The interaction of predictive processing and similarity-based retrieval interference: an ERP study

ABSTRACT

Language processing requires memory retrieval to integrate current input with previous context and making predictions about upcoming input. We propose that prediction and retrieval are two sides of the same coin, i.e. functionally the same, as they both activate memory representations. Under this assumption, memory retrieval and prediction should interact: Retrieval interference can only occur at a word that triggers retrieval and a fully predicted word would not do that. The present study investigated the proposed interaction with event-related potentials (ERPs) during the processing of sentence pairs in German. Predictability was measured via cloze probability. Memory retrieval was manipulated via the position of a distractor inducing proactive or retroactive similarity-based interference. Linear mixed model analyses provided evidence for the hypothesised interaction in a broadly distributed negativity, which we discuss in relation to the interference ERP literature. Our finding supports the proposal that memory retrieval and prediction are functionally the same.

KEYWORDS:

• Language
• memory retrieval
• interference
• prediction
• predictive processing
• interaction
• ERP

Introduction

Two phenomena have been identified as influencing factors during sentence processing: prediction and similarity-based interference. While predictable input is easier to process (see e.g. Federmeier & Kutas, 1999; Konieczny, 2000), interfering distractors during memory retrieval can lead to difficulty (see e.g. Jäger et al., 2017; Lewis et al., 2006; Van Dyke & McElree, 2006). The combined investigation of memory retrieval, indicated by interference effects, and prediction is interesting because they correspond to contrary streams of processing: While interference arises when bottom-up information calls for a retrieval of previously encountered information, prediction can be defined as pre-activation of upcoming information due to top-down activity. Here, we propose that memory retrieval and predictive pre-activation are two sides of the same coin, as they are functionally the same: Both activate memory representations. This proposal is based on the interaction of memory retrieval and prediction. Behavioural studies have found that high predictability cancels memory-induced processing difficulty (Campanelli et al., 2018; Husain et al., 2014). We present – to the best of our knowledge – the first event-related potentials (ERP) study that investigated whether interference effects can only arise at unpredicted words. Highly predicted words should be pre-activated completely, which would render retrievals redundant. Without a retrieval, there can be no retrieval interference. In the remainder of the introduction, we introduce interference, prediction, and their interaction before introducing the present study.

Memory retrieval and interference effects during sentence processing

To successfully process a sentence or discourse, it is necessary to access previously encountered words, so that new words can be integrated with ones that have already been processed. The cognitive mechanisms involved in this process have been described in detail in models of cue-based parsing (Lewis et al., 2006; McElree, 2000; Van Dyke & Lewis, 2003; Vasishth et al., 2019, for detailed discussions of the models and their differences, see Lissón et al., 2021; Nicenboim & Vasishth, 2018). The cue-based parsing framework presents memory retrieval as a cue-based process, which compares retrieval cues generated at the retrieval site with the features of the words in memory. Each new word or constituent is encoded into memory as a feature bundle (e.g. case, number, gender, etc.). Importantly, not only the constituent itself is represented, but also predicted constituents that are required to form a grammatical sentence. When the noun toy is encountered in (1) (from Lewis et al., 2006), two memory representations are constructed: one depicting its features and the other a prediction of the corresponding predicate, i.e. a verb.

1. Melissa knew that the toy from her uncle in Bogotá arrived today.

When the verb arrived is processed, retrieval cues are generated to integrate it into the sentence context. The cues are matched in parallel against all (recent) memory representations. A sufficient match results in retrieval of the predicate prediction and the integration of verb and noun. Retrieval can be impaired by similarity-based interference from cue overlap, i.e. an overlap between the retrieval cues and features of more than one word or prediction recently stored in memory (e.g. Lewis et al., 2006; Nairne, 2002; Van Dyke & McElree, 2011). The correct to-be-retrieved word/prediction is called the target of the retrieval and other words/predictions in memory may function as interfering distractors during the retrieval process.

Numerous behavioural studies have provided evidence for similarity-based interference (see Jäger et al., 2017 for a review and meta-analysis). For example, Van Dyke and McElree (2006) used a dual-task paradigm to study interference from extra-sentential memory items during sentence comprehension. They presented it-clefts that required a retrieval operation to integrate the matrix sentence verb with its object (see 2).

(2a)	table – sink – truck
(2b)	It was the boat that the guy who lived by the sea sailed / fixed in two sunny days.

In load conditions, three nouns functioning as distractors (see 2a) were presented before a self-paced sentence reading task (for sentences such as 2b). In high interference conditions, the distractors were plausible objects of the matrix verb, and in low interference conditions, they were implausible (i.e. it is possible to fix, but not to sail a table / sink / truck). No load conditions, in which participants read the sentences without an additional task, were included to ensure that there were no confounding differences in comprehension of the verbs. Reading times for the critical verb carrying the interference manipulation were longest for high interference conditions with memory load. These results were interpreted as evidence for processing difficulty due to similarity-based retrieval interference at the verb in the presence of distractor items (Van Dyke & McElree, 2006).

Van Dyke (2007) showed that distractors that match only a subset of retrieval cues could nevertheless influence retrieval and cause processing difficulties. Furthermore, it was found that distractors occurring between target and retrieval site (retroactive interference) have more impact on processing than distractors that are processed before the target (proactive interference, Martin & McElree, 2009).

Most of the work on similarity-based interference has been on English. Nevertheless, morphologically richer languages, such as German, are interesting to study in this context as they mark features such as number and gender explicitly. Nicenboim et al. (2018) found retrieval interference during sentence processing in German (see also Vasishth & Drenhaus, 2011; Pregla et al., 2021). The authors used a self-paced reading paradigm and found weak evidence for an interference effect during the retrieval of a subject in the presence versus absence of distractor nouns that shared the same number feature.

Only a few ERP studies have found effects of similarity-based retrieval interference. The first widely acknowledged ERP evidence¹ was presented by Martin et al. (2012), who studied how distractors affected the processing of noun phrase ellipses in Spanish (see 3). To successfully process the ellipsis, the antecedent camiseta (“t-shirt”) needs to be retrieved at the position of the determiner otra (“another”). The authors manipulated the (mis)match between the gender of the target (camiseta, “t-shirt”[feminine]), the distractor (falda, “skirt”[feminine] / vestido, “dress”[masculine]) and the retrieval cues generated at the retrieval site (otra, “another”[feminine] / otro, “another”[masculine]).

(3)	Marta se compró la camiseta que estaba al lado de la falda / del vestido y Miren cogió otra / otro […] para salir de fiesta. “Marta bought the t-shirt [feminine] that was next to the skirt [feminine] / dress [masculine] and Miren took another [feminine] / another [masculine] to go to the party.”

They reported a sustained, broadly distributed negativity for ungrammatical ellipses (camiseta [feminine] … otro [masculine]) compared to grammatical ellipses (camiseta [feminine] … otra [feminine]). Within grammatical ellipses, this effect was modulated by the gender of the distractor: A mismatching distractor (vestido, “dress”[masculine]) led to a more negative response compared to a matching distractor (falda, “skirt”[feminine]). It was concluded that in grammatical sentences, the retrieval of the target, i.e. the antecedent, was interrupted by the mismatching features of a mismatching distractor. In a follow-up study, Martin et al. (2014) changed the syntactic structure of their material to render the antecedents of the ellipses, i.e. the target of the retrieval, less distinct. In this study, grammatical ellipses with matching distractors and ungrammatical ellipses elicited an early anterior negativity compared to grammatical ellipses with mismatching distractors. This was regarded as an effect of similarity-based interference. Lee and Garnsey (2015) found a sustained frontal negativity at the verb during the resolution of a long distance subject-verb dependency when a highly interfering distractor was present. Tanner et al. (2017) found that in ungrammatical sentences, the P600 at the verb was reduced when a distractor led to an illusion of grammaticality. This inconsistent pattern of results across studies makes it clear that the neurocognitive correlates of interference require further investigation.

Prediction during sentence processing

The investigation of prediction during language and in particular sentence comprehension has been a major focus in psycholinguistics and neurolinguistics for decades. In ERP research, prediction and prediction errors have been studied predominantly by means of the N400 component (but see, for example, Van Petten & Luka, 2012, for a different view), a negative deflection that peaks around 400 ms post stimulus onset (e.g. Kutas & Federmeier, 2011; Kutas & Hillyard, 1980). It was found that the N400 amplitude is negatively correlated with the cloze probability of a word in a given sentence context (e.g. Kutas & Hillyard, 1980, 1984), a measure of predictability during sentence comprehension (Taylor, 1953). Furthermore, Federmeier and Kutas (1999) found that the N400 amplitude indexes the degree of mismatch between features that have been pre-activated and the actually encountered word.

Some researchers have investigated prediction effects before a prediction can be confirmed or falsified. Wicha et al. (2003) studied whether participants predicted the gender of a noun during sentence comprehension in Spanish and whether that prediction affected processing of the pre-nominal article (see 4). In (4), canasta (“basket”) is highly predictable from the context. In contrast, corona (“crown”) is not predictable (because it is semantically incongruent). The factors predictability and article (mis)match were fully crossed. The nouns were presented as line drawings within an otherwise typical word-by-word presentation paradigm.

(4)	Caperucita Roja lleveba la comida para su abuela en una / un canasta / corona muy bonita Pero el lobo llegó antes que ella. “Little Red Riding Hood carried the food for her grandmother in a[fem.] / a[masc.] basket[fem.] / crown[fem.] [very pretty]. But the wolf arrived before her.”

At the noun, no interaction between predictability and gender (mis)match with the article was found. Unpredicted nouns (the corresponding line drawings, respectively) elicited an increased N400 compared to predicted nouns/drawings. Drawings preceded by a mismatching article led to a more negative response than drawings that matched the article’s gender. An article that mismatched the gender of a predicted (but yet not seen) noun led to an N400 effect compared to an article that matched the gender of the predicted noun. These results were the first that indicated a prediction effect before the prediction was confirmed or falsified. In a follow-up study, Wicha et al. (2004) used words instead of line drawings as critical stimuli and found a late positivity for the articles instead of the N400 effect.

In a prominent study, DeLong et al. (2005) investigated article effects such as those observed by Wicha et al. (2003, 2004) in English. In sentences like (5), they found the expected N400 effect for unpredicted (airplane) vs. predicted nouns (kite). The same pattern was found for the corresponding articles: An N400 effect for an compared to a.

(5)	The day was breezy so the boy went out to fly a kite / an airplane.

A large-scale replication attempt of the DeLong et al. (2005) results failed (Nieuwland et al., 2018). This again raised the question of whether and to what extent predictive processing is employed during sentence comprehension (but for findings supporting predictive processing at pre-nominal positions, see e.g. Fleur et al., 2020; Freunberger & Roehm, 2017; Hosemann et al., 2013; Nicenboim et al., 2020; Szewczyk & Schriefers, 2013). Additionally, in contrast to the N400 effects for unpredicted stimuli, P300 effects can arise for highly predicted stimuli in highly constraining contexts such as The opposite of black is … (Roehm et al., 2007). The highly predictable word white in the above context elicited a pronounced positivity around 300 ms post stimulus onset compared to words that violated that prediction (yellow, nice; for similar results, see Kutas & Iragui, 1998).

The interaction of memory retrieval and predictive processing

The interplay of memory-based difficulty and prediction-based facilitation during language comprehension has primarily been of interest to researchers in the context of the discussion of locality and anti-locality effects.² Locality effects arise due to increased processing difficulty with increased distance between dependent constituents (see e.g. Gibson, 2000). Anti-locality effects are a processing facilitation at the second element of a dependency due to its increased predictability with increasing distance, mainly seen in head-final languages (e.g. Konieczny, 2000; Vasishth & Lewis, 2006).

However, it was shown that even in head-final languages increased processing costs could turn anti-locality effects into locality effects. Vasishth and Drenhaus (2011) conducted three experiments in German, which examined argument-verb integration within relative clauses. Reading times for verbs obtained by self-paced reading and eye-tracking increased as the distance between verb and argument increased. In an ERP study using the same design, Vasishth and Drenhaus (2011) found a left localised fronto-central negativity for longer compared to shorter verb-argument distance. The authors linked this negativity to the Left Anterior Negativity (LAN) as a marker of increased memory load during long-distance dependency resolution. Vasishth and Drenhaus (2011) explained the observed locality effects with similarity-based retrieval interference. To increase the distance between verb and argument, they had introduced additional noun phrases. These noun phrases were distractors during the retrieval. It can be concluded that Vasishth and Drenhaus (2011) provided the first evidence of similarity-based retrieval interference in German using self-paced reading, eye-tracking and ERPs. However, they also observed reading time effects before the verb. They speculated that the argument and verb could have been integrated before the verb was actually encountered. Therefore, Vasishth and Drenhaus (2011) were perhaps the first to formulate the idea that pre-activation could bring forward later processing steps, which is highly relevant for the present proposal.

Further evidence for locality effects in a head-final language came from Levy and Keller’s (2013) Experiment 2. They showed that syntactic complexity of the intervening material between argument and verb resulted in locality effects in German (but see Vasishth et al., 2018, for a Bayesian reanalysis and failed replication attempts of Levy & Keller, 2013). Similarly, Husain et al. (2014) showed that in Hindi, anti-locality effects only arise when there is a strong prediction for a specific word to occur as the second element of a dependency. When it was only possible to predict the word class of the upcoming word, locality effects were found. Husain et al. (2014) interpreted this result to indicate that “[…] the specific verb is so predictable in a complex predicate, the verb could already be integrated into the predicted VP when the nominal host is processed.” (Husain et al., 2014, p. 12).

Campanelli et al. (2018) provided further support for this notion. They extended the Van Dyke and McElree (2006) dual-task interference paradigm (see 2, above), in which memory cost is manipulated via the match between retrieval cues generated at the critical verb and extra-sentential distractors. An additional predictability manipulation varied the predictability of the critical word via semantic context. In sentences like (6b), the predictability of the critical matrix sentence verb performed / created was manipulated by varying the subject noun person / choreographer. The authors measured the predictability of the verb via its cloze probability.

(6a)	website – handbag – password
(6b)	It was the dance that the person / choreographer who lived in the city performed / created early last month.

No effects were found at the critical verb. The sentence-final spillover region early last month was read faster in load than no load conditions. Within the no load conditions, high predictability led to faster reading times than low predictability. Within the load conditions, a marginal effect of interference and an effect of prediction were found. Residual log reading times were fastest in interference conditions with high predictability and slowest in interference conditions with low predictability of the critical verb. This result provided further converging support for the notion that memory retrieval is unnecessary at highly predictable words and thus similarity-based interference effects cannot arise. As Campanelli et al. (2018) put it: “If we think of memory retrieval as a gradual accumulation of information in the focus of attention, expectation would exercise its effect via an advance accumulation of evidence before the retrieval is initiated. Such a head start for selection of information would reduce retrieval interference by boosting the availability of the target word relative to its competitors.” (p. 1440). Taking this idea even further and in a similar vein to Vasishth and Drenhaus (2011) and Husain et al. (2014), we assume that a highly predicted word is already retrieved and integrated at the site where that prediction was generated. Consequently, we propose that at a highly predicted, i.e. fully pre-activated, word triggers no memory retrieval. The present study aimed to provide the first ERP evidence for this claim.³

The present study

The present study examined the interplay between prediction and memory retrieval – or more precisely similarity-based retrieval interference – in German. The study had two aims: A) Provide further evidence for interference effects in German; B) Provide initial ERP evidence that pre-activation can prevent memory retrieval. We manipulated gender interference and predictability during the processing of an article (see 7). In German, articles are overtly marked for case, number and gender. Therefore, they are highly suited to manipulate interference and investigate pre-nominal prediction effects. Interference was represented by two fully crossed factors: interference type (retroactive, proactive) and interference degree (high, low).⁴ Prediction, i.e. predictability, was measured via cloze probability obtained by a sentence completion pretest.

(7)	Im Glas liegt eine Raupe und in der Schachtel sitzt ein Käfer. Peter befreit den Käfer aus der Schachtel. “In the glas there lies a caterpillar [feminine] and in the box there sits a beetle [masculine]. Peter frees the [masculine] beetle [masculine] from the box.”

Two nouns, the target Käfer (“beetle”[masculine]) and the distractor Raupe (“caterpillar”[feminine]) were introduced in a context sentence. In the second sentence, the target was repeated with a definite article (as it was already introduced). We used two sentences as stimuli because we did not want to construct complex sentences including multiple clauses to introduce target and distractor and initiate the retrieval. We hypothesised that the processing of the critical article would be influenced by its own predictability within the broader context and retrieval operations that it might trigger. In line with the literature on pre-nominal effects, we expected to find N400 effects. An unpredicted article signals that the prediction for the upcoming noun needs to be revised (DeLong et al., 2005; Fleur et al., 2020; Nicenboim et al., 2020; Otten & Van Berkum, 2009; Van Berkum et al., 2005; Wicha et al., 2003, 2004; but see Kochari & Flecken, 2019; Nieuwland et al., 2018). This revision might involve memory retrieval of potential candidates for the noun based on the context sentence. This retrieval would operate across a sentence border. However, cue-based parsing models do not distinguish between words from different sentences. Therefore, whether interference effects emerge or not should not be affected by this. If a distinction between words of one sentence and another is assumed, it would affect the target noun and the distractor noun in our materials equally as they were both introduced in the sentence before the critical sentence.

Hypotheses

We hypothesised that the retrieval triggered by an unpredicted article would be influenced by interference. Interference degree (high vs. low interference) was assumed to influence retrieval as it affects the diagnosticity of the gender cue provided by the article to distinguish between target and distractor (cue-diagnosticity, Nairne, 2002). In low interference conditions, the gender cue unambiguously matched the target noun and retrieval should be easy. In contrast, in high interference conditions, the gender cue of the article was ambiguous as it matched the target and distractor noun, so it was complex to identify the target. In the context sentence, the nouns were introduced in two different orders to vary the recency of the target at the retrieval site, resulting in retroactive and proactive interference conditions (see “Materials and methods” section for further details). In line with the literature (Martin & McElree, 2009; Van Dyke & McElree, 2011), we expected to see reduced interference effects for proactive compared to retroactive interference. In sum, we expected interference effects in the N400 time window of unpredicted articles, but no interference effect at predicted articles.

At the noun position, we expected a predictability effect in the P300 and/or N400 time window. While the P300 amplitude would be larger for predicted than for unpredicted nouns, the N400 amplitude would be larger for unpredicted nouns than for predicted nouns. Interference effects at the noun position were not assumed, as the noun would not trigger retrievals.

Materials and methods

Participants

Forty-four right-handed participants (8 male, mean age: 22.75 years, range: 18 – 35) gave written informed consent and received either course credit or 20 Euros for their participation. All were native (Austrian) German speakers with normal or corrected-to-normal vision and no history of neurological or psychiatric disorders. None of the participants took part in the sentence completion pretest. Four participants had to be excluded because of technical issues, three were excluded because their response accuracy was below 75% and five participants were excluded because of excessive artefacts (more than 25% of the critical sentences were affected). The data of thirty-two participants (5 male, mean age: 22.34 years, range: 18 – 35) was used for final analyses.

Materials

Eighty-five sets of sentence pairs were constructed in a 2 (interference degree: high, low) × 2 (interference type: retroactive, proactive) factorial design (see Table 1 for an example). Two nouns, which either matched or mismatched according to their gender features, were introduced in a context sentence and one of them was repeated with a definite article as the object of the target sentence. However, both of the context nouns were semantically suitable to be the object of the target sentence. Target sentences were identical across conditions. The article and the noun in the target sentences were the critical positions for ERP analyses. In German, articles and corresponding nouns must have congruent gender, number and case features. So, in high interference conditions ((1) and (2) in Table 1), the article could grammatically be followed by each of the two nouns in the context, while in low interference conditions ((3) and (4) in Table 1), the article was only compatible with one noun in the context. We manipulated the order in which the nouns were introduced in the context sentence; resulting in two types of interference: retroactive and proactive interference. In retroactive interference conditions, the interfering distractor is linearly closer to the retrieval site, i.e. the article, than the target noun (target … distractor … retrieval site). In proactive interference conditions, the interfering distractor was encountered before the target noun was presented (distractor … target … retrieval site).

Table 1. Example material and cloze values (pretest results) for the article and noun within the target sentence displayed beside the corresponding context sentences.

Conditions	Context sentences	Cloze values
			Article	Noun
(1) high retroactive interference	In der Schachtel sitzt ein Käfer und im Glas liegt ein Wurm. “In the box there sits a beetle[masc.] and in the glas there lies a worm[masc.].”	mean (sd) range	0.77 (0.11) 0.56 – 1.0	0.45 (0.24) 0.0 – 0.94
(2) high proactive interference	Im Glas liegt ein Wurm und in der Schachtel sitzt ein Käfer. “In the glas there lies a worm[masc.] and in the box there sits a beetle[masc.].”	mean (sd) range	0.84 (0.07) 0.65 – 0.95	0.58 (0.23) 0.12 – 1.0
(3) low retroactive interference	In der Schachtel sitzt ein Käfer und im Glas liegt eine Raupe. “In the box there sits a beetle[masc.] and in the glas there lies a caterpillar[fem.].”	mean (sd) range	0.4 (0.25) 0.0 – 0.94	0.89 (0.08) 0.59 – 1.0
(4) low proactive interference	Im Glas liegt eine Raupe und in der Schachtel sitzt ein Käfer. “In the glas there lies a caterpillar[fem.] and in the box there sits a beetle[masc.].”	mean (sd) range	0.5 (0.21) 0.05 – 0.89	0.95 (0.07) 0.58 – 1.0
	Target sentence
For all conditions	Peter befreit den Käfer aus der Schachtel. “Peter frees the[masc.] beetle[masc.] from the box.”

Critical words are underlined. Sd = Standard deviation.

Sentence completion pretest

A sentence completion pretest was conducted to obtain cloze values for the critical article and the noun in context (Kutas & Hillyard, 1984; Taylor, 1953). This was done to achieve a measure of predictability of the article and noun within the specific sentence context. The material was split into four lists, with only one sentence pair from each lexical set occurring per list. In addition, there were two versions of each list: one truncated the target sentence before and one after the article, resulting in eight different versions. The different lists/versions were constructed to ensure that the pretest participants stayed naïve towards the manipulation. No fillers were included in the pretest. A total of 144 university students participated in the pretest (16 to 20 per pretest version; 39 male, mean age = 22.5 years, range: 18 - 40). Pretest participants were native speakers of (Austrian) German and were instructed to intuitively complete the sentence pairs. None of the pretest participants was included in the main experiment analyses.

Across conditions, the obtained cloze values ranged from 0 to 1 for both the article and noun. The cloze values are shown in Table 1. Pairwise contrasts were computed with paired t-tests in R (R Core Team, 2019, Version 4.03). The cloze value of the article in the high interference conditions was higher overall than in the low interference conditions (retroactive interference: diff = 0.37, p < 0.001, 95% CI = 0.32–0.43; proactive interference: diff = 0.33, p < 0.001, 95% CI = 0.29–0.38). This is a logical consequence of the fact that, no matter which of the two nouns of the high interference context the participants chose to continue the sentence, the article had the same form because the nouns shared the same gender feature. In contrast, the cloze values for the article in the low interference conditions were lower because the article form varied based on the decision regarding which of the context nouns should follow. In both high and low interference conditions, the participants showed a recency effect, i.e. they chose an article congruent with the last-mentioned noun more often than with the first-mentioned noun (see the higher cloze values for proactive interference conditions compared to retroactive interference; high interference: diff = 0.06, p < 0.001, 95% CI = 0.04–0.09; low interference: diff = 0.1, p < 0.001, 95% CI = 0.07–0.14). In the high interference conditions, this recency effect is difficult to interpret as the nouns share the same gender, but it is numerically small. It might be caused by an implicit bias related to verb semantics, i.e. one of the nouns might be more closely associated with the verb.

The cloze values for the target noun are lower overall in the high interference conditions than in the low interference conditions (retroactive interference: diff = - 0.43, p < 0.001, 95% CI = −0.49 – −0.37; proactive interference: diff = −0.37, p < 0.001, 95% CI = −0.43 – −0.32). This reflects the fact that the article is ambiguous in the high interference conditions and thus it can be followed by either of the two nouns leading to lower cloze values for both of them. In the low interference conditions, the article only matches one noun, so that the matching noun has a higher cloze value. Like the article results, the noun cloze values show a recency effect, i.e. higher cloze values for proactive interference conditions than retroactive interference conditions (high interference: diff = 0.13, p < 0.001, 95% CI = 0.09–0.16; low interference: diff = 0.07, p < 0.001, 95% CI = 0.05 – 0.08).

The distribution and combinations of the cloze values, respectively, are shown in Figure 1. In Figure 1, it is apparent that our materials contain more data points for higher cloze values than for low cloze values. Furthermore, the distribution of cloze values varies between conditions; this is caused by the above-mentioned probabilities of articles and nouns dependent on condition.

Figure 1. Distribution of cloze values for the article (A and C) and noun (B and D) as obtained by the sentence completion pretest.

Cloze values were strongly correlated with the factor interference degree (article cloze and interference degree: R = 0.71, p < 0.001, 95% CI = 0.65–0.76; noun cloze and interference degree: R = −0.75, p < 0.001, 95% CI = −0.79 – −0.69). Correlations between cloze values and interference type were weak (article cloze and interference type: R = −0.17, p = 0.003, 95% CI = −0.27 – −0.06; noun cloze and interference type: R = −0.18, p = 0.001, 95% CI = −0.29 – −0.07). Correlations were computed with the R function cor.test() with method Pearson and are presented in Figure 2.

Figure 2. Correlation of predictors: Article cloze, noun cloze, interference type and interference degree. This plot was generated with the R package ggcorrplot (Kassambara, 2019).

In summary, the participants of our pretest used the probability of an article with a specific gender feature to complete the sentence pair. This probability was mostly determined by the characteristics of the experimental conditions. Also, article and noun cloze values revealed an offline recency effect. Participants had a tendency to use an article that matched the gender of the last-mentioned noun or the last-mentioned noun itself to continue the sentence. This recency effect indicates that the order of introduction of the nouns influenced their predictability stronger than the target sentence verb suggesting that verb bias was a minor factor in our materials. The pretest results were used as predictors in our linear mixed model analyses (for details, see section “Statistical analysis”).

Procedure

Eighty of the original eighty-five items were presented in the main study. The exclusion of five items was done to achieve an even number of items and was not based on pretest results. Two pseudo-randomised lists were constructed so that each participant saw only two versions of one item (either version 1 and 4 or version 2 and 3 in Table 1, the same approach was used by e.g. Bornkessel-Schlesewsky et al., 2011). The reason for this was to maximise the amount of data that can be collected from a single participant (compared to using a Latin square design) as preparation for EEG recordings is time and labour intensive. To minimise the risk that participants would remember one version when they saw the other, we presented the two versions of an item with at least 90 trials between them. This was achieved by presenting the two versions of one item in different experimental blocks that did not directly follow each other. To avoid order effects, half of the participants were presented a backward version of the lists. Each of the lists consisted of 240 sentence pairs (160 critical and 80 filler sentence pairs). The first sentences of the filler sentence pairs were structurally identical to the context sentences of the critical sentence pairs. The second sentences of the fillers were constructed to be coherent to the first sentences but did not include a referential link to it (e.g. In der Pfanne brät ein Steak und im Topf kocht eine Suppe. Luis hat Hunger und freut sich auf das Essen. “In the pan there fries a steak and in the pot there boils a soup. Luis is hungry and is looking forward to the meal.”). All sentence pairs (both critical and filler pairs) were grammatical. A third of all sentence pairs in both lists (80 out of 240) was followed by a “yes” / “no” comprehension question. There were three types of questions: a) about the action or state of one noun phrase within the context sentence (Liegt der Wurm im Glas? “Is the worm in the glass?”), b) about which of the noun phrases was used in the target sentence (Befreit er den Käfer? “Does he free the beetle?”) or c) the content of the target sentence after the critical words (Befreit er ihn aus der Schachtel? “Does he free it from the box?”). Note that the provided example questions are hypothetical questions that correspond to the example item presented in Table 1. Each participant was asked to answer at most one question per item. Questions were not repeated within lists. Half of the questions required a “yes” answer indicated by a right or left mouse click (balanced across participants).

Participants were instructed to silently read the sentence pairs, which were presented on the centre of a computer screen in a word-by-word manner. Words were presented in white letters on a black background and participants were asked to minimise eye movements, blinks and general movements. A trial began with a 400 ms fixation cross followed by 200 ms blank screen. Words were presented for 400 ms each with a 100 ms inter-stimulus interval between words; after the last word of the first sentence in the pair, there was a 1000 ms inter-stimulus interval. To encourage participants to read carefully, a third of the trials was followed by a “yes” / “no” comprehension question, which the participants had five seconds to answer. After that (or in trials without a task), there was a self-paced inter-trial interval, during which participants could blink as needed before pressing a button to start the next trial. All sessions began with a practice block of five sentence pairs to familiarise participants with the procedure. The following six experimental blocks were separated by short breaks. Each experimental session lasted approximately 2.5 h; participants spent approximately 75 min on the task. The study was conducted in accordance with the Declaration of Helsinki (World Medical Association, 2009). We do not claim accordance to the current version of the Declaration of Helsinki as it requires preregistration.

Electroencephalogram recording and preprocessing

The electroencephalogram (EEG) was recorded from 58 active scalp electrodes positioned according to the standard 10–20 system and attached to an elastic cap (ActiCap from Easycap GmbH, Herrsching, Germany). EEG recordings used an ActiCHamp amplifier (Brain Products GmbH, Gilching, Germany) with a sampling rate of 500 Hz and AFz as the ground electrode. Six electro-oculogram (EOG) electrodes were positioned above and below both eyes and at the outer canthus of each eye. Impedances were kept below 10 kOhm. All electrodes were online referenced to the left mastoid and re-referenced to the average of both mastoids offline. The raw data were filtered with a bandpass of 0.1–30 Hz. After an ICA was conducted to identify and reduce eye movement artefacts, artefact rejection excluded epochs (−200 to 1000 ms around critical word onset) where peak-to-peak amplitude for EEG channels exceeded 150 μV. ERPs were computed in epochs of −200 to 1200 ms relative to critical word onset. Preprocessing steps were carried out using MNE python (Gramfort et al., 2013) and the python package philistine (Alday, 2019b).

Statistical analysis

We performed (generalised) linear mixed model analyses using the R (R Core Team, 2019, Version 4.0.3) package lme4 (Bates et al., 2015, Version lme4_1.1-25) to analyse both the comprehension question accuracy as well as the electrophysiological data. All models included random intercepts for participants and items. Whether a more complex random effects structure (random slopes) was appropriate for the models was tested with hierarchical model selection (see section “Results”). Models with and without specific random slopes were compared using likelihood-ratio tests provided by the anova() function in R. Fixed effects structure is described below. Whether effects and interactions were significant was determined by comparing models with and without that effect/interaction with the R function anova(). Visualisations of effects of interest were generated with the ggplot2 package (Wickham, 2016, Version ggplot2_3.3.3). Prior to the analyses, we excluded two items that were inconsistently constructed, leaving us with data for 78 items. Below we describe the details of the analyses of the comprehension question accuracy and the electrophysiological data, in turn.

Comprehension question accuracy

We used generalised linear mixed models, i.e. logistic mixed effects regression, to analyse comprehension question accuracy. Although comprehension question accuracy was not part of our hypotheses, we wanted to investigate whether/how our experimental manipulations (interference type, interference degree) influenced comprehension accuracy and corresponding reaction times. As fixed effects, our models included the log-transformed reaction time, the factors interference type, interference degree, and the interaction of the factors. A model with a three-way interaction did not converge. Both factors were sum contrast coded (interference type: retroactive was coded as 1, proactive was coded as −1; interference degree: high was coded as 1, low was coded as −1). With this contrast coding, model estimates for one factor level represent the difference between this factor level and the grand mean of a factor (for detailed discussion of contrast coding, see Schad et al., (2020)). Cloze probability of article and/or noun was not included in these analyses because we did not expect that predictability of a specific word within two sentences would affect overall comprehension.

Electrophysiological data

We used linear mixed models to analyse the activity of single trial EEG data for the critical article and following noun. As both investigated ERP components, the P300 and the N400, are known to have a centro-parietal scalp distribution, we averaged the activity of a set of 15 centro-parietal electrodes (Cz, C1/2, C3/4, CPz, CP1/2, CP3/4, Pz, P1/2, P3/4) for all planned analyses.⁵ All models included the baseline EEG activity from 200 ms pre-stimulus until stimulus onset as a continuous predictor (Alday, 2019a); this corresponds to the standard N400 time window of the preceding word. We do not interpret baseline effects as they were included in the models as a covariate instead of traditional baseline correction (for detailed discussion, see Alday, 2019a). Inclusion of other predictors/fixed effects was determined by our hypotheses, which we summarise here for convenience. At the article, we hypothesised to find an interference effect at low cloze articles that modulates the N400 amplitude, but no interference effect at high cloze articles (i.e. an interaction of predictability and interference). Therefore, the models for the article investigated the N400 time window from 300 to 500 ms post word onset and included fixed effects for centred continuous article cloze probability, interference type, and their interaction. Interference degree was not included as a fixed effect in our analyses of the electrophysiological data as it was strongly correlated with cloze probability in our material (see section “Sentence Completion Pretest”). Interference type was modelled as a sum contrast coded two-level factor (retroactive was coded as 1, proactive was coded as −1). Under lme4 default settings, the model produced a convergence warning. In line with the R documentation regarding convergence issues, we refitted the model with all available optimisers. All but the default optimiser showed no warnings. We continued with the model using the optimiser bobyqa. For the noun, we hypothesised to find predictability effects for high cloze compared to low cloze nouns with more pronounced amplitudes in the P300 (200 - 400 ms) and/or N400 time window (400 - 600 ms) and therefore fitted models which included the continuous centred noun cloze probability.

Results

The data and analysis scripts for this experiment can be accessed at https://osf.io/ju4mr/.

Comprehension question accuracy

Participants were encouraged to answer the comprehension questions correctly and within the time limit (five seconds). As noted above, three participants were excluded for accuracy rates below 75%. For the remaining participants, comprehension accuracy was good (mean: 85.4%, range: 76.2 - 95%), thus we can assume that they processed the sentences attentively. A generalised linear mixed model with log reaction time and the interaction of interference type and interference degree and random intercepts for participants and items revealed that correct answers were associated with a faster response time than incorrect answers (1941ms vs. 2426 ms, z = −7.97, p < 0.001). No other effects were significant in this model. There was no evidence that adding by-participant random slopes for interference degree improved fit (χ²= 0.04, p = 0.98). A model with by-participant random slopes for interference type resulted in a singular fit warning and a model with by-participant random slopes for reaction time did not converge. Models with by-item random slopes were not fit after by-participant random slopes were not fruitful.

Electrophysiological data

Article

The interaction of interference type and centred article cloze was not significant, i.e. a model with the interaction did not provide a better fit than a model with just the main effects of the predictors (χ²= 3.06, p = 0.08). Both models included only by-participant and by-item random intercepts because random slopes did not improve the interaction model. Inclusion of by-participant random slopes for interference type produced a singular fit warning and inclusion of by-participant random slopes for centred article cloze did not improve fit (χ²= 0.92, p = 0.63). Nevertheless, we investigated the interaction further because it reflects our main hypothesis. Figure 3 shows the difference between interference types within high and low cloze articles.

Figure 3. Panel A) ERPs elicited by high and low cloze articles at electrode CPz. Article onset is at 0 ms and noun onset at 500 ms. For visualisation, it was necessary to dichotomise the continuous article cloze probability (high cloze > 0.66; low cloze < 0.33). Negativity is plotted upwards. Panel B) Topographic map of difference wave voltage in μV for brain responses to high cloze articles (retroactive interference – proactive interference) averaged across the N400 time window (300–500 ms after article onset). All topographic maps were generated using code by Craddock (2017). Panel C) Topographic map of difference wave voltage in μV for brain responses to low cloze articles (retroactive interference – proactive interference) averaged across the N400 time window (300–500 ms after article onset).

Visual inspection of Figure 3 suggests that interference types differed from each other at low cloze, but not at high cloze. For low cloze articles, retroactive interference resulted in a negativity compared to proactive interference. There was no difference between interference type conditions for high cloze articles. The topographic map presented in Figure 3 C suggests that the negativity had a broad distribution with a fronto-central focus. Therefore, the selected region of interest (ROI) of centro-parietal electrodes was suboptimal to investigate this negativity. We used fronto-central electrodes (Fz, F1/2, F3/4, FCz, FC1/2, FC3/4, Cz, C1/2, C3/4) in a subsequent post-hoc analysis. For the fronto-central ROI, the interaction model provided better fit to the data than a model including only main effects (χ²= 4.8, p = 0.028). The results of the interaction model for the fronto-central ROI with by-participant and by-item random intercepts are shown in Table 2. A model with by-participant random slopes for interference type produced a singular fit warning. A model with by-participant random slopes for centred article cloze did not improve fit (χ²= 3.09, p = 0.21).

Table 2. Article. Results (fronto-central ROI, 300 – 500 ms).

Linear mixed model fit by maximum likelihood. t-tests use Satterthwaite’s method [“lmerModLmerTest”] Formula: amplitude ∼ 1 + baseline + centred article cloze * interference type + (1|participant) + (1|item)
	AIC	BIC	logLik	deviance	df.resid
	27602.5	27654.5	−13793.2	27586.5	4933
Scaled residuals:
	Min	1Q	Median	3Q	Max
	−3.8926	−0.6176	0.0050	0.6349	4.3552
Random effects:
	Groups	Name	Variance	Std.Dev.
	item	(Intercept)	0.08162	0.2857
	participant	(Intercept)	0.64477	0.8030
	Residual		15.29765	3.9112
Number of obs: 4941, groups: item, 78; participant, 32
Fixed effects:
	Estimate	Std.Error	df	t value	Pr(>|t|)
(Intercept)	−1.12056	0.16630	44.52378	−6.738	<0.0001
baseline	−0.14168	0.01095	4930.60294	−12.941	<0.0001
c. art. cloze	−0.20882	0.35262	3297.46089	−0.592	0.5538
interf. type1	−0.13571	0.11294	4874.41668	−1.202	0.2296
cloze:interf1	1.00360	0.45801	4834.17496	2.191	0.0285

c. art. cloze = centred article cloze, interf. type = interference type, cloze:interf1 = centred article cloze: interference type1.

The estimate of the interaction in Table 2 indicates that the retroactive interference condition was more influenced by an increase of article cloze than the proactive interference condition. When article cloze increased, the amplitude for retroactive interference became approximately 1 μV more positive. That led to a negativity for retroactive interference compared to proactive interference for low cloze articles, but no such difference for high cloze articles (see Figure 3). This negativity had a broad distribution with a frontal focus (see Figure 3 C).

In Figure 4, ERPs for low cloze articles at nine electrode sites are shown to visualise the morphology of the observed negativity. Visual inspection of the ERP waveforms reveals that the negativity has a broad distribution and is visible at all presented electrode sites (electrode F1 might be the only exception). This stands in contrast to the topographic map (Figure 3 C) which showed a frontal focus for the averaged activity in the investigated time window (300–500 ms). At frontal electrodes, conditions start to differ from each other earlier (around 290 ms post word onset) than at central and posterior sites. Due to this earlier effect onset, conditions differ across the whole time window. At central and parietal electrodes, the negativity is only visible from approximately 350 to 500 ms. These different effect onsets at different electrode sites might be the reason for the frontal focus of the effect in the topographic map (Figure 3 C), which is not reflected in the ERP waveforms presented in Figure 4.

Figure 4. ERPs elicited by low cloze articles at nine selected electrodes. Article onset is at 0 ms and noun onset at 500 ms. The shaded areas represent the time window for analysis (300–500 ms post article onset). For visualisation, it was necessary to dichotomise the continuous article cloze probability (high cloze > 0.66; low cloze < 0.33). Negativity is plotted upwards.

Noun

For the noun, we analysed two time windows: a P300 time window ranging from 200 to 400 ms and an N400 time window from 400 to 600 ms. We present the results for the two time windows in turn below.

A model of the electrophysiological activity in the P300 time window of the noun that included centred noun cloze provided better fit than a model without it (χ²= 14.3, p = 0.0002). We present the results of that model in Table 3. The model included by-participant and by-item random intercepts. A more complex random effect structure did not improve model fit (comparison base model and model with by-participant random slopes for centred noun cloze: χ²= 0.74, p = 0.69).

Table 3. Noun. Results (centro-parietal ROI, 200 – 400 ms).

Linear mixed model fit by maximum likelihood. t-tests use Satterthwaite’s method [“lmerModLmerTest”] Formula: amplitude ∼ 1 + baseline + centred noun cloze + (1|participant) + (1|item)
	AIC	BIC	logLik	deviance	df.resid
	28949.9	28988.9	−14468.9	28937.9	4945
Scaled residuals:
	Min	1Q	Median	3Q	Max
	−7.0992	−0.6476	−0.0232	0.6576	3.9490
Random effects:
	Groups	Name	Variance	Std.Dev.
	item	(Intercept)	0.1145	0.3383
	participant	(Intercept)	2.1065	1.4514
	Residual		19.7590	4.4451
Number of obs: 4951, groups: item, 78; participant, 32
Fixed effects:
	Estimate	Std.Error	df	t value	Pr(>|t|)
(Intercept)	2.35666	0.26756	33.52909	8.808	< 0.0001
baseline	−0.04359	0.01248	4940.35357	−3.492	0.0005
c. noun cloze	0.91016	0.24012	2629.39616	3.790	0.0002

c. noun cloze = centred noun cloze.

The main effect of centred noun cloze in Table 3 shows that P300 amplitude increased as centred noun cloze increased (see Figure 5 A). The effect had a broad distribution with a left posterior focus (see Figure 5 B).

Figure 5. Panel A) ERPs at the noun for centred noun cloze effects in the P300 and N400 time windows (200 - 400 ms and 400 - 600 ms, respectively) at electrode CPz. Noun onset is at 0 ms. For visualisation, it was necessary to dichotomise the continuous noun cloze probability (high cloze > 0.66; low cloze < 0.33). Negativity is plotted upwards. Panel B) Topographic map of difference wave voltage in μV (high cloze – low cloze) averaged across the P300 time window (200–400 ms after noun onset). Panel C) Topographic map of difference wave voltage in μV (high cloze - low cloze) averaged across the N400 time window (400–600 ms after noun onset).

We present the results of a model of the electrophysiological activity in the N400 time window of the noun that included centred noun cloze (see Table 4). Inclusion of centred noun cloze improved model fit compared to a model without it (χ²= 21.4, p < 0.0001). A model with by-participant random slopes for centred noun cloze did not improve model fit compared to the model with just random intercepts for participants and items (χ²= 1.48, p = 0.48).

Table 4. Noun. Results (centro-parietal ROI, 400 – 600 ms).

Linear mixed model fit by maximum likelihood. t-tests use Satterthwaite’s method [“lmerModLmerTest”] Formula: amplitude ∼ 1 + baseline + centred noun cloze + (1|participant) + (1|item)
	AIC	BIC	logLik	deviance	df.resid
	27907.0	27946.1	−13947.5	27895.0	4945
Scaled residuals:
	Min	1Q	Median	3Q	Max
	−3.8407	−0.6317	−0.0147	0.6451	4.7244
Random effects:
	Groups	Name	Variance	Std.Dev.
	item	(Intercept)	0.2700	0.5196
	participant	(Intercept)	0.5611	0.7490
	Residual		16.0061	4.0008
Number of obs: 4951, groups: item, 78; participant, 32
Fixed effects:
	Estimate	Std.Error	df	t value	Pr(>|t|)
(Intercept)	0.42578	0.15644	42.52072	2.722	0.009
baseline	−0.22661	0.01124	4943.19045	−20.164	< 0.0001
c. noun cloze	−1.02249	0.22110	3606.07157	−4.625	< 0.0001

c. noun cloze = centred noun cloze.

The main effect of centred noun cloze in Table 4 indicates that the N400 amplitude increased (became more negative), when noun cloze increased (an increase of one unit of cloze was accompanied by approximately 1 μV increase in N400 amplitude, see Figure 5 A). The effect had a broad distribution (see Figure 5 C).

Discussion

We have presented an ERP study that investigated the interplay of prediction and interference at an article within sentence pairs in German. Additionally, prediction was examined at the following noun. At the position of the article, a linear mixed model estimated an interaction between interference and predictability. At unpredicted articles, retroactive interference led to an increased negativity compared to proactive interference. At predicted articles, interference types did not differ. For the noun, effects of noun cloze in both the P300 and N400 time window were estimated. In the P300 time window, high cloze nouns elicited a more positive response than low cloze nouns. In contrast, in the N400 time window, high cloze nouns resulted in a more negative response than low cloze nouns. In the following, we will discuss the results of the article and noun in turn.

Article

The hypothesised interaction of interference type and predictability at the article was only significant in an analysis of activity at fronto-central electrodes. Unpredicted articles led to a negativity in the retroactive interference condition compared to the proactive interference condition. This effect is compatible with the assumption that under low predictability, there is no or very limited pre-activation. Therefore, the encounter of an unpredicted word induces memory retrievals and the possibility of interference. In contrast, under high predictability, interference types did not differ. This result pattern is in line with our hypothesis. It provides support for our proposal that pre-activation and memory retrieval are two sides of the same coin as both activate memory representations. Pre-activation of a word brings forward processing steps, such as retrieval and integration (see Vasishth & Drenhaus, 2011, Husain et al., 2014). Freunberger and Roehm (2017) provided first evidence for the cost of pre-activating the noun at a pre-nominal adverb. Future work should investigate further when pre-activation happens and whether retrieval and integration cost can be found at these positions. The encounter of a pre-activated word does not add information, thus no additional processing is needed. Consequently, without memory retrieval, there is no retrieval interference. Similarly, in the predictive coding theory, fully predicted input gets “explained away” by predictions, i.e. it is not processed further (Bastos et al., 2012; Clark, 2013; Feldman & Friston, 2010; Friston, 2005, 2010; Rao & Ballard, 1999).

The electrophysiological correlates of similarity-based retrieval interference

We hypothesised to find that interference modulates the N400. The observed negativity had a broad distribution and had an earlier onset at right, frontal electrodes, unlike the centro-parietal N400. Our finding aligns well with the interference ERP literature reporting frontal negativities for interference effects with the more detrimental condition eliciting the greater negativity (Vasishth & Drenhaus, 2011; Lee & Garnsey, 2015; Martin et al., 2014). Therefore, our study provides converging support that interference effects during language comprehension manifest in (frontal) negativities. The lateralisation of that negativity is unclear so far: Vasishth and Drenhaus’ (2011) negativity was left lateralised, Lee and Garnsey (2015) and Martin et al. (2014) found bilateral negativities and we observed a broadly distributed negativity which was more pronounced on the right. The frontal negativities elicited by interference have been discussed as Left Anterior Negativities (LAN, Lee & Garnsey, 2015; Vasishth & Drenhaus, 2011) and/or Nrefs (Lee & Garnsey, 2015; Martin et al., 2014). The LAN is a negative deflection with a frontal, left localised distribution. Typically, it is evoked by morpho-syntactic violations and arises from 300 to 500 ms post word onset (e.g. Coulson et al., 1998). However, it has been suggested that the LAN indexes increased memory load during language processing in general (e.g. Kluender & Kutas, 1993; see Münte et al., 1998, for a similar, but sustained memory-based negativity). This interpretation would fit the negativities seen in interference research. However, the negativity in the present study showed a broad rather than left anterior distribution, rendering it dissimilar to the LAN.

The Nref is a sustained negative deflection with a broad distribution with a frontal focus that starts somewhere around 200 - 400 ms post word onset (Nieuwland et al., 2019; Van Berkum, 2009; Van Berkum et al., 2007). Generally, it is interpreted as a marker of the difficulty that arises during reference resolution in contexts with multiple possible antecedents (e.g. “the man” in a two men context, Van Berkum et al., 1999). This kind of referential ambiguity seems to be similar to a retrieval with multiple possible candidates, i.e. similarity-based retrieval interference. The morphology of the negativity in the present study is similar to the Nref. However, our negativity was too short-lived to be an Nref. Future work is needed to further evaluate the electrophysiological correlates of similarity-based retrieval interference. The shifts in negativity topography and latency observed in interference-related contexts across studies might reflect a functionally similar underlying component with slightly different manifestations dependent on the context in which it is observed (Bornkessel-Schlesewsky & Schlesewsky, 2019).

Interference effects in German

So far only a few studies found interference effects in German (Vasishth & Drenhaus, 2011; Nicenboim et al., 2018; Pregla et al., 2021), although a few other studies aimed to find such effects as well (e.g. Jäger et al., 2015; Laurinavichyute et al., 2017). The difference between retroactive and proactive interference conditions for unpredicted articles in our study provides converging support for interference effects in German. Furthermore, our study is only the second to report ERP results of interference effects in German (Vasishth & Drenhaus, 2011).

Lack of a main effect of prediction

In our linear mixed model of the interaction of prediction and interference, no main effect of predictability at the article was found. Such pre-nominal prediction effects have been discussed controversially (DeLong et al., 2005; Fleur et al., 2020; Kochari & Flecken, 2019; Nieuwland et al., 2018; Otten & Van Berkum, 2009). Our experimental design differed from previous studies because the to-be-predicted noun phrase was introduced verbatim in the context sentence. This could be the reason why we did not observe a pre-nominal prediction effect. Recently, a cross-linguistic Bayesian meta-analysis of studies investigating pre-nominal effects in English, Dutch and German found clear evidence for small pre-activation induced effects at articles (Nicenboim et al., 2020).

Limitations

While the observed result pattern fits our hypothesis and previous findings regarding the interaction of prediction and memory retrieval/interference (Campanelli et al., 2018; Husain et al, 2014), it is important to keep in mind that it was only significant in our post-hoc analysis. Additionally, we have identified the following limitations in our design: Firstly, data for low cloze articles which drove the interaction was sparse (see Figure 1 A, C and noisiness of the ERP waves in Figure 3 A). A challenge for future work will be to employ items with an even distribution of predictability of the critical words. Secondly, we could only model the interaction of interference type (retroactive vs. proactive) and prediction. We did not fit models that included interference degree (high vs. low) and cloze values as a marker of predictability because those two predictors were strongly correlated in our materials. Previous studies of proactive interference and retroactive interference found the latter to be more detrimental during sentence processing (Martin & McElree, 2009; Van Dyke & McElree, 2011). Therefore, the contrast between the interference types proactive and retroactive was still suitable for our study, but a comparison of high and low interference conditions could have generated additional insights. Future work should employ designs were interference degree and predictability are not correlated.

Noun

At the position of the noun, we found P300 and N400 effects. Both were sensitive for differences in predictability. We discuss these findings in turn.

Effects in the P300 time window

High cloze nouns led to a more positive P300 than low cloze nouns. This effect is in line with the P300 literature reporting predictability effects during sentence comprehension when only one single best completion was expected (Roehm et al., 2007, Kutas & Iragui, 1998; Vespignani et al., 2010). As is shown in Figure 1 B and D, high noun cloze values stem mainly from conditions with low interference degree. In these conditions, the gender of the article unambiguously identified the noun before it was presented. This might have led to an extremely strong prediction, which resulted in the P300 for these highly predicted nouns compared to low cloze nouns.

Effects in the N400 time window

In the N400 time window of the noun, we found a predictability effect, which had the opposite direction than expected: A more negative response for high cloze nouns than for low cloze nouns. The absence of the expected predictability effect, i.e. a reduced N400 amplitude for predicted versus unpredicted words, could be due to priming. Priming is known to reduce N400 amplitude (Kutas & Federmeier, 2000). Independent of its predictability in the target sentence, the critical noun could have been primed by its verbatim presentation in the context sentence. Therefore, predicted as well as unpredicted critical nouns could have elicited a reduced N400 because they have been primed by their previous encounter in the context sentence (see Brouwer et al., 2012, for a related discussion on the absence of N400 effects in semantic role reversal anomalies because of priming).

We assume that the observed effect might be an instance of the so-called repeated name penalty. The repeated name penalty describes the increased processing demands that arise when co-reference to a prominent entity is realised via repetition of the noun phrase rather than with a pronoun (Gordon et al., 1993). It has been explained by means of the informational load hypothesis: A referential expression should not carry more informational load than is necessary to identify the antecedent (Almor, 1999). Based on our pretest results, we can conclude that in our materials there was a strong preference for the second entity of the first sentence to be referenced again in the second sentence (recency effect). Therefore, in the second sentence, a less informative referential expression would have been sufficient to identify the antecedent and thus the repeated name penalty arose. Electrophysiologically, the repeated name penalty manifests in more pronounced N400 amplitudes for repeated names that refer to a prominent entity compared to repeated names that refer to a less prominent entity (Camblin et al., 2007; Swaab et al., 2004). However, Hoeks and Brouwer (2014) proposed that the repeated name penalty might be an Nref effect rather than an N400 effect. They discussed the Nref as a general marker of referential processing. While the distribution of the negativity that we found and its onset suggest that it might be an Nref effect, its duration (approximately 200 ms) does not support that assumption. In previous studies, the repeated name penalty disappeared when the antecedent was introduced within a coordinated construction (Dirk (and Becca) pushed the desk into the corner because Dirk needed room for the filing cabinet., Swaab et al., 2004). But in an ERP study using full noun phrases as referential expressions, Burkhardt and Roehm (2007) found no differences between repeated noun phrases that referred to entities that were introduced as a single noun phrase or within a coordinated construction. So, while our finding is not fully consistent with the repeated name penalty literature, the observed negativity at the noun position is similar to the repeated name penalty.

Conclusion

The hypothesised interaction of memory retrieval and prediction was found in a broadly distributed negativity elicited by an article in the present ERP study. Interference effects were found under low predictability, but not under high predictability of the word at the retrieval site. This interaction suggests that the pre-activation of a fully predicted word prevents the initiation of a memory retrieval at said word. However, an unpredicted word triggers retrieval, which can be influenced by interfering distractors in memory. This finding provided initial evidence for our proposal that prediction and memory retrieval rely on the same underlying process as both result in the activation of memory representations.

Acknowledgements

The authors would like to thank Shravan Vasishth, Jonas Diekmann and Phillip M. Alday for helpful discussions related to the research reported here. Thanks are also due to Anja Bergmair, Nadine Furtner and Veronica Baldin for their contributions to the experimental materials and to Barbara Sophie Hartl and again Anja Bergmair for their assistance during data acquistion.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This research was supported by a research grant awarded to P.S. by the University of Salzburg. I.B.-S. acknowledges the support of an Australian Research Council Future Fellowship (FT160100437).

Notes

1 ERP results from Vasishth and Drenhaus (2011) could also be interpreted as showing similarity-based retrieval interference. We discuss their study below, together with other studies examining the interaction of interference and prediction.

2 An exception is Staub (2010) who investigated relative clauses in English. He found increased processing demands in object relative compared to subject relative clauses at two positions (determiner of the relative clause-initial noun phrase and relative clause verb). He concluded that the determiner effect was due to a disconfirmed prediction and the verb effect was due to retrieval cost. Given these results, Staub (2010) formulated the need for an account that reconciles memory and prediction, such as the activation-based model of cue-based retrieval by Lewis and Vasishth (2005). Both effects observed by Staub (2010) could be derived from the Lewis and Vasishth (2005) model.

3 The ERP results from Vasishth and Drenhaus (2011) provided evidence for locality effects only.

4 However, due to correlations between interference degree and cloze probability, only interference type was used in the ERP analyses (see below).

5 But see the selection of electrodes for a post-hoc analysis below.