Prediction in SVO and SOV languages: processing and typological considerations

1 Introduction

A good deal of psycholinguistic research has been devoted to determining the kinds of information that guide parsing decisions. For example, Frazier (1987) examined the effects of proximity and structural complexity, and Altmann and Steedman (1988) addressed discourse context. Ueno and Polinsky (2009) have shown how headedness may affect processing in languages with different word orders, teasing apart some apparently universal processing biases (i.e., the pro-drop bias) from those pertaining to specific language type (e.g., the intransitive bias in OV languages). Of special interest in the present study is research that has dealt with the effects of head information on processing, and in particular, prediction.

A number of linguistic theories have advocated that the head constrains and determines the kind of information that will be available in the sentence because of its power to determine relationships with the other constituents (e.g., Chomsky 1965; Grimshaw 1990; Jackendoff 1972).^[1] Psycholinguistic research has tried to determine empirically whether the information supplied by the head is crucial for parsing, and a number of head-driven parsers have been proposed, some more and some less radically head-driven (e.g., Pritchett 1991, 1992; Trueswell and Tanenhaus 1994a, 1994b). On the other hand, other studies have pointed out that the structural (incremental) analysis of a phrase may be initiated before its head is reached. That is, it has been proposed that processing of verb-argument structure can happen before the verb itself is reached (e.g., Konieczny 1996; Mazuka and Itoh 1995). For example, Bornkessel et al. (2004) reported that morphological case can be used to assign a theta-role interpretation to a noun phrase (NP) before the verb is encountered. Kamide and Mitchell (1999) have shown that some pre-head incremental processing occurs in Japanese (i.e., in early dative attachment disambiguation), and that head-driven accounts of parsing do not quite capture the data. However, they conclude that their pre-head parsing account cannot be seen as any more generally applicable than head-driven accounts, and they speculate that parsing mechanisms may differ across languages as a consequence of the specific structural configuration of the language. Thus, the comprehension system will be configured as “a head-driven system when it is needed to handle head-initial languages, whereas it takes the alternative pre-head driven form when dealing with the head-final languages, like Japanese” (Kamide and Mitchell 1999: 657).

The verb has often been argued to play a crucial role in incremental parsing and also in production (see Melinger et al. 2009 for a summary of relevant studies). The general assumption is that verbs are better predictors of their arguments than vice versa. This is because objects are listed in lexical co-occurrence frames for each verb, but a co-occurring verb, or set of verbs, is not listed for each noun (cf. Gentner 1981). Noun phrases and their thematic roles, for example, agent, patient, instrument, etc., do not in general select for, and predict, particular predicates, because they are compatible with too many alternatives. A verb, by contrast, activates a finite set of co-occurrence frames along with frequency-based or contextual preferences. If the verb occurs before one or more of its arguments, then the activated frames can be gradually reduced and any preferences confirmed or disconfirmed as the input continues. If the arguments occur before the verb, then alternative co-occurrence frames will not generally be activated and reduced, but instead, the intended co-occurrence frame for the verb will be selected at the verb and in a way that takes account of the preceding arguments. Some pre-head parsing may occur before the head is reached, but full assignment of all constituents must happen at the verb and will be determined by the specifications of that verb. In short, verbs are assumed to be better predictors of nouns than nouns are of verbs.

Ueno and Polinsky (2009) point out that this asymmetry in verb-argument processing is fully captured by Pritchett’s (1992) head-driven parsing model, where it is proposed that syntactic attachment happens at the head. As the parser moves along the string in an SOV language, it encounters different items that must be held in memory before it reaches the verb. For an SVO language, like English, only the subject argument comes before the verb. With SOV both subject and object and possibly other complements and adjuncts precede the verb. SOV constructions should therefore exert an extra (memory-based) processing cost prior to integrating the pre-verbal constituents with the verb when the verb is encountered. However, speakers of SVO languages also burden their working memory, not incrementally, but immediately as soon as they encounter the verb because they need to hold predicted items (including co-occurrence frames) in memory, which adds to the online processing cost.

Gibson (1998) appeals to this added processing cost resulting from online prediction in working memory in his explanation for the dispreferred ordering of direct objects before subjects in German and Finnish, two languages with rich case marking and considerable word order flexibility. Object before subject is highly dispreferred, and Gibson argued that this is because a preceding case-marked object NP predicts a subject in these languages, whereas a subject does not predict an object, resulting in a memory-based preference for SO over OS (see also Levy 2013). By this logic, the English SVO pattern would be better for prediction and worse for working memory, with respect to the verb-object relationship, while the Japanese SOV would involve less working memory demands prior to the verb but would permit fewer online predictions. However, this view may be overly simple, and as Hawkins (2022) argues, there are also other factors that encourage or discourage the positioning of one category prior to another in addition to prediction and working memory load, including (1) the agent versus patient theta-role assigned to a noun phrase, and (2) its animacy or inanimacy.

The precise role of online prediction in Japanese is complicated further by an important typological consideration documented in Polinsky and Magyar (2020). They point out that verb-final languages have much lower ratios of verbs to nouns in their lexicons than verb-early languages, and more generally that the verb-to-noun ratio declines proportionately as the verb moves further to the right in the clause across languages. This means, they argue, that more information content is given earlier in processing about the event in question, by a rich and differentiated set of nouns in SOV languages, whereas verb-early languages describe the event early on the basis of their richer set of verbs and the co-occurrence frames they activate.^[2]

Nouns in SOV languages can accordingly be predictive of further details of an event, and even of upcoming verb tokens selected from the smaller set available in this language type (cf. further Strunk et al. 2015). In the extreme case Polinsky and Magyar (2020) point out that there are many instances in Japanese of semantically rich nouns co-occurring with semantically empty “light” verbs like suru (do), which add little or nothing semantically to an event description following the lexical content of the accompanying noun, as in guuguuru suru (i.e., ‘to Google’, literally google do) (Polinsky and Magyar 2020). Light verbs are found in many SOV languages (cf. Amberber et al. 2010), though they do also occur in certain non-SOV languages (including English) for structural and processing reasons that may be independent of online prediction (cf. Hawkins 2019).

Prediction and incremental argument interpretations have been much discussed in Surprisal Theory (Hale 2001; Levy 2008). Surprisal is measured by the degree to which a word is unpredictable due to its inconsistency with previous assignments of structure and meaning that have been made at earlier points in the string. There is evidence that the processing load in the interpretation of a word is often proportional to its surprisal (e.g., Frank et al. 2015; Smith and Levy 2013). Most of the work done in this theoretical framework has been done on English or similar languages (cf. Hörberg 2016). The role of prediction (and its close cousin surprisal) may have been overestimated because of the type of language examined, and when structurally different languages are considered, the extent of prediction/surprisal may need to be reconsidered.

Indeed, a study on German (a mixed VO/OV language) demonstrated that both prediction and memory-load restrictions in integration play a role in the processing of the same syntactic structure within the language (Levy and Keller 2013). The results were framed within the debate over locality (which claims that additional material that needs to be integrated with a head increases processing effort through distant and non-local dependencies; see Gibson 1998, 2000) versus anti-locality (which claims that additional material facilitates processing; see Surprisal Theory and related models; Konieczny 2000; Konieczny and Döring 2003; Levy 2008). In Levy and Keller’s study, locality and anti-locality effects were examined in verb-final constructions in German, while controlling for lexical identity, plausibility, and sentence position. Clear anti-locality effects were found in the main clause, where presence of a preceding dative argument facilitated processing at the final verb (contra the predictions of Gibson 1998).

However, in subject-extracting relative clauses with identical linear ordering of verbal dependents, Levy and Keller (2013) found locality effects in addition to anti-locality effects. Specifically, sentence processing was facilitated when the verb was preceded by a dative argument alone and hindered when both the dative argument and an adjunct preceded the verb. Apparently, under some circumstances, locality effects can partially override expectation-based facilitation. Levy and Keller (2013: 215) offered the explanation that “native speakers of verb-final languages are simply more practiced, and therefore, more skilled at comprehending non-local syntactic configurations”. They also noted other studies that appear to offer empirical support for this interpretation. For instance, average total-sentence dependency lengths have been shown to be considerably longer in German than in English (Gildea and Temperley 2010; Park and Levy 2009). Second, Vasishth et al. (2010) found that German speakers are better than English speakers at tracking multiple incomplete noun-verb dependencies induced by multiple center embedding. Levy and Keller (2013: 215) concluded that “it appears that theories of syntactic complexity may need to posit memory costs which are a function of a speaker’s linguistic experience rather than fixed and universal”.

The proposal that parsing may be quite different in languages with different word orders was first made in Hawkins (1995), using a different kind of theory and data, namely distinct grammatical patterns in SVO and SOV languages, for which he hypothesized a distinct set of processing routines that could make sense of the grammatical differences (see Hawkins 1995, 2014).^[3] Specifically, Hawkins argued that speakers of SVO and SOV languages use distinct parsing strategies deriving from the different verb positions of the verb and from the verb’s activation of possible and preferred co-occurrences. In an SVO language, these co-occurrences will gradually be selected from when post-verbal material is parsed, and the preferences will be confirmed or disconfirmed. For SOV both activation and selection will occur simultaneously at the clause-final verb, and the verb will gather in its arguments by looking backwards rather than forwards in the string.^[4]

The parsing theory of Hawkins (1995) did not use the language of prediction (or surprisal) versus integration, since these were developed much later in the psycholinguistic literature, but the activation potential of the verb which he was assuming, and the possible time lag he proposed between activation and selection of a verb’s co-occurrences in SVO languages versus their simultaneity in SOV, meant that what he was proposing in more modern psycholinguistic parlance was essentially this: an SVO language activates and predicts one or more co-occurrence frames at the verb, then selects and confirms or disconfirms any preferred one subsequently; an SOV language activates its co-occurrence frames at the verb and uses them to select and integrate preceding items simultaneously with the processing of the verb. If we think in terms of Ferreira and Chantavarin’s (2018) proposed synthesis of prediction and integration models in psycholinguistics,^[5] the co-occurrences activated at the verb in SVO languages predict possibilities and preferences into which subsequent material can be integrated; in SOV these possibilities and preferences are not predictive at the verb, but instead serve to integrate the verb with its co-occurring arguments when the verb has rich semantic content.

Each of these VO and OV strategies has been argued to have advantages and disadvantages for processing (Hawkins 2014). An SVO language permits more prediction of possible and likely co-occurrences, but there is then additional prediction-related memory cost prior to the integration of post-verb material, and misassignments (and garden paths) can and do occur. An SOV language avoids these by integrating its verbs with already encountered arguments and by immediately removing unintended co-occurrence possibilities, but in the process, there is an online delay in assigning arguments to the verb (i.e., there is an “un-assignment” of argument structure contra the efficiency principle of Maximize Online Processing; see Hawkins 2014), plus there is additional argument-related memory load prior to the verb. These competing efficiencies in processing form part of the explanation for why VO and OV languages are both productive and roughly equally frequent across the globe.^[6]

In addition, they can have equally minimal and efficient domains for the processing of basic phrase structure (Hawkins 2004), or, using the theory of Dependency Grammar (Hays 1964; Hudson 1984), they can have equal Dependency Length Minimizations (Futrell et al. 2015).

The fundamental question for prediction and integration that has not yet been sufficiently addressed empirically is whether it is only the verb that can activate and predict its co-occurring phrases, or whether and to what extent noun phrases can also predict the verb, prior to its occurrence at the end of the clause. In an SOV order, any such prediction would have to be based on pre-verbal information, such as the rich lexical content of preceding nouns mentioned above (Polinsky and Magyar 2020), or the case marking and associated theta-role of an NP (see Bornkessel 2002; Bornkessel and Schlesewsky 2006), which may be compatible with the selectional and subcategorization restrictions of only a limited set of verbs. There is also the issue of how/when the discourse context could make a particular verb predictable, as could speakers’ real-world knowledge about the probabilities of certain events and linguistic knowledge about the frequencies of noun-verb co-occurrences in their previous linguistic experience. In other words, SOV language users are still plausibly engaging in some prediction and attempting to anticipate the verb, and if so, we might expect similar processing outcomes in VO and OV languages, suggesting that sentence processing patterns (and prediction) may not be primarily guided by the position of the verb.

The traditional head-driven parsing assumption is that only the verb activates and predicts the co-occurrences of that verb, and that object NPs or other strictly subcategorized phrases do not activate and predict the set of verbs with which they co-occur. But it may often be possible for semantically rich nouns or real-world knowledge and previous linguistic experience, or discourse context, to activate predictions regarding the appropriate actions that apply to certain entities that are encountered. Online predictions may then look similar to the grammatically and lexically based parsing predictions that derive from the asymmetrical structuring of lexical entries for verbs and nouns.

Our key hypothesis is that we should nonetheless see a greater and more systematic role for online prediction in an SVO language like English, compared with SOV Japanese, because these languages differ in the relative order in which object nouns and verbs are processed. Both sets of speakers will have prior real-world knowledge of events and of noun-verb co-occurrence frequencies, and both will presumably have similar discourse processing skills that enable them to integrate the current sentence into its discourse context. Both will also, we assume, have an asymmetrical lexical listing for nouns and verbs, whereby nouns are listed alongside verbs in the mental lexicon but not vice versa. The crucial, and plausibly only, difference between English and Japanese therefore involves the relative order of access to these two asymmetrically listed categories, and this should, we hypothesize, result in a more productive and systematic role for prediction in English than in Japanese.

The empirical questions addressed in this study are therefore: how predictive are object NPs of their following verbs, and does prediction actually work in both directions, from verb to object NPs, and from object NPs to verb, and to what extent does it do so in contexts with different co-occurrence constraints (high vs. low)?

2 The current study

Our key hypothesis was that prediction will be a more robust and regularly used parsing strategy in English than in Japanese, since the most predictive category, the verb, appears earlier in English than in Japanese. It is important to note that the current study does not manipulate discourse context. Instead, we examined only decontextualized sentences in order to test whether there is evidence for the differential use of prediction in the online processing of basic transitive sentences within these two language types.

Note also that we have chosen English and Japanese as representatives of the SVO (head-initial) and SOV (head-final) language types that are commonly assumed in language typology in order to test our hypothesis about the crucial role of verb position in relation to prediction in processing. English is a prototypical SVO language; it exhibits some variability with respect to the positioning of (auxiliary) verbs and subjects (Never have I seen such a thing), but not with respect to the verb and direct object, and there is universal agreement that SVO is its “dominant” order typologically (Dryer 2013b; Lehmann 1978), as well as its “basic” order in different models of grammar (cf. Brown and Miller 1996). Japanese is a prototypical SOV language (Kuno 1973, 1978), indeed a prototypically “rigid” SOV language (Greenberg 1963), with grammatically permitted albeit relatively infrequent reorderings of SO to OS before V (Yamashita 2002), and also some observed postposings of elements to the right of the verb that occur in Japanese conversation (Clancy 1982; Ono 2006; Whitman 2000). These two languages are well-placed typologically, therefore, to form the basis for this initial test of whether different verb positions are indeed associated with differences in prediction in language processing. If our hypothesis is supported, more fine-tuned studies can then be conducted on languages that are less consistently SVO than English, for example, Polish (Siewierska 1993) with its more variable word orders alongside a dominant SVO, and on languages that are less rigidly verb-final than Japanese and combine OV and VO orders. The different grammatically permitted positionings of the verb within these languages would provide a rich set of stimuli for within-language comparison and for further testing of our hypothesis.

To examine our key hypothesis, we conducted three experiments on English and Japanese. The first compared native speakers of English to native speakers of Japanese, using a cloze task. English participants heard SV_ and had to produce what they thought was the best (direct object) completion. Japanese participants heard SO_ and had to produce what they thought was the best (verb) completion. Because this experiment compared the production of nouns to verbs crosslinguistically, we believed that two further experiments were necessary to augment Experiment 1’s findings. In Experiment 2, we examined native speakers of English, who also completed a cloze task. One group read SV_ and had to provide the best written (direct object) completion. The other group read OS_ (using the it-cleft construction, see below) and had to provide the best-written (verb) completion. This allowed us to examine noun versus verb prediction within one of our languages. Following the key hypothesis, we expected to observe less prediction in the OS_ order (i.e., nouns are not (as) good predictors of verbs). Finally, in Experiment 3, we again compared native speakers of English and Japanese, but this time, unlike the cloze completions in Experiment 1 (which were different, i.e., nouns vs. verbs) we used direct translations, which meant that participants were presented with the exact same stimuli. In this case, English participants were presented with SV_ and had to produce a direct object. Japanese participants were presented with S_V and also had to produce a direct object. We expected that Japanese participants would show less evidence of prediction, even in this sentence type in which participants heard both S and V, since the processing routines associated with rigid verb-finality are not normally designed to use the verb as a predictor in online sentence comprehension but instead as an integrator of already encountered material.

3 Experiment 1

In Experiment 1, we utilized an online version of the cloze task in which participants heard a sentence fragment and had to complete the sentence as naturally as possible with the word that best completed the sentence. We refer to the task as “online” because in addition to assessing the cloze probability of the words produced, we also assessed the reaction time, that is, the voice onset time. Voice onset time is the duration from the end of the stimulus sentence fragment to the onset of the word that the participant produced.

The cloze task has been the gold standard psycholinguistic task used to assess prediction. Staub et al. (2015) showed a relationship between the cloze probability of words produced and voice onset time.^[7] In particular, there was a negative relationship, such that, higher cloze probability responses were associated with shorter voice onset times, indicating that more predictable responses were produced faster. Based on these findings, we hypothesized that Japanese speakers (SOV) would show differences compared to English speakers (SVO). More specifically, we expected that SOV speakers would produce either lower cloze probability responses with similar voice onset times or similar cloze probability responses but would be slower to do so. Both of these alternatives would be indicative of less prediction.

4 Experiment 2

Experiment 1 compared SO_ Japanese sentences to SV_ English sentences in a cloze task and showed that Japanese speakers engaged in less prediction. More specifically, across the full set of sentences, cloze probabilities were equal but Japanese speakers had significantly slower reaction times. When analyses focused on the most predictable sentences, Experiment 1 found both lower probability responses and longer reaction times. However, this study leaves open some alternative explanations (e.g., nouns may be easier to produce than verbs, cultural differences in background knowledge, etc.). Thus, in Experiments 2 and 3, we attempted to narrow down and reduce the number of differences in our comparisons so as to try and confirm our interpretation of the results of Experiment 1.

In Experiment 2 we accordingly set out to examine differences in prediction within one of our languages, English and we compared the production of nouns to the production of verbs in this language. In order to do so, we took advantage of the English it-cleft sentence structure (e.g., It was the match the ailing team ______), and compared it to (e.g., The ailing team forfeited the _______). The rationale for the hypothesis (i.e., that less prediction would be observed in OS__) is based on the idea that nouns are not (as) good predictors of verbs, as verbs are of nouns. Second, we also obtained new materials from Staub et al. (2015), which gave us 50 critical items. Half of these were deemed by the original authors to be high constraint (cloze probability > 0.50) and half were low constraint (cloze probability < 0.50). The OS_ sentences were created by taking the highest cloze probability noun from the SV_ completions and inserting it into an it-cleft sentence and removing the verb. All sentences were simple transitive sentences, except that they had an adjective prior to the subject noun (see Section D, Supplementary materials). We had intended to run the study in the laboratory, as in Experiment 1. However, very early in the data collection process, the global pandemic hit. Thus, the experiment was conducted on Qualtrics, which prevented the collection of reaction times. Thus, Experiments 2 and 3 only analyzed mean cloze probabilities.

5 Experiment 3

In Experiment 3, we compared SV_ English sentences to S_V Japanese sentences. The rationale for doing so was to consider the extent to which crosslinguistic differences in processing are observed when the stimuli are the same. The key hypothesis underlying this study is that prediction is generally more difficult in head-final languages, primarily due to the position of the verb, and that the verb serves to integrate material that precedes it and that has already been processed rather than to predict what lies ahead and has not yet been encountered. Because of this typological difference, we expect that Japanese speakers will have developed general processing routines for the integration of prior material and engage in less prediction than English speakers, but that Japanese speakers do nonetheless engage in some prediction on the basis of pre-verbal material. Given these assumptions, it is an empirical question whether Japanese speakers will produce equal cloze probability responses to those found in English when they are given a subject noun and a verb (and ample time), or whether their usage of a language with general processing routines that have responded to the verb-final grammar and afford less prediction will mean that they do not engage in prediction to the same extent as English speakers, even when stimuli contain the same items for cloze probability completions. Accordingly, we used the same materials as in Experiment 2, and translated them directly into Japanese. The experiment was again conducted on Qualtrics. English speakers provided a noun for an SV_ sentence fragment and Japanese speakers provided a noun for an S_V fragment with its missing pre-verbal direct object.

Before presenting the details of this experiment, some further clarification and an analogy may be helpful about the logic that underlies it. It is becoming increasingly clear in crosslinguistic processing that speakers of structurally different languages adjust their general processing routines to the grammars of the relevant languages and they tolerate e.g., more versus fewer center embeddings and incomplete noun-verb dependencies (cf. Gildea and Temperley 2010; Levy and Keller 2013 on German vs. English), or, as we are arguing here, they develop different strategies for the efficient processing of the verb and its arguments. We believe that languages with early verbs in the clause encourage a general processing strategy in which speakers predict the different options that lie ahead and assign certain probabilities to them based on context, real-world knowledge, structural simplicity, and so on. Verb-final clauses, on the other hand, lead to different routines in which the co-occurring items are already there, and what the verb needs to do is gather them in, integrate them with the relevant co-occurrence frame, check for compatibility, and so on. This same integration routine applies when a verb’s argument occurs in an earlier sentence in the text and is “deleted” from the current clause being processed, which is grammatically possible in Japanese. There is no prediction of numerous unencountered possibilities here, therefore. To use an analogy, the English strategy is that of a fisherman sitting beside a muddy English pond and having to predict what kind of fish are likely to be in there and what kind of hook, bait and net might be needed to catch something. The Japanese strategy, by contrast, is that of a fisherman who can clearly see the koi swimming in the clear water before him and who then adjusts the size of his net and the length of his pole to make them compatible with gathering in the fish that he already sees and wants. If your verb is early, the processor adjusts to making regular predictions; if it is late, it adjusts to regular integrations of previous material, i.e., two very different processing routines, albeit both ultimately deriving from verbal co-occurrence frames in the mental lexicon and from the online activation of these when the verb is encountered. The issue in Experiment 3, then, is whether we can see any evidence, even when the stimuli for a cloze probability completion test are the same, for these two different general processing routines of English and Japanese.

5.6 Discussion

This experiment showed the most consistent pattern of all three experiments. We observed clearly and significantly higher cloze probability responses for English speakers as compared to Japanese speakers, and the same finding held for all analyses. Here we compared the production of object nouns across the two languages, based on the same subject noun phrase and verb (i.e., direct translations). In this experiment compared to Experiment 1, we hypothesized that the main difference between English and Japanese would be due to the general tendency of Japanese speakers to engage in less prediction, given their (dominant) experience with a head-final language. The results were consistent with this assumption.

There are two additional points that we think are important to raise here. First, in examining the results of this experiment, we observed two or three items that showed clear cross-cultural differences. One example was The ailing team forfeited _____. The modal response in English was the game (cloze probability = 0.33) and the modal response in Japanese was fighting spirit (cloze probability = 0.37). Thus, the verb forfeited would seem to be more psychological in Japanese than in English. We do not view these differences as a major issue for our conclusions because (1) they affect a small number of items, approximately 10 %, and (2) there is as much variability within language as between languages (see Section E, Supplementary materials). The whole notion of cloze probability, and specifically higher cloze probability, is that the more consensus or consistency in responses, the more predictable the sentence fragment is. Our analyses average across items and participants to assess whether, at a more coarse-grained level, there is more or less evidence for prediction. Nuanced differences are common even within a given language and are not central to the main arguments.

The second point concerns a slightly different way to analyze the data from this experiment. Up to this point, we have focused on mean cloze probabilities (averaged across participants) and the mean cloze probabilities of the modal responses (i.e., the most frequently issued response for a particular item), again averaged across participants. However, we also examined the cloze probabilities (by item) of the second most frequent responses, the third most frequent responses, the fourth most frequent responses, and a final category that contained only singleton responses (see Figure 1). What these item results nicely show is that in high-constraint items, Japanese participants produce significantly (1) fewer modal responses (i.e., fewer “first” responses) and (2) more singletons. In contrast, in low-constraint items, the only significant difference is in the number of “other singletons”.

Figure 1:

Mean cloze probability averaged across items. The left panel shows results from high-constraint items and the right panel shows results from low-constraint items. Error bars show the standard error of the mean.

6 General discussion

The focus of the current paper has been on crosslinguistic differences in prediction, specifically with respect to verb-object relations. We hypothesized that there would be less evidence for prediction as a parsing strategy in head-final (SOV) Japanese than in SVO English since there is an asymmetry in the mental lexicon between verbs and nouns: the former always make explicit reference to co-occurring arguments, which consist regularly of nouns, whereas nouns do not generally make explicit reference to their accompanying verbs. In a verb-early language, these co-occurrences will be activated early in online processing, whereas in an SOV language, they will be activated late, when the verb is encountered. Nouns clearly do make some predictions for upcoming verbs in verb-final languages, and there is typological evidence for higher ratios of nouns to verbs in the lexicons of these languages, which has been argued to provide informationally richer and more predictive semantic processing online (Polinsky and Magyar 2020). But the predictive potential of this higher noun-to-verb ratio is, we argue, weaker and less systematic than that provided by the basic asymmetry between verbs and nouns and by their differences in linear ordering.

Our three experiments provided support for this key hypothesis that there would be more prediction in English compared to Japanese overall, despite the fact that there is clearly some prediction taking place from nouns to verbs in Japanese, as seen in the cloze probabilities for this language. Head-final languages place the verb at the end of the sentence, so activation and selection of verb co-occurrence frames occur only after the arguments that select the relevant co-occurrence frame have already been processed. The verb then serves an integrating function gathering in this previous material, checking it for compatibility with its listed co-occurrences in the mental lexicon, and does not predict a range of possibilities that have not yet been encountered. By contrast, in English, the head comes earlier, so prediction of an upcoming object is routinely possible.

Another relevant factor in all of this is the speaker’s experience actually speaking and hearing a head-initial or head-final language. If an individual is a native speaker of a head-final language (and prediction is not as robust a parsing strategy in that language), then these individuals will tend to show less prediction even when given a noun and a verb (as was shown between languages in Experiment 3). Future studies would now be valuable that examine languages with more variable orderings of S, V and O (e.g., Slavic languages like Polish) and with more variable verb-finality than Japanese, i.e., non-rigid SOV languages (Greenberg 1963), in order to see whether languages with less “dominant” SVO and SOV still show the same or similar differences with respect to prediction that we have shown here across English and Japanese and within English.

Moreover, in this study, we tested Japanese native speakers who were also bilingual in English and studying at a primarily English-speaking university (i.e., an immersion situation). At this juncture, we do not know exactly whether and how proficiency, as well as frequencies and usage patterns in a head-initial language, affect prediction ability in a native head-final language, and vice versa (but see Filipović and Hawkins 2013, 2019; Filipović 2019 for a multi-factor model of bilingualism that proposes ways to address such questions). In the context of the current study, it would be worth testing monolingual Japanese speakers situated in Japan who do not have any knowledge of another, head-initial language. However, if Japanese bilinguals in an immersion situation show the effects we have reported in this paper, then monolinguals would almost certainly show even greater differences compared with native English monolinguals than we have reported here.

We believe that the results across all three experiments support our key hypothesis. In the remainder of the General Discussion, we focus particularly on the cloze probability data from the high-constraint items (both overall means and means from modal responses). The expectation was that English SV_ should show the highest cloze probabilities (i.e., the most evidence of prediction). This is what we found. The results across experiments are moderately variable (0.54–0.75), but this is in part due to some clear differences between our experiments (e.g., different items and different modalities between Experiment 1 vs. Experiments 2 and 3). When we examined the comparisons between English SV_ and the other three comparison groups, we observed lower mean cloze probabilities and the differences ranged from 0.09 to 0.16, and all resulted in large effect sizes (see Table 9 for a summary of the data).

Finally we pooled the data from Experiments 2 and 3, and ran a linear mixed effects model on them. We acknowledge issues involved in between experiment comparisons and primarily because in Experiment 2 one group of (native) English speakers produced verbs in OS_ structures and in Experiment 3 the (native) Japanese speakers produced objects in S_V structures. Thus, both groups were engaged in tasks that were atypical of their presumed dominant language experience. We combined the data (from Experiments 2 and 3) and coded one variable based on the category of the response provided (object vs. verb), and another based on language (English vs. Japanese). If there is an effect of response category, it would be driven by the English speakers (Experiment 2 OS_ order), and if there is an effect of language, it would be driven by Japanese speakers in which they provided an object in response to a subject and a verb (Experiment 3 S_V order). This is actually a strong test of our key hypothesis because (1) an effect of language would be driven by a situation in which Japanese speakers produced a direct object (based on a subject and a verb, which should lead to greater prediction), and (2) an effect of response category would be driven by English speakers, who produced a verb (based on a direct object and subject, which should lead to less prediction). We ran a linear mixed effects model in R (version 3.6.1), similar to the previous analyses. The results are shown in Table 10. Both variables are significant. However, the effect of language was larger than the effect of the response category. Again, this result is consistent with our key hypothesis.

Before addressing further theoretical implications and conclusions from this study, there are some alternative explanations that can be excluded based on prior research. Staub et al. (2015) demonstrated that the relationship between cloze probability and reaction time was not due to (1) the frequency of the words produced or (2) semantic associations between words in the sentence and the words produced. Instead, the only variable that produced a significant effect in terms of lower reaction times and higher cloze probability responses was the length of the sentence fragment. In our study, this is not a concern because the sentences were direct translations. Therefore, if the main patterns of results are not due to low-level lexical factors or semantic associations, then as Staub et al. concluded, predictability in sentence contexts is derived from the combinatorial nature of incremental parsing, combined with “background” real-world knowledge.

7 Limitations

We see three main limitations in the current study. First, we were not able to collect reaction time data for Experiments 2 and 3, and thus, this is an important next step in this line of research. At the same time, it is important to note that the use of “online” studies via the internet is a potential avenue forward, and greatly reduces the amount of work required for assessing prediction in monolingual speakers across the world, particularly if reaction time data could be also collected. Second, our Japanese speakers were all proficient in English as a second language, and it is currently unknown whether proficiency in an SVO language might impact prediction in an SOV language. This issue could be addressed by testing monolingual speakers of Japanese and comparing them to bilinguals. Third, our Japanese materials were translations from English, and therefore, there may be some unknown (lexical or cultural) factors in which the translation is not equivalent. We think that these potential issues are on the minor side, and we carefully examined responses on a by-item level for anomalies. Finally, in the current study, we made no attempt to manipulate discourse context. The goal of this initial set of studies was to assess whether prediction operates differently in these two languages in decontextualized situations. It is an open question whether there are or are not differences in prediction between English and Japanese based on contextual factors.

8 Theoretical implications and conclusions

We have argued that SVO languages, like English, and SOV languages, like Japanese, have both advantages and disadvantages for processing. SVO permits more prediction of possible and likely co-occurrences at the verb, but there is then an additional working memory cost in online processing before possible co-occurrences from the mental lexicon are eliminated by post-verbal material, and there are also misassignments (and garden paths) that can result from these possibilities before they are eliminated. An SOV language like Japanese avoids these by integrating its verbs with already encountered arguments and by immediately removing unintended co-occurrence possibilities. But in the process, there is an online delay in assigning arguments to their verbs, and there is less actual online prediction.^[9]

These competing efficiencies in processing form part of the explanation that Hawkins (2004, 2014 has given for why SVO and SOV languages are both frequent across the globe (see fn.6). This also follows from his earlier calculations (see Hawkins 1990, 1994) that SVO and SOV have equally minimal and efficient domains for the processing of basic phrase structure and head ordering.

Our argument, based on typological distributions for different verb positions (see again fn.6) and on the experimental results reported here, is that prediction is not necessarily the key factor that facilitates ease of processing in language (see also Onnis et al. 2022 for a similar conclusion). We have presented arguments that predictions add to working memory and may be incorrect, and that the world’s languages show no evidence of favoring basic word orders that would optimize prediction by positioning verbs and other heads first or even early in their respective phrases. But prediction is certainly one of the factors that make processing easier, and in particular it appears to be more relevant for the “look ahead” SVO type, like English, and less relevant for “look back” SOV languages like Japanese. More generally we advocate a model of competing motivations in both processing and grammar, along the lines of MacWhinney et al. (2014), within which we see predictive processing as one of several cooperating and competing forces that determine both ease of processing, as evidenced by experimental and corpus results, and also preferences for grammatical conventions and their crosslinguistic distributions (see the Performance-Grammar Correspondence Hypothesis of Hawkins 2004, 2014 and fn.3 above). The purpose of this paper was to show the greater role of prediction in an SVO language like English compared to Japanese.