lexical predictions may be discouraged in all but the highest certainty situations. Even the potential facilitatory benefit of a correct prediction to processing the predicted word and the intervening material may be outweighed by potential reanalysis cost.

The cost of uncertainty in predictive processing

The probability of a misprediction increases as uncertainty about a sentence increases, making uncertainty a potentially important factor in whether readers make predictions about long-distance dependencies. A reader’s uncertainty at a given point in a sentence can be measured with entropy, which quantifies the distribution of different ways a sentence could continue [20]; for example, via the number of different continuations provided at a particular point in the sentence in a sentence-completion task. A link between entropy and predictive processing cost has been observed in reading times. Encountering a word that triggers entropy among plausible sentence continuations has been associated with longer reading times, suggesting that uncertainty creates processing cost via the prediction and maintenance of multiple potential continuations in memory [28, 29]. Large reductions in entropy have also been associated with increased reading times, suggesting that large shifts in expectation also incur a processing cost [20, 24, 30, 31].

The relationship between entropy and processing cost is less well-studied in event-related potentials (ERP). Several experiments have examined the effect of contextual constraint on predictive processing, and constraint can be interpreted as a measure of entropy at a target region of a sentence. However, constraint in such studies is usually quantified using the highest probability word given at the target region in a sentence-completion task (a cloze test); for example, see [32–34]. The degree of entropy among different cloze test completions is usually not taken into account.

Entropy among cloze test completions is important if we assume that it represents memory activations in a single reader [35]. Consider a sentence where the most strongly activated word at a given point has a cloze probability of 36% and that two other activated words each have a probability of 32% (high entropy). A reader would be highly uncertain about how the sentence will continue, given there are three almost equally probable options. Compare this to a sentence where the strongest activated word also has a probability of 36%, but where there is a large spread of other activated words each with a probability of no more than 5% (lower entropy). The reader can now be relatively certain about how the sentence will continue, even though the target word has the same probability as in the first sentence. Under the highest-cloze-probability definition of constraint used in many ERP studies, both sentences would be grouped together as weakly constraining, even though the difference in uncertainty will have different consequences for predictive processing.

The highest-cloze-probability definition of constraint may explain inconsistency among ERP findings about the effect of constraint on processing cost. In particular, studies have observed a larger anterior late positivity (PNP) for unexpected words in strongly versus weakly constraining contexts, suggesting that higher certainty leads to predictions which are more costly when disconfirmed [32, 33, 36]. The PNP has been linked to the suppression of disconfirmed sentence representations, potentially shedding light on the mechanism driving the processing costs observed in reading times [37]. However, not all studies of the PNP find an effect of constraint, or have observed it in the opposite direction [34, 38–42].

Few studies have directly tested entropy in ERP, as opposed to constraint. In an exploratory analysis of Thornhill and Van Petten’s (2012) data, a reduction in entropy at the target word was found to explain some variance in the PNP, although this did not survive a confirmatory analysis [23]. In Piai et al.’s (2013) study of Dutch particle verbs, verbs that took large number of particles were not found to elicit differences in the left anterior negativity (LAN) compared to verbs that took a small number of particles. While entropy among particles was not explicitly quantified, the findings can be interpreted as showing that differences in lexical uncertainty did not result in differences in working memory burden, as indexed by the LAN. The processing cost of disconfirmed predictions in relation to entropy was not investigated. Given that uncertainty may be a deciding factor in predictive processing of long-distance dependencies and that ERP has the potential to reveal mechanisms driving processing cost, we conducted two ERP experiments examining the effect of entropy on predictive processing.

The current experiments

The aim of the experiments was to test whether lexical entropy would affect readers’ willingness to predict a long-distance verb particle. We compared sentences which required a particle to be semantically plausible, but varied in how many different particles could complete the sentence. Since we were specifically interested in the effect of entropy among the different particles’ lexical representations, it was strictly controlled: In the example experimental item (2) below, only a single particle was semantically plausible in conditions (a) and (b). In conditions (c) and (d), at least two particles were plausible. We hypothesised that readers should experience more uncertainty about the meaning of the sentence in (c/d), and thus be less likely to predict a specific particle given the increased potential for a misprediction and its associated cost:

(2). 1 plausible particle, expected:
1. a. Der ordentliche Professor fuhr mit seinem Vortrag trotz
  The orderly professor carried with his lecture despite
  regelmäßiger Störungen immer ordnungsgemäß fort, da er für
  regular interruptions always properly on, as he for
  seine Unaufgeregtheit bekannt war.
  his unflappability known was.
1 plausible particle, violation:
1. b. *Der ordentliche Professor fuhr mit seinem Vortrag trotz
  The orderly professor carried with his lecture despite
  regelmäßiger Störungen immer ordnungsgemäß mit, da er für
  regular interruptions always properly with, as he for
  seine Unaufgeregtheit bekannt war.
  his unflappability known was.
2+ plausible particles, expected:
1. c. Der ordentliche Buchhalter fuhr seinen zuverlässigen
  The orderly accountant turned his reliable
  Computer bei der Arbeit immer ordnungsgemäß herunter/hoch,
  computer at work always properly off/on,
  da er für seine korrekte Arbeitsweise bekannt war.
  as he for his correct work practices known was.
2+ plausible particles, violation:
1. d. *Der ordentliche Buchhalter fuhr seinen zuverlässigen
  The orderly accountant turned his reliable
  Computer bei der Arbeit immer ordnungsgemäß mit, da er für
  computer at work always properly with, as he for
  seine korrekte Arbeitsweise bekannt war.
  his correct work practices known was.

We used the N400 and the post-anterior positivity (PNP) to measure processing cost at the particle. In particular, we were interested in ERP evidence of processing cost when an unexpected, semantically implausible particle was encountered in conditions (b) and (d). We focused our inference on these conditions because the target particles both had a cloze probability of zero. This allowed us to rule out the possibility that any ERP difference observed at the particle was due to differences in cloze probability, which correlate strongly with amplitude of the N400 and the PNP [33, 43]. While cloze probability is a measure of a word’s predictability, it does not necessarily equate to whether a word is predicted: Preactivation of a word by its context—which increases cloze probability [35]—also makes that word easier to access in the lexicon and integrate into the building sentence representation once it is seen, without it having been specifically predicted. We assume a difference between preactivation and prediction, where preactivation can reflect activation of several different items in memory, while prediction reflects commitment to one specific item, for example by adding it to working memory [1, 2, 17, 18]. Differences in ERP amplitude associated with a target word’s cloze probability may therefore also be driven by non-predictive processes triggered by having seen that word. In contrast, a difference between two target words with the same cloze probability but differing in entropy—the distribution of words not seen but preactivated—should be informative about predictive processing. The target words in (b) and (d) were always identical, and the pre- and post-critical regions were matched between conditions to rule out any influence of the preceding or following words on the ERP. The pre-critical region was not informative about the identity of the particle.

We use the term “zero probability” when referring to the implausible particles to reflect the fact that the sentence context is highly unlikely to preactivate implausible particles in memory as they would be meaningless in the context. However, there exists evidence from the N400 demonstrating that readers can be sensitive to differences within the very low end of the probability spectrum, including for implausible words [25]. These activations are unlikely to be captured by a cloze test and thus have a cloze probability of zero, but their preactivation may still affect ERP amplitude. This would be a caveat to our “matched zero probability” target particles. The contribution of very low probability words to entropy and ERP amplitude needs future investigation, but for the current study we use the term “zero probability” to refer broadly to a situation where, for the purposes of studying the strength of people’s probabilistic representations of the context of a sentence, the contribution of preactivation from an implausible particle is effectively zero.

To detect the cost of violated predictions, we examined two ERP components: the N400 and the anterior post-N400 positivity (PNP). The N400 is a negative deflection in the ERP occurring at around 250–500 ms after a word is seen and is highly correlated with a word’s probability given the preceding sentence context [43]. Amplitude of the N400 decreases at each new, context-congruent word in a sentence [44, 45] and increases at any unexpected word, inversely proportional to its probability [43, 46]. This has led to the hypothesis that sentence context allows the preactivation of probable words that are then easier to process once encountered, and that the N400 reflects the lexical access or update cost incurred by words with low preactivation [43, 47].

Previous studies involving uncertainty have tested the effect of contextual constraint on the N400, but have not found that constraint affects its amplitude when the cloze probability of the target words is matched [32, 33, 41, 46, 48]; but see [49, 50]. That is, two words with the same low probability will trigger N400s of the same amplitude, even if one of them is found in a strongly constraining context and the other in a weakly constraining context. Lexical access and probabilistic update models of the N400 explain this as occurring because each low probability word has received no preactivation, meaning that lexical access is equally unfacilitated [43] or that the shift in model update is equally large [47]. This presents a testable null hypothesis for the current study: if the N400 is not affected by constraint—which should reflect entropy—then we should not expect to see a difference in N400 amplitude between 1- and 2+particle violations. However, given that few previous studies of the N400 have directly tested entropy, it is possible that we could see a larger N400 in the 1-particle condition reflecting greater update or more difficult lexical access after a misprediction.

The PNP is a positive amplitude deflection in the anterior region of the scalp in the 600–900 ms window [32, 33, 46]. The PNP appears to be functionally distinct from the more well-known posterior P600 [32, 33], which has been associated with conflict detection and repair processes [51–56]. In contrast to the P600, the PNP has been associated with suppression of disconfirmed sentence representations [33, 37] and appears only to be sensitive to plausible but unexpected words, suggesting that suppression may only occur if there is a viable, alternative way to update the event representation [32, 33, 57]. If this is the case, we may not expect to see a PNP modulation by our implausible particles.

However, in addition to inconsistent findings about the effect of constraint on the PNP already mentioned [34, 38–42], the plausibility feature has only ever been tested in verb-noun dependencies. A context-violating noun may be implausible at multiple levels of representation (e.g. animacy, thematic role, semantic), whereas our implausible particles present a purely semantic violation. It is thus less clear how the PNP might be affected. If purely semantic violations can elicit the PNP, we may see a higher cost (larger PNP) of having to suppress a disconfirmed sentence representation at the violation particle in the 1-particle condition than in the 2+particle condition. If not, then we may see no difference associated with entropy if the PNP is not triggered by the implausible violations.

We tested these hypotheses in two ERP experiments: in Experiment 1, we used the 2 × 2 design set out in example (2) to compare the effect of the implausible particles on the N400 and PNP, but also to confirm that the particles could elicit an N400 and PNP in the implausible versus plausible conditions. In Experiment 2, we attempted to replicate the findings of Experiment 1 using a simpler design with a larger number of participants and experimental items. Data, code, and materials for both experiments can be found at https://osf.io/h75jm/.

Experiment 1

Experiment 1 tested whether entropy would affect readers’ predictions about an upcoming verb particle. We inferred that predictions had been made by measuring processing cost at prediction violations. We predicted that a larger N400 and PNP would reflect the greater cost of recovering from a mis-prediction in the 1-particle condition than in the 2+particle condition where a prediction was less likely. Experiment 1 was pre-registered at https://osf.io/qbna2; see also S1 Appendix.

Materials and methods

Participants.

Fifty-four participants were recruited. The sample size was determined by recruiting as many participants as was possible during one university semester. Four participants were excluded due to a large number of their target EEG trials being contaminated by muscle and/or blink artefacts. This left a total of 50 subjects (6 male), with a mean age of 25 years (range: 17–40 years, SD = 5 years). In line with university policy, all participants were reimbursed for their time either financially or in the form of credit points toward their studies. All participants were right-handed German native speakers, with no known history of developmental or current language, neurological, or psychiatric disorder, and had not participated in the cloze test. The study was conducted in line with the principles of the Declaration of Helsinki and all participants provided written consent to participation in the study. In accordance with German law, IRB review was not required as the study involved only healthy adult participants.

Materials.

For each particle verb, two sentences were constructed, as seen in Example (2). The position of the base verb and particle in each sentence pair was matched. The particle was the target word. Within each sentence pair, at least two words before the particle were identical. 103 sentence pairs were constructed and presented as a cloze test to 30 German native speakers (mean age 25 years, SD 6 years, range 18–41 years) on a desktop computer in our in-house lab using the Ibex software [58]. The sentence pairs were divided into two lists, such that each participant only saw one condition from each item. The particle of the sentence was replaced by a gap, which participants were asked to fill with the first word that came to mind.

The items were then ranked in terms of how well they fulfilled the criteria that the 1-particle condition elicited only one particle and the 2+particle condition elicited at least two particles with similar probability. To rank the items, cloze probabilities were calculated for each particle completion. Other kinds of completions were grouped into categories (e.g. prepositional phrases, adjectives, nouns) and a cloze probability was calculated for each category. Items were then ranked by entropy among the responses (lowest to highest in the 1-particle condition; highest to lowest in the 2+particle condition). For the 2+particle condition, further weight was given to items where the two highest ranked particles were close in cloze probability. This ranking scheme left a final set of 40 plausible items fulfilling the criteria of the experiment. Cloze probability and entropy statistics are summarised in Table 1.

Download:

Table 1. Cloze probability and entropy summary statistics for Experiment 1.

https://doi.org/10.1371/journal.pone.0267813.t001

To create the violation conditions, two German native speakers selected particles that were not integrateable into the sentence context, including illicit verb-particle combinations. No participant in the cloze test produced any of these illicit particles. The same particle was used in both violation sentences within each item. The 40 critical sentences used in the analysis were split into four lists in a Latin square design, such that each participant only saw one of the four conditions for each critical item and thus a total of 10 items per condition. Critical sentences were pseudo-randomly interspersed with 98 filler sentences. Sixty-two filler sentences contained plausible sentences in a variety of lengths and tenses, some of which contained non-separated particle verbs (e.g. hochgefahren, switched on). The remaining 36 fillers were particle verb sentences from the cloze test that did not fulfill the 1- vs. 2+particle criteria. In total, each participant saw a total of 138 sentences. The ratio of plausible to implausible sentences was approximately 3:1. The order of presentation of sentences within each list was pseudo-randomised via the presentation software. Each sentence was followed by a yes/no question appeared that probed different regions of the sentence; for example, the question for example (2) above was “Verlief der Vortrag ungestört?” did the lecture proceed uninterrupted?.

Procedure.

Participants sat in a shielded EEG cabin approximately 60 cm from a 56 cm presentation screen. The experimental paradigm was built and presented using Open Sesame [59]. Each experimental session began with an instruction screen advising participants that they would read sentences presented word-by-word and that after each sentence, they would answer a question using a video game controller. Participants were instructed to answer as quickly and accurately as possible. Each experimental session began with four practice trials.

Each trial in the experiment began with a 500 ms fixation cross in the centre of the screen followed by a blank screen jittered with a mean of 1000 ms and standard deviation 250 ms. Each sentence was presented word-by-word for a duration of 190 ms per word plus 20 ms for each letter. The target word was always presented for 700 ms regardless of length. The inter-word interval was 300 ms. A comprehension question appeared after each sentence and was answered via the video game controller, which triggered the next trial. Breaks were offered after every 30 sentences. The testing session including EEG setup lasted approximately two hours.

EEG recording and preprocessing.

The EEG recordings were made in the Department of Linguistics at the University of Potsdam, Germany, in a purpose-built, electro-magnetically shielded EEG cabin using a 32-lead system and electrodes arranged on the head based on the international 10–20 system [60]. Electrode impedances were kept below 10kΩ throughout the experiment. EEG was recorded at a sampling rate of 512 Hz and online filtered with a low-pass filter of 138 Hz and using the left mastoid as a reference.

Raw EEG recordings were downsampled offline in BrainVision Analyzer 2, Version 2.1.2, to 500 Hz for ease of processing and interpretation. Zero phase shift IIR Butterworth filters were applied with a low-pass cut-off at 0.01 Hz (order of 2, time constant of 15.92) and a high-pass cutoff at 30 Hz (order of 2, no time constant). A notch filter was applied at 50 Hz. The full recording was then segmented into epochs from sentence onset to question onset. Ocular correction using restricted Infomax was applied to the sentence epochs using automatic independent component analysis (ICA) with a meaned slope algorithm. The reference electrodes were two electrodes placed at the left outer canthus and above the left eye to record horizontal and vertical eye movements, respectively. The bound number of blinks was 60 with a convergence bound of 10⁻⁷. The number of ICA steps was 512. Components were found using sum of squared correlations with the horizontal and vertical ocular electrodes [61–63]. The total value to delete was 30%.

The EEG was then re-referenced to the average of the two mastoids. The corrected segments were further segmented into 1200 ms epochs representing a period of 200 ms before the onset of the target word (the particle), and 1000 ms after onset. EEG segments with muscle artefact or irreparable eye-blink or movement artefact were automatically marked for 200 ms before and after each respective artefact, defined as exceeding:

a maximum voltage step of more than 50 μV
a maximum absolute difference of 200 μV in a 100 ms interval
a minimum amplitude of -100 μV
a maximum amplitude of 100 μV
a minimum low activity of 1 μV in a 100 ms interval.

Marked segments were then visually inspected and discarded if they indicated muscle artefact or a technical issue. This resulted in the exclusion of 6.20% of the 1000 target trials used in the statistical analysis. A further 0.50% of trials were excluded due to question response times over 10 seconds (indicating a technical problem) and 0.30% were not recorded due to experimenter error. The data were then exported and baseline-corrected in R using the package eeguana [64].

Analysis.

A linear mixed effects model with full variance-covariance matrices to account for the individual variability of subjects and items was fit to data from the violation conditions (b) and (d) using the brms package for R [65]. The dependent variable for the N400 was mean amplitude across the electrodes Cz, Pz, CP1, and CP2 in the 250–500 ms time window. The dependent variable for the PNP was mean amplitude across the electrodes Fz, FC1, and FC2 in the 600–900 ms window. For Experiment 1, these regions were chosen post hoc after we deviated from the pre-registered single-electrode analysis (see S1 and S2 Appendices). The regions were generally consistent with the distribution of the N400 and PNP observed in the predictive processing literature [32, 33, 66], although some studies observe a more anterior PNP [33, 36]. Since they were chosen post hoc, the regions of interest in Experiment 1 were used as pre-registered regions in Experiment 2.

The predictor ‘number of plausible particles’ was effect contrast coded to reflect our expectation for larger ERPs at 1-particle violations: 0.5 (1-particle), -0.5 (2+particle). The decision to use a categorical predictor rather than raw entropy was made because the precise functional relationship between entropy and ERP amplitude was not known, and we felt unable to safely assume it was linear.

To constrain the model to plausible values and avoid overfitting, we placed regularising priors of N(0, 1) on each parameter [67–70]. This prior with a normal distribution centered on zero indicates to the model that, for example, the effect of the number of plausible particles will most likely be small or zero, with either a positive or negative sign, and has a 95% probability of lying between −2 and 2μV.

To quantify how much evidence we had for any N400 and PNP effect, we conducted Bayes factor analyses comparing models with the predictor number of particles (M1) versus a reduced model without this predictor (M0), i.e. BF₁₀. A Bayes factor of approximately 1 would indicate no evidence in favour of either model. As the Bayes factor increases over 3, evidence strengthens in favour of M1 [71, 72]. As the Bayes factor decreases under 0.3, evidence strengthens in favour of M0. Since the Bayes factor is sensitive to the choice of prior [73, 74], we conducted a sensitivity analysis by computing Bayes factors for a range of priors to determine how these affected our conclusions [75, 76]. We used a range of priors that assumed a priori effect sizes ranging from small, between −0.2 and 0.2μV, to large, between −10 and 10μV.

To confirm that the violation particles elicited an N400 and PNP relative to their plausible counterparts in line with previous studies [77], we fit the above models to all conditions (a-d) and added the predictor ‘plausibility’. Plausibility was effect contrast coded to reflect our expectation for larger ERPs in the implausible conditions: 0.5 (implausible), -0.5 (plausible) and a Bayes factor was computed at a prior of N(0, 1). Since the plausible particles differed in cloze probability and identity from each other and from the violation particles, we used this analysis as a sanity check only and did not conduct any further analyses.

Results

Mean amplitude relative to the onset of the target particle is plotted in Fig 1.

Download:

Fig 1. Experiment 1: ERP results.

A. Average amplitude following stimulus onset shows an N400 in the posterior region of interest (Cz, CP1, CP2, Pz) for violations of both the 1-particle (red) and 2+particle (blue) conditions relative to the plausible (grey) conditions. Ribbons show 95% confidence intervals. These intervals were calculated by fitting a linear model of the form amplitude ∼ condition with by-subject and by-item intercept adjustments at each timestep of the ERP recording and extracting the standard error of the relevant conditions. The analysed N400 time window is overlaid. B. A voltage subtraction map of the N400 time window shows a small difference in amplitude between the 1-particle (a) and 2+particle (b) conditions in the region of interest (dashed square). C. A PNP for the 1-particle condition is seen in the anterior region of interest (Fz, FC1, FC2) 600–900 ms window after stimulus onset, but not in the 2+particle or plausible conditions. Ribbons show 95% confidence intervals. D. A voltage subtraction map of the PNP time window suggests amplitudes in the 1-particle condition were more positive in the mid-frontal region (dashed triangle) than in the 2+particle condition.

https://doi.org/10.1371/journal.pone.0267813.g001

In Fig 1A and 1B, the 250–500 ms window of the ERP shows a strong negative deflection over the posterior scalp for the implausible conditions relative to the plausible conditions, consistent with an N400. Implausible particles also elicited a large positive deflection in the 600+ ms time window relative to plausible particles over the posterior scalp, consistent with a P600.

Fig 1C and 1D shows a positive deflection in the post-N400 period of the ERP (i.e. after 500 ms) in fronto-central electrodes consistent with a PNP in the 1-particle condition relative to the 2+particle condition and relative to the plausible conditions.

Analysis of the N400.

At implausible particles, the N400 was more negative in the 1-particle condition than in the 2+particle condition and the mean of the posterior was −0.34μV with a 95% credible interval (CrI) of [−1.10, 0.39] μV (Fig 2A). The Bayes factor at a prior of N(0, 1) favoured neither the null or alternative hypotheses about an entropy-related difference in the N400, BF₁₀ = 0.57. The sensitivity analysis suggested there was moderate evidence against a priori hypothesised large effect sizes, but was inconclusive about small effect sizes (Fig 2B). There was strong evidence that the N400 was more negative for implausible than plausible particles, .

Download:

Fig 2. Experiment 1: Posterior estimates and Bayes factor analysis for the planned and exploratory comparisons.

A. The posterior distribution represents the estimated change in N400 amplitude associated with violations in the 1-particle condition relative to the 2+particle condition. The point and errorbar reflect the posterior mean and 95% credible interval. B. Bayes factors computed under a range of priors comparing models with and without a predictor for the 1- vs. 2+particle condition manipulation. A Bayes factor of 1 indicates no evidence in favour of either model. Bayes factors below 1 indicate evidence against a difference in N400 amplitude associated with the manipulation. C. Posterior distribution of the difference in PNP amplitude between the violation conditions. D. Bayes factors under a range of priors for the PNP effect.

https://doi.org/10.1371/journal.pone.0267813.g002

Analysis of the anterior post-N400 positivity (PNP).

At implausible particles, the posterior mean indicated more positive amplitude in the 1- vs. the 2+particle condition, , 95% CrI [−0.25, 1.77] μV (Fig 2C). However, the Bayes factor at a prior of N(0, 1) favoured neither the null or alternative hypotheses about an entropy-related effect on the PNP, BF₁₀ = 1.50. The sensitivity analysis suggested effect sizes of approximately 1μV were more plausible than smaller or larger effects, although the Bayes factors even at 1μV were inconclusive (Fig 2D). The PNP was more positive for implausible than plausible particles, although the Bayes factor suggested only anecdotal evidence for this conclusion, .

Interim discussion

We compared sentence pairs where either a single verb particle or a small set of verb particles were plausible continuations given the context. We hypothesised that lower entropy would encourage commitment to a lexical prediction in the 1-particle condition, and that violating this prediction would be more costly relative to the 2+particle condition where higher uncertainty discouraged prediction. We predicted that this cost would be reflected in a larger N400 and PNP. Contrary to this prediction, evidence about the effect of entropy on both the N400 and PNP was inconclusive. The comparison of plausible versus implausible particles elicited a difference in both component in line with the typical plausibility effect observed in previous research, including research on German particle verbs [77], confirming that the experiment worked in principle, although evidence for the effect on the PNP was less conclusive than expected.

While Bayes factor evidence for the effect of entropy on the N400 was inconclusive, the magnitude of the effect was small, which could be considered consistent with the null hypothesis of lexical access and probabilistic update accounts [42, 43, 47]. Under these accounts, the N400 at zero-probability words should not be affected by uncertainty caused by multiple activated words. One caveat to this conclusion is that, if there were truly no N400 effect, one could have expected stronger support for the null hypothesis from the Bayes factor analysis, especially since the cloze probability manipulation was able to yield strong evidential support. One reason that the Bayes factor may not have been able to distinguish between the null and alternative hypotheses is that the study was not sufficiently powered to detect or rule out an effect of entropy. While the study had a sample size of 50—larger than many ERP studies in the literature—a small number of critical trials (10 per participant) likely led to a large amount of noise in the averaged ERP data.

Bayes factor evidence for the PNP was also inconclusive, but the effect size was larger and consistent in sign with previous findings on the effect of constraint (larger PNP for violations in high-constraint settings) [32, 33, 36]. This may suggest that lower entropy led to a prediction in the 1-particle condition, eliciting a higher suppression cost once this prediction was violated. Higher entropy in the 2+particle condition may have discouraged predictions, avoiding this cost. However, again, one could have expected stronger evidence from the Bayes factor. The findings would therefore also be consistent with a null effect in line with current PNP accounts suggesting that the component should not be sensitive to entropy if the input is implausible [33, 57].

One possible explanation for the weak Bayes factors is that variability was introduced into the data by individual differences in dealing with the implausible verb particles. Particles are more syntactically and semantically ambiguous than the implausible nouns used in previous PNP research and participants may have tried to reanalyse them in different ways, while others made no attempts at revision. For example, an implausible verb particle such as make [a story] *on could potentially be revised as a preposition make [a story [ on the war in Iraq]]…. The fact that a large posterior P600 was also elicited by the violations may indicate that such a syntactic reanalysis was attempted.

However, any conclusions from Experiment 1 are limited by a small number of critical trials, even though the sample size was relatively large for an ERP study. To test whether a larger study would provide more conclusive evidence for the PNP effect and against the N400 effect, we used the results of Experiment 1 to pre-register a second experiment with double the number of participants and more than double the number of experimental sentences.

Experiment 2

Experiment 2 sought to replicate the findings of Experiment 1 in a larger sample, using a larger number of experimental items. Since Experiment 1 already confirmed the presence of the expected ERP components at plausible vs. implausible particles, to further increase the probability of detecting an effect, we removed the plausible conditions to create a two-condition design. This meant that the implausible conditions could be split into two rather than four lists, doubling the number of critical trials each participant would see. In line with Experiment 1, we expected that we would not see a difference in N400 amplitude between 1-particle and 2+particle violations, but that the PNP might be larger in the 1-particle than the 2+particle condition. Experiment 2 was pre-registered at https://osf.io/y6k2d; see also S1 Appendix.