The effect of decay and lexical uncertainty on processing long-distance dependencies in reading

Lexical pre-activation in long-distance dependency formation

Contextual cues in a sentence are used to predictively pre-activate probable words and features in memory, such that processing of a predictable word can begin before that word is seen (Kuperberg & Jaeger, 2016; DeLong, Urbach & Kutas, 2005; Van Berkum et al., 2005; Wicha, Moreno & Kutas, 2004; Nicenboim, Vasishth & Rösler, 2020). Pre-activation therefore represents a processing advantage at predictable versus unpredictable words, as reflected by shorter reading times (Ehrlich & Rayner, 1981; Staub, 2015; Kliegl et al., 2004) and decreased event-related potential (ERP) components (Kutas & Hillyard, 1980, 1984; Kutas & Federmeier, 2011). It has also been proposed that strong pre-activation may trigger pre-integration of a specific lexical item into the building sentence representation in working memory (Ness & Meltzer-Asscher, 2018; Lewis & Vasishth, 2005; Vasishth & Lewis, 2006).

However, evidence for the pre-activation of lexical content in long-distance dependency formation is sparse. While there is evidence that specific lexical items are pre-activated by their context, pre-activation in such studies is generally only tested at the immediately preceding word or within the noun phrase (DeLong, Urbach & Kutas, 2005; Van Berkum et al., 2005; Wicha, Moreno & Kutas, 2004; Nicenboim, Vasishth & Rösler, 2020). To investigate longer distance dependency formation, researchers have demonstrated evidence that the left anterior negative (LAN) ERP component is larger at the initiation of long versus short syntactic wh-dependencies, suggesting that anticipation of a long dependency leads to greater working memory load (Fiebach, Schlesewsky & Friederici, 2002; Phillips, Kazanina & Abada, 2005). Applied to lexical pre-activation, a study of Dutch particle verbs hypothesised that verbs that take a large number of possible particles (e.g. “spannen”, to tense, which can take at least seven particles) should trigger pre-activation of those particles, placing a larger demand on working memory than verbs with a small set size (e.g. “kleuren”, to colour, which can take only two; Piai et al., 2013). When a verb-particle dependency is initiated by a verb that takes particles, the LAN should therefore be larger for large- versus small-set verbs. Instead, the authors observed that while the LAN was larger for verbs that took particles than those that did not, it did not differ between small and large set size. The authors concluded that the particles themselves were not pre-activated, but rather that readers anticipated the possibility of a downstream particle and stored the verb to facilitate its retrieval if a particle was encountered. Another plausible interpretation is that readers anticipated a particle and generated a syntactic prediction for its position, but not for its specific lexical identity. Together, this evidence suggests that readers pre-activate the syntactic structure of long-distance dependencies, but not long-distance lexical content.

Reading time studies have offered a different perspective on long-distance lexical pre-activation: complex predicate constructions in Hindi and Persian succeeded in eliciting a set size-type difference in reading times, which were faster at a target verb when a specific verb continuation was predictable than when no specific verb was predictable (Husain, Vasishth & Srinivasan, 2014; Safavi, Husain & Vasishth, 2016). Although these studies measured reading times at the target verb, the sentence stimuli in the Hindi study—including the target verb—were identical across conditions. Only the head noun differed, meaning that reading time differences at the target verb could reasonably be attributed to differences in pre-activation at the noun, rather than to differences in integrating the verb into different contexts. There is thus evidence that readers may pre-activate the lexical content of particle verb-type dependencies, although findings are inconsistent.

Delaying dependency resolution

Dependencies in English tend to be resolved relatively quickly (Futrell, Mahowald & Gibson, 2015), but this is often not the case in languages such as Dutch, Hindi, Persian, and German. This means that if dependent lexical content is pre-activated, pre-activation must be sustained over a potentially large amount of intervening sentence material. Processing of the intervening sentence material can have either a facilitatory or a hindering effect on processing of the dependency, as proposed by different theoretical accounts.

A hindering effect of delaying dependency resolution is predicted by accounts suggesting that processing intervening sentence material places a larger demand on working memory. The introduction of new discourse referents in particular has been associated with a locality effect in dependency processing, where reading of the distant word becomes slower the more new discourse referents are introduced. Slowed reading is proposed to reflect the cost of storing and integrating the new referents (Gibson, 1998, 2000), retrieval interference (Lewis & Vasishth, 2005; Vasishth & Lewis, 2006), and/or decay of constituent activation over time (Gibson, 1998, 2000; Lewis & Vasishth, 2005; Vasishth & Lewis, 2006; Vosse & Kempen, 2000), all contributing to longer retrieval time at the distant word.

A facilitatory effect of delaying dependency resolution may occur when the additional sentence material provides additional information as to the position and the identity of the distant word. This results in easier processing of the distant word, as reflected in faster reading times, otherwise known as an antilocality effect (Vasishth & Lewis, 2006). The facilitatory effect of increasing distance is captured by surprisal theory. Surprisal theory provides an information theoretic account of the difficulty of processing each new word in a sentence, represented by the negative log probability of that word appearing given the preceding context (Levy, 2008; Hale, 2001). According to surprisal theory, the building context of a sentence generates a set of licensed continuations. Each new word encountered triggers an update of the probability distribution of these continuations, and the degree of update is proportional to the difficulty of processing the new word; that is the greater the update, the greater the processing difficulty or ‘surprisal’. In broader terms, this means the more constraining a sentence is, the fewer likely possible continuations it will have, meaning lower surprisal and easier processing at a predictable word. Conversely, at an unpredictable word, surprisal and thus processing difficulty will be higher. Thus, surprisal theory predicts that the greater the amount of information separating two dependent words, the more predictable and easy to process the distant word will become.

The sources underlying antilocality and locality effects—predictability and working memory load respectively—may even interact. There is some evidence that the negative effect of high working memory load may only be apparent in weakly predictive contexts and that otherwise, antilocality effects are observed (Husain, Vasishth & Srinivasan, 2014; Konieczny, 2000; Levy & Keller, 2013). For example in German, it was found that reading times at the clause-final verb of a relative clause were faster when the verb was delayed by one additional constituent than when it was not delayed (an antilocality effect), but that reading times slowed down when the verb was delayed by two additional constituents (a locality effect; Levy & Keller, 2013). The authors reasoned that the relative infrequency of adding the second constituent (according to a corpus analysis) actually reduced predictability, making the effects of increased working memory load more pronounced. Casting doubt on these results, however, is a replication attempt finding only locality effects, regardless of what information preceded the verb (Vasishth et al., 2018).

More direct tests of an interaction between predictability and working memory load have been conducted in Hindi and Persian. In Hindi, increasing the separation within noun-verb complex predicates facilitated the reading of highly predictable verbs, but slowed the reading of low-predictable verbs, suggesting that high predictability outweighed the effect of additional working memory load introduced by the intervening sentence material (Husain, Vasishth & Srinivasan, 2014). However, this load/predictability interaction was not replicated in analogous constructions in Persian, where higher working memory load induced by additional sentence material slowed reading of the distant verb, regardless of the verb’s predictability (Safavi, Husain & Vasishth, 2016). One difference between the Hindi and Persian studies was the type of information used to manipulate the separation distance of the complex predicate dependencies. The Persian study used a relative clause and a prepositional phrase as an intervener (Safavi, Husain & Vasishth, 2016). Both relative clauses and prepositional phrases introduce new discourse referents and interference, both of which are predicted to burden working memory resources and slow reading (Gibson, 1998, 2000; Lewis & Vasishth, 2005), although new discourse referents may not be the only source of slowing in longer dependencies (Gibson & Wu, 2013). In comparison, the separation in the Hindi experiments was increased with adverbials, which instead may have increased evidence for the position and lexical identity of the upcoming verb (Hale, 2001; Levy, 2008). Altogether, these findings suggest that while readers may pre-activate the lexical entry of an upcoming dependent word, if appearance of that word is delayed, its predictability may play an important role in how the intervening information impacts processing.

The current experiments

We tested the decay/predictability interaction using German particle verbs, which are complex predicates similar to the constructions used in previous studies of Hindi and Persian (Husain, Vasishth & Srinivasan, 2014; Safavi, Husain & Vasishth, 2016). German particle verbs are comparable to English particle verbs in that they are composed of a base verb (e.g. “räumen”, to tidy) and a particle (e.g. “auf”, up) which can be separated (Müller, 2002). In German, however, the particle must appear after the direct object if the verb is transitive, usually at the right clause boundary (e.g. “Er räumte den Raum auf” he tidied the room up, but not “*Er räumte auf den Raum” he tidied up the room; Müller, 2002). Particle verbs form a very strong dependency because the full meaning of the verb “aufräumen” (to tidy up) can only be interpreted once both the verb and particle are known. Delaying appearance of the particle therefore creates a very strong structural expectation if the context makes a particle necessary, but potentially also a strong lexical expectation for a specific particle. In English particle verb constructions, the delay between a base verb and its particle is usually not very long; consider to tidy up vs ?/* to tidy the mess left after the party on Saturday up. In German, however, long-distance separations are common.

To manipulate lexical predictability of the distant particle, we compared base verbs that could take a large number of particles (10+) with verbs that can take only a small number of particles (six or fewer). We hypothesised that the set of potential particles would be pre-activated at the verb and that a larger set of particles would create more uncertainty (weaker predictability) about the eventual identity of the particle. Large set verbs therefore formed a low predictability condition and small set verbs a high predictability condition. Note that throughout the remainder of the article, we use set size as a proxy for predictability. Set size also relates to entropy, which we introduce in detail as it becomes relevant in the Cloze Test section. To induce decay between the verb and its particle, we manipulated distance with a neutral adjectival modifier. Critically, the modifier added no interference or working memory load through the introduction of new discourse referents (Gibson, 1998, 2000; Lewis & Vasishth, 2005), and did not provide semantic clues about the lexical identity of the dependency resolution. Any effects of the intervener on reading time were therefore attributable to temporal decay alone.

The design was based on the study of Dutch particle verbs (Piai et al., 2013). The Dutch study found no evidence of a modulation of LAN amplitude according to set size. We reasoned, however, that the distinction between small and large particle set sizes may have been too small: small set verbs took two to three particles and large set verbs, at least five. We therefore categorised our German verbs into small set verbs that took up to six particles, and large set verbs that took at least 10 particles. The current experiments therefore tested the hypotheses that (i) verbs that take particles trigger pre-activation of those particles; (ii) that delaying the appearance of the particle would slow reading times through temporal decay; but that (iii) higher predictability would make reading times at the particle less likely to be affected by decay.

We tested the hypotheses in self-paced reading and eye tracking experiments, both to confirm that any effects seen were not limited to a particular experimental method, but also because the two methods provide complementary information. Self-paced reading has the advantage of forcing readers to view each word in the sentence, whereas eye tracking allows words to be skipped and re-read. In the current study, the target word, a particle, was very short and may therefore have been more likely to be skipped, making self-paced reading data valuable in examining reading time effects at the particle. On the other hand, eye tracking has the advantage of more closely resembling natural reading and is able to measure phenomena such as regressive eye movements to previous regions of the sentence, and forward saccades to upcoming regions of the sentence. This allows us to generate hypotheses about the cognitive processes underlying slower or faster reading of a particular word and complements observations made in self-paced reading.

Predictions

It is well-established that more predictable words are associated with faster reading times than less predictable words, and thus we expected to see faster reading times for small set (more predictable) versus large set (less predictable) particles. With respect to our two distance conditions (short vs long), at short distance the predictions of surprisal theory and decay are the same: small set (more predictable) particles should be read faster than large set (less predictable) particles. This is reflected in both panels of Fig. 1, where predicted reading times for small set particles are always faster than those for large set particles.

Where the predictions of surprisal theory and decay diverge is in the long-distance condition. Under surprisal theory, the long-distance condition should produce an antilocality effect (faster reading times) at both small set and large set particles, as illustrated in Fig. 1A. We attempted to quantify these predictions by computing surprisal values for the particles; however, the particular particle verbs used in the experiment were likely too infrequent in the corpora used and the parser’s surprisal estimates were unreliable.¹ Instead, Fig. 1A represents informal predictions for the surprisal account. In the absence of formal quantifications for whether surprisal theory would predict an antilocality effect for our sentences, these predictions should be taken as an approximation of surprisal theory’s general claim that long distance should always result in faster reading times and that higher lexical predictability should sharpen expectations (Levy, 2008).

In contrast, under a decay account, reading times in the long-distance conditions should depend on how predictable the particle is. For more predictable (small set particles), pre-activation should be stronger to begin with and thus less affected by decay at long distance, whereas weaker pre-activation for less predictable (large set) particles may be more susceptible to decay, resulting in a locality effect (slower reading times) at long versus short distance. To quantify the effect of decay on reading time, we conducted a simulation using the decay parameter of the LV05 model (Lewis & Vasishth, 2005). Note that the full LV05 model was not used as it is primarily a model of interference, which we were not testing in the current study. To quantify predictability in the simulation, we assumed a finite pool of spreading activation for all of the plausible particle continuations. Dividing the finite pool of spreading activation among fewer particles meant a higher starting activation per particle in the small set than in the large set condition. Figure 1B shows that the simulation predicted a larger slow-down between small and large set size in the long distance condition than in the short distance condition. Code for the simulation is included in the R script in the paper’s OSF repository, see “Appendix 1”.

Methods

Cloze test

In order to confirm that our sentence stimuli (i) elicited particles, (ii) that more particles were elicited by the large set condition than the small set condition, and to (iii) quantify the predictability of the target particle, a cloze test was conducted. An initial total of 48 items, each with four conditions (a–d), was truncated just before the particle such that the verb and the direct object of the sentence were known. German native speakers completed the truncated sentences in a paper-and-pencil cloze test (N = 126, 25 male, mean age 25 years, standard deviation 7 years, range 17–53 years). The 48 sentences were split into four lists such that each participant saw only one condition from every item. The target sentences were randomly interspersed with 63 filler sentences, giving a total of 111 sentences per cloze test. Participants were instructed to complete each truncated sentence with the word or words that first came to mind.

The results of the cloze test yielded 24 items that achieved the required experimental manipulation: a particle was always elicited and more particles were elicited in the large than in the small set condition. It should be noted that in 8% of the stimuli, the highest cloze particle was not used as the target particle. This was because the target particle had to be matched across conditions and the highest cloze particle in one condition was therefore not always the highest cloze particle in another condition. Wherever possible, however, the highest cloze particle was used. Means and 95% confidence intervals of Beta distributions corresponding to the cloze probabilities for each factor level are presented in Table 1.

Table 1:

Summary cloze statistics for the final set of 24 items.

Condition	Cloze probability		Entropy
	Mean	95% CI	Mean	95% CI
Small set	0.51	[0.28, 0.73]	1.10	[1.09, 1.12]
Large set	0.55	[0.35, 0.75]	1.20	[1.19, 1.22]
Short distance	0.52	[0.31, 0.73]	1.15	[1.14, 1.16]
Long distance	0.53	[0.32, 0.75]	1.15	[1.13, 1.16]

DOI: 10.7717/peerj.10438/table-1

Note:

The 95% CIs reflect confidence intervals of each cloze probability distribution.

Cloze probabilities provided a measure of how predictable the target particles in each condition were. To determine whether the cloze probability of the particle differed between small and large set conditions, a logistic mixed model was fit in brms (Buerkner, 2017) in R (R Development Core Team, 2018) to the cloze probabilities of the target particles, with factor levels contrast coded as follows: small set −0.5/large set 0.5, short distance −0.5/long distance 0.5. The brms zero/one inflated Beta family was used for the likelihood to account for the presence of 0s and 1s in the data. Regularising priors were selected for each of the predictors set size, distance, and their interaction: β ∼ Normal(0, 0.25). The full prior and model specification can be found in the code provided, see “Appendix 1”. The model did not suggest that either set size, distance, or an interaction of the two influenced cloze probability. As can be seen in Fig. 2, the posteriors for the probability of giving the target particle were more or less centred on zero, meaning that neither set size, distance, or their interaction made people any more or less likely to give the target particle.

The set size manipulation was intended to induce uncertainty about the upcoming particle’s lexical identity; the higher the uncertainty, the less predictable the particle. One useful way of quantifying uncertainty is with entropy. Entropy is a measure of how much information is carried by a new input in light of all possible outcomes.² In our case, the new input is the particle. In a sentence context where many particles are plausible and cloze probability is uniformly low across all the plausible particles, we assume that uncertainty about the identity of the upcoming particle is high. Thus, each of the plausible particles carries a large amount of information about the meaning of the sentence and entropy is high. In a sentence where only few particles are plausible and one particle is much more probable than the others, we assume that uncertainty about that particle’s identity and the meaning of the sentence is low, and so encountering the high-probability particle will be less informative: this is a low entropy situation.

To determine whether uncertainty (and thus entropy) was higher in the large set condition, a lognormal regression model was fitted to the entropy values with the same contrast coding as for the cloze probability analysis. The brms hurdle lognormal family was used for the likelihood function to account for zeros in the data. Regularising priors were used for the predictors set size, distance, and their interaction: β ∼ Normal(0, 0.01). This model did not suggest that entropy varied with set size, distance, or their interaction, as can be seen in Fig. 2, although the mean entropy was a little higher in the large than the small set condition.

This analysis raised an immediate problem with the experimental design. The categorical predictor set size used in the planned analysis was intended as a proxy for entropy and predictability, where a large set size was supposed to reflect high entropy and thus lower predictability. Although these categories may have reflected the number of particles licensed by each base verb, the results of the cloze test suggested they did not represent the range of particle completions provided by readers at the particle site. This can be seen in Fig. 3: although the average entropy was higher in the large set than in the small set condition, both conditions contained high and low entropy sentences. In other words, there was no difference in predictability of the particle between the small and large set conditions. We therefore present an analysis of entropy as a continuous predictor instead, since this maps better to our planned manipulation of predictability (high entropy = low predictability and vice versa). For transparency, we present both the planned “categorical” analysis and the exploratory “continuous” analysis.

Question response accuracy and reaction times

Mean accuracy and reaction times to responses to comprehension questions in all four conditions are set out in Table 2.

Planned analysis

Exploratory analysis

Entropy as a continuous predictor

In an exploratory analysis, entropy at the particle was refitted as a continuous predictor and its effect on reading speed examined. Descriptive statistics for reading times in each distance condition are shown in Table 5. Mean reading times according to entropy have been split into high and low categories by median-split for summary purposes, but entropy was used as a continuous predictor in the statistical model.

Table 5:

Experiment 1: Summary self-paced reading times by condition using entropy as a continuous variable.

Condition	Reading time (ms)
	Mean	95% CI
(a) Low entropy, short distance	443	[420, 466]
(b) Low entropy, long distance	438	[416, 461]
(c) High entropy, short distance	433	[413, 455]
(d) High entropy, long distance	443	[422, 466]

DOI: 10.7717/peerj.10438/table-5

Note:

For the purpose of these summary statistics only, entropy was sorted into high and low categories via median-split. The mean and 95% confidence interval (CI) per condition are presented.

Mean reading times across the whole sentence for both experiments are plotted in Fig. 4. One feature of these data that should be mentioned is that base verbs for sentences with higher entropy at the particle site had a higher corpus frequency than base verbs in sentences with lower entropy at the particle site (to compare verb frequency, we divided sentences into high and low entropy categories via a median split; see Table A1 in “Appendix 2”). Higher corpus frequency of the base verb should have resulted in faster reading times at the verb in high entropy sentences (Kliegl et al., 2004; Rayner & Duffy, 1986), but this was not the case in either experiment. The lack of a frequency effect at the base verb is discussed in the “General Discussion”.

The priors and model specification remained the same as for the planned analysis. The model coefficients are summarised in Table 6. As can also be seen in Fig. 5, zero is well within the 95% credible interval for the posterior of the all predictors. The Bayes factor analysis found evidence for the null hypothesis for each of the predictors. In other words, there was evidence against an effect of entropy, distance, and their interaction on reading speed.

Table 6:

Experiment 1: Self-paced reading speed estimates for the exploratory analysis with entropy as a continuous predictor.

Predictor	$\hat{\beta }$ (words/s)	95% CrI	Bayes factors (BF10)
			Informative	Planned	Diffuse
Intercept	2.51	[2.32, 2.69]	–	–	–
Entropy	−0.04	[−0.13, 0.05]	0.51	0.14	0.07
Distance	−0.02	[−0.11, 0.07]	0.42	0.10	0.05
Entropy × distance	−0.02	[−0.15, 0.10]	0.52	0.05	0.01

DOI: 10.7717/peerj.10438/table-6

Note:

As for the planned analysis, the reciprocal transform means that $\hat{\beta }$ represents the model’s estimated effect for each of the predictors in words per second. A positive sign therefore indicates faster reading (more words per second) and a negative sign, slower reading. The 95% Bayesian credible interval (CrI) gives the range in which 95% of the model’s samples fell. Bayes factors are presented for a range of β priors including, from left to right: more informative than the prior used in the planned analysis, N(0, 0.1); the prior used in the planned analysis, N(0, 0.5); and more diffuse than the prior used in the planned analysis, N(0, 1). BF₁₀ indicates the Bayes factor for the full model (1) against a reduced model (0). Bayes factors of less than 0.3 indicate evidence for the reduced model, while Bayes factors greater than 3.0 suggest evidence for the full model.

Reading speed predicted by the model is plotted in Fig. 6. The numerical pattern suggests an interesting mix of the two hypotheses: when predictability was high (low entropy), reading speed was faster at long distance in line with the surprisal account. In contrast, when predictability was low (high entropy), the pattern more closely resembles that predicted by decay. However, these patterns are not further interpreted as the outcome of the statistical analysis did not support an interaction.

Interim discussion

Neither the planned nor the exploratory analyses were consistent with the predictions in Fig. 6. With respect to the planned (categorical) analysis, one potential explanation may lie in the very small differences in cloze probability and entropy at the particle site, meaning that entropy between set size conditions was effectively matched at that point in the sentence. Examples of entropy differences between condition means discussed elsewhere in the literature include 0.38 or 0.50 bits (Levy, 2008), 0.57 bits (Linzen & Jaeger, 2016), and reductions of up to 53 bits (Hale, 2006). In comparison, our between-category difference was only 0.10 bits. However, the examples given from the literature are derived from syntactic entropy of the rest of the sentence, while ours were based on lexical entropy at the particle. Nonetheless, while the small between-category difference in entropy may explain why we did not see a statistical difference in reading times between the large and small set categories, it does not explain why we still saw no difference when entropy was used as a continuous predictor. We turn now to the eye tracking results for further information.

Participants

Sixty German native speakers were recruited, of which one was excluded due to the presence of a neurological disorder. The remaining 59 (13 male) were free of current or developmental reading or language production disorders, hearing disorders, or vision impairments that could not be corrected without impeding the eye-tracker (e.g. glasses and contacts occasionally caused reflection preventing accurate calibration of the eye-tracker, meaning that these participants had to be excluded if they were unable to read without visual correction). The mean age of the participants was 26 (SD = 6, range = 18–47) and all were university educated. All participants provided written informed consent in accordance with the Declaration of Helsinki. In accordance with German law, IRB review was not required.

Materials

The experimental materials and presentation lists were identical to those used in the self-paced reading study.

Procedure

Right eye monocular tracking was conducted using an EyeLink 1000 eye-tracker (SR Research) with a desktop-mounted camera and a sampling rate of 1,000 Hz. The head was stabilised using a chin and forehead rest which set the eyes at a distance of approximately 66 cm from the presentation monitor. The experimental paradigm was built and presented using Experiment Builder (SR Research). The 22-inch presentation monitor had a screen resolution of 1680 × 1050. Sentences were presented in size 16-point Courier New font on a pale grey background (hex code #cccccc). Each experimental session began with calibration of the eye-tracker, which was repeated if necessary during the experiment. The experimental sentences were preceded by six practice sentences. Participants fixated on a dot at the centre left of the screen before each sentence was presented. Once they had finished reading, they fixated on a dot at the bottom right of the screen. Each of the experimental sentences was followed by the same yes/no question used in the self-paced reading study, which the participant answered using a gamepad. Each session lasted approximately 30 min.

Data analysis

Sampled data were exported from DataViewer (SR Research) and pre-processed in R using the em2 package (Logačev & Vasishth, 2013). Trials containing blinks or track loss were excluded. Linear mixed-effects models with full variance–covariance matrices estimated for the random effects of participant and item were fitted using brms (Buerkner, 2017) in R (R Development Core Team, 2018) separately to data for each of four reading time measures: first fixation duration, first pass reading time, total fixation time, and regression path duration. This range of measures was selected as both early and late measures have been found to be affected by predictability (Kliegl et al., 2004; Boston et al., 2008), although perhaps earlier measures are more sensitive (Staub, 2015). The target region of the sentence was the particle plus the immediately preceding word, since the particles were usually short (two to three letters) and therefore not always fixated. As for Experiment 1, the spillover region was not analysed, but mean reading times across the whole sentence are presented in Fig. 4. The preceding rather than the following word was chosen because the target particle was at the right clause boundary. The dependent variables were first fixation duration, first pass reading time, total fixation time, and regression path duration at the particle, log transformed as indicated by the Box Cox procedure. The predictors set size and distance were effect contrast coded: −0.5 (small set/short distance), 0.5 (large set/long distance). The model priors were as follows:

• β₀ ∼ Normal(5.7, 0.5)

• β_1,2,3 ∼ Normal(0, 0.5)

• υ ∼ Normal(0, σ_υ)

• γ ∼ Normal(0, σ_γ)

• σ_υ, σ_γ ∼ Normal₊ (0, 1)

• ρ_υ, ρ_γ ∼ LKJ(2)

• σ ∼ Normal₊ (0, 1)

The prior distribution of the intercept was determined using domain knowledge that mean reading time is approximately 300 ms (5.7 on the log scale) and that 95% of reading times should fall within a range of 110 and 812 ms. We expected the effect of the predictors would mostly lie somewhere between a speed-up of 190 ms and a slow-down of 513 ms. Priors for the random effects parameters were as shown above. The full model specification can be found in the code in the accompanying code, see “Appendix 1”.

Question response accuracy and reaction times

Mean response accuracy and reaction times for the comprehension questions in all four conditions are set out in Table 7.

Planned analysis

Exploratory analyses

Entropy as a continuous predictor

As for the self-paced reading analysis, models were refit using entropy as a continuous predictor. Descriptive statistics for each reading time measure are shown in Table 10. Mean reading times according to entropy have been split into high and low categories by median-split for summary purposes, but entropy was used as a continuous predictor in the statistical model.

Table 10:

Experiment 2: Summary eye-tracking reading times by condition using entropy as a continuous variable.

Measure	Condition	Reading time (ms)
		Mean	95% CI
First fixation duration
	(a) Low entropy, short distance	279	[265, 295]
	(b) Low entropy, long distance	264	[250, 279]
	(c) High entropy, short distance	293	[277, 311]
	(d) High entropy, long distance	317	[299, 335]
First pass reading time
	(a) Low entropy, short distance	317	[297, 338]
	(b) Low entropy, long distance	287	[270, 306]
	(c) High entropy, short distance	321	[300, 343]
	(d) High entropy, long distance	357	[334, 381]
Total fixation time
	(a) Low entropy, short distance	357	[332, 385]
	(b) Low entropy, long distance	321	[299, 346]
	(c) High entropy, short distance	376	[348, 407]
	(d) High entropy, long distance	416	[385, 449]
Regression path duration
	(a) Low entropy, short distance	354	[329, 382]
	(b) Low entropy, long distance	325	[301, 351]
	(c) High entropy, short distance	358	[332, 386]
	(d) High entropy, long distance	402	[373, 433]

DOI: 10.7717/peerj.10438/table-10

Note:

For the purpose of these summary statistics only, entropy was sorted into high and low categories via median-split. The mean and 95% confidence interval (CI) per condition are presented

The model estimates can be seen in Table 11 and the model posteriors in Fig. 7. The Bayes factor analysis found evidence for an effect of entropy on first fixation duration, first pass reading time, and total fixation time, in that increasing entropy slowed reading times. With more informative priors, Bayes factors suggested evidence for the effect of entropy in each of these three measures was strong. At the planned (non-informative, regularising) prior for regression path duration, Bayes factor evidence for an effect of entropy was inconclusive. However, when the more informative prior was used, evidence for an effect of entropy on regression path duration was strong. The Bayes factors for the remaining predictors (distance, entropy × distance) were in favour of the null hypothesis, regardless of which prior was used.

Table 11:

Experiment 2: Model estimates for the exploratory analysis with entropy as a continuous predictor.

Measure	Predictor	$\hat{\beta }$ (log ms)	95% CrI	Bayes factors (BF10)
				Informative	Planned	Diffuse
First fixation duration
	Intercept	5.66	[5.55, 5.76]	–	–	–
	Entropy	0.08	[0.03, 0.13]	23.88	4.65	2.15
	Distance	0.01	[−0.05, 0.07]	0.28	0.06	0.03
	Entropy × distance	0.04	[−0.04, 0.11]	0.32	0.01	0.00
First pass reading time
	Intercept	5.76	[5.61, 5.90]	–	–	–
	Entropy	0.08	[0.03, 0.13]	17.71	4.49	1.86
	Distance	0.00	[−0.06, 0.07]	0.27	0.06	0.03
	Entropy × distance	0.02	[−0.06, 0.10	0.19	0.00	0.00
Total fixation time
	Intercept	5.87	[5.70, 6.04]	–	–	–
	Entropy	0.12	[0.04, 0.21]	24.65	4.77	2.78
	Distance	0.00	[−0.06, 0.07]	0.32	0.07	0.04
	Entropy × distance	0.01	[−0.08, 0.09]	0.22	0.00	0.00
Regression path duration
	Intercept	5.85	[5.67, 6.02]	–	–	–
	Entropy	0.10	[0.03, 0.18]	12.58	2.91	1.18
	Distance	0.01	[−0.05, 0.08]	0.35	0.07	0.03
	Entropy × distance	0.04	[−0.06, 0.12]	0.41	0.01	0.00

DOI: 10.7717/peerj.10438/table-11

Note:

$\hat{\beta }$ represents the model’s estimated effect for each of the predictors on the log scale. The log transform means that estimates with a positive sign indicate slower reading times and that readers who are slower on average will be more affected by the manipulation than faster readers. The 95% Bayesian credible interval (CrI) gives the range in which 95% of the model’s samples fell. Bayes factors are presented for a range of β priors including, from left to right: more informative than the prior used in the planned analysis, N(0, 0.1); the prior used in the planned analysis, N(0, 0.5); and more diffuse than the prior used in the planned analysis, N(0, 1). BF₁₀ indicates the Bayes factor for the full model (1) against a reduced model (0). Bayes factors of less than 0.3 indicate evidence for the reduced model, while Bayes factors greater than 3.0 suggest evidence for the full model.

The predicted vs observed interactions of distance and entropy are plotted in Fig. 8. Numerically, the pattern of reading times again appeared to be a mixture of the predictions of surprisal theory and the decay simulation based on the LV05 model. However, the results of the statistical analyses did not support an interaction of entropy and distance, and so this pattern is not further interpreted.

Interim discussion

The planned analysis with the categorical predictor set size again did not find any support for our hypotheses that temporal activation decay would be more prominent when lexical predictability was low. Reconfiguring set size as the continuous predictor entropy, however, found support for the hypothesis that increased uncertainty about the lexical identity of the particle would slow reading times. However, there was still no evidence that temporal decay influenced reading times, either alone or in interaction with entropy.

When was the particle pre-activated?

Within each experimental item, all words were identical except for the verb, meaning that the only information influencing uncertainty at the particle site was the verb. This supports the possibility that the difference in reading time observed at the particle could have resulted from differences in particle pre-activation at the verb. However, it is also possible that pre-activation was triggered by the combination of the verb and its direct objects. For example the fragment ‘Nach dem Gespräch stellte er die Kandidatin…’ (Following the interview, he put the candidate…) should be sufficient to anticipate the most likely verb-particle combinations. The lexical pre-activation of particles is unlikely to have been triggered by information between the direct object and the particle site (e.g. ‘aus England’, from England), since this region did not add any information about the identity of the particle. It is therefore possible to conclude that pre-activation occurred at the latest before the pre-critical region, suggesting that lexical pre-activation can be sustained over multiple intervening words that do not form part of the particle verb constituent (cf. studies where evidence for lexical pre-activation is only observed at the immediately preceding word or within the noun phrase: DeLong, Urbach & Kutas, 2005; Van Berkum et al., 2005; Wicha, Moreno & Kutas, 2004; Nicenboim, Vasishth & Rösler, 2020).

One feature of interest in the data, and perhaps in further support of particle pre-activation at the verb, is the fact that base verbs associated with higher entropy at the particle were higher in frequency, and yet were not read faster. High word frequency is strongly correlated with faster reading time (Kliegl et al., 2004; Rayner & Duffy, 1986). A potential explanation for the lack of a speed-up is that a larger number of pre-activated particles made the meaning of the verb more ambiguous, which in turn led to slower reading and cancelling out of the expected speed-up associated with higher frequency. This hypothesis requires testing, however.

Assuming that particle pre-activation underlies the effects observed in eye-tracking, our findings present a contradiction to the hypothesis that verbs that take particles are maintained in working memory to facilitate retrieval once the particle is finally encountered (Piai et al., 2013). If this were the case, we should not have observed an effect of predictability at the particle, since there is no reason to think that one verb, already activated and integrated into the sentence parse, should have required more resources to retrieve than another. It may indeed be that high entropy verbs are somehow more difficult to integrate than low entropy verbs, but it is difficult to conceive of why without invoking activation of associated lexical or syntactic information, including particles. Maintenance of the verb in working memory therefore does not account for the eye-tracking results observed reported here.

Temporal activation decay

The evidence against an effect of temporal decay in both the self-paced reading and eye tracking experiments is consistent with findings suggesting that decay is not an important factor influencing reading and memory recall times (Lewandowsky, Oberauer & Brown, 2009; Engelmann, Jäger & Vasishth, 2019; Vasishth et al., 2019). In comparison to the sentences used in distance manipulations in previous studies, our sentences used simple adjectival modifiers that deliberately avoided the introduction of interference or new discourse referents. This allowed us to isolate decay as an explanatory factor; however, it is possible that the modifiers were not long enough to introduce a detectable effect of decay. That said, it would have been difficult to construct longer interveners without reintroducing interference or working memory load, which supports the idea that interference and working memory load are indeed the more important source of processing difficulty in longer sentences, rather than temporal decay. Alternatively, it could be argued that the difficulty in constructing longer sentences without introducing interference or working memory load means it is difficult or impossible to test decay in isolation, and thus that we cannot know what the true effect of decay is. However, if the effect of decay is so small that it is undetectable in the face of interference and working memory load, and these factors are almost unavoidable in constructing long dependencies, then one could argue that decay does not play a major role in processing difficulty.

Another possible explanation for not having detected a decay effect is that the difficulty in creating experimental items meant there were only 24 experimental items in total. In the Latin square design, this meant that each participant saw only six target trials per condition. If the effect of decay is indeed very small, future experiments should include more trials per participant in order to detect the effect.

Data and code

All data and code necessary to reproduce our analyses are available here: https://osf.io/yg5wx/.

Particle verb frequencies

Frequencies were computed for both the base verb and the particle verb as a whole using the Tübingen aNotated Data Retrieval Application, TüNDRA (Martens, 2013). The treebank used was the automatic dependency parse of the German Wikipedia with over 48.26 million sentences. Frequencies are presented as the incidence of the verb or particle verb per 1,000 words. As can be seen in Table A1, while the frequencies of the verb+particle constructions were comparable, frequency of the base verb was notably higher in the high entropy condition.