Glossa Psycholinguistics

1. Introduction

One of the central issues in the research on sentence processing has been parsing choices for locally structurally ambiguous input, as, for example, in (1). (1) illustrates the object/subject ambiguity, in which the underlined noun phrase (NP) the sentence can be attached in two ways: it can serve as the direct object of the verb in the adverbial clause, as illustrated in (1a); or it can be attached as the subject of the main clause, as illustrated in (1b). Accordingly, this NP is ambiguous with regard to which predicate it receives its thematic role from, i.e., the embedded verb or the main verb.¹

Numerous studies (e.g., Adams et al., 1998; Ferreira & Henderson, 1991; Frazier & Rayner, 1982; Staub, 2007) observed a Garden Path (GP) effect in sentences such as (1b), namely, slower reading times at the disambiguating main verb became, compared to an unambiguous baseline. This effect suggests that the ambiguous NP is initially attached at the direct object position of the subordinate verb, necessitating reanalysis upon the arrival of the main verb.

An early, influential suggestion made by Frazier (1978, 1987a) to account for the parsing of local ambiguities was that, at points of ambiguity, initial attachment choices are guided by syntactic considerations alone, with no role for lexical, semantic or contextual information. Parsing at this stage is determined by two purely structural principles: Minimal Attachment, namely, building the least complex structure possible; and Late Closure, namely, attaching incoming material to recent constituents. According to Minimal Attachment and Late Closure, in object/subject ambiguities, such as (1), the NP is attached as a direct object, as this obeys Minimal Attachment (gives rise to the simplest structure) and Late Closure (the NP is attached to a recent constituent). In this article, we do not discuss Minimal Attachment further, but rather focus on Late Closure, which is relevant for the present data.

An alternative to this strategy based on structural principles holds that the processor’s aim is to satisfy grammatical relations in such a way as to achieve maximal interpretation during incremental processing (e.g., Crocker, 1992; Frazier & Clifton, 1996; Kamide et al., 2003; Pritchett, 1988, 1992; Sturt & Crocker, 1997). According to this view, attachment decisions should be defined over thematic structure, rather than constituent structure, with the aim of maximizing thematic role assignment. For instance, a widely observed phenomenon where comprehenders posit a gap for a fronted wh-phrase at the earliest available position (What did you say that Mary was reading ___?) was first explained by a purely structural principle, i.e., the Active Filler Strategy (de Vincenzi, 1991; Frazier, 1987b). In contrast, Pritchett (1992) and Gibson, Hickok, and Schütze (1994) argued that no special strategy is required; instead, the need to assign thematic roles as soon as possible guides such filler-gap formation. While, in English, these two alternative proposals make similar predictions, Aoshima, Phillips, and Weinberg (2004) showed that the predictions differ in Japanese, a verb-final language. Aoshima et al. (2004) examined sentences with embedded clauses in Japanese, in which a fronted dative wh-phrase can, in principle, be associated either with a gap in an embedded object or main object position (e.g., which student-dat class teacher-top (____gap1) [principal-nom (____gap2) book-acc read] library-at librarian-dat told. ‘The class teacher told the librarian at the library which student the principal read a book for.’). The main gap position is the closest to the filler structurally, whereas the embedded position allows the earliest thematic role assignment. The authors found that Japanese speakers preferred to interpret the fronted wh-phrase as if it was displaced from the embedded clause, meaning that the processing of filler-gap dependencies is driven by the need to satisfy thematic requirements, rather than by purely structural principles.

In the current study, we further test the degree to which attachment decisions are guided by the pressure to maximize thematic relations and, hence, interpretation, during processing. In examining thematic assignment maximization, we use the implementation of this requirement formulated by Pritchett (1988, 1992). According to Pritchett, parsing performance crucially depends on grammatical competence, meaning that, at every point during processing, a global principle of grammar – the Theta Criterion – drives the parser, through its local application. Accordingly, he proposes the Theta Attachment principle, given in (2).

As it turns out, in most structures with local ambiguity examined to date, maximization of thematic assignment as per Theta Attachment (namely, maximal interpretation) is confounded with attaching incoming input to recently built structure (Late Closure). For example, the object reading in object/subject ambiguities, as in (1), will be preferred both according to Late Closure, as the NP is attached to a recent constituent, and according to the principle of maximizing interpretation, as this structure entails assignment of a thematic role to the NP by the adverbial clause verb, and enables the reader to interpret this NP as part of a proposition (‘Mary was reading the sentence’). Thus, previous studies of local ambiguities could not tease apart structural vs. thematic considerations in attachment choices.

However, there are structures with local ambiguity in which the predictions of structural vs. interpretational principles differ. Consider the local ambiguity exemplified in sentences (3–4) (Pritchett, 1992).

In these sentences, the main verb is a ditransitive verb (here, gave), taking both a goal and a theme argument; the goal argument is modified by a subject relative clause which includes an optionally transitive verb (here, eat). In such sentences, when the NP following the second verb (the fudge) is encountered, it can, in principle, be attached in two ways, as illustrated in (5). It can either be attached to the recent, optionally transitive verb inside the relative clause, with the first verb awaiting its third argument, as in (5a); in this case, another NP has to appear for the sentence to be grammatical, as in (4) above. Alternatively, the ambiguous NP can be attached to the first, ditransitive verb in the main clause, in which case the second verb will be interpreted intransitively, as in (5b). In this case, the sentence is grammatical without additional material. We will call this type of ambiguity main clause / relative clause (MC/RC) ambiguity, as the ambiguous NP can be analyzed as either an argument in the main clause, or an argument in the relative clause.²

For the type of sentence exemplified in (5), models which make predictions based on structural vs. thematic considerations differ. Structural models which highlight locality of attachment, including Late Closure, predict invariable low attachment of the ambiguous NP, as in (5a). In contrast, according to thematic assignment maximization principles, such as Theta Attachment, there is optionality with regard to the attachment of the post-verbal NP. This is so since when this NP is encountered, it can receive a thematic role either from the matrix verb gave or from the embedded verb eat. Since the NP receives an interpretation in both analyses, no preference is predicted, and a chance distribution of attachment choices is expected.

Note that, at least according to the Theta Attachment principle of Pritchett (1988, 1992), the fact that the Theme role is obligatory for the higher, ditransitive verb (gave), whereas it is optional for the lower verb (eating), is not relevant at the point of ambiguity. Such global considerations do not have a role in the initial attachment decision. The only relevant factor is maximization of thematic relations, namely, whether the NP can receive a thematic interpretation. This is in contrast with Gibson (1991), according to which assignment of obligatory thematic roles is preferred over assignment of optional ones. In particular, the model presented in Gibson (1991) incorporates both thematic and structural considerations. This model assumes a parallel parser, where structures are evaluated against each other using (a) thematic assignment maximization in accord with the Theta Criterion, specifically, the assignment of obligatory thematic roles, and (b) locality, with a preference for local attachments over more distant ones. According to this model, in the MC/RC sentences, the structure with local attachment will satisfy locality, but not thematic assignment, as it leaves the main verb with an unassigned obligatory theta role; in contrast, in the structure with non-local attachment, all obligatory thematic roles are assigned, but locality is not satisfied. Therefore, the structures with local and non-local attachment will have the same weight, predicting, similar to Pritchett’s Theta Attachment, a chance distribution of attachment choices.

Importantly, the theories of thematic assignment described above, and, specifically, Theta Attachment, view thematic roles categorically: is there a role that can be assigned? However, other theories hold a probabilistic view of thematic assignment, i.e., they consider how common it is for a particular predicate to have an argument. Therefore, a factor whose effect can be tested here is verb bias, namely, the statistical information reflecting verbs’ subcategorization preferences.

Note that if verb biases drive attachment choices, then the transitivity bias of the embedded, optionally transitive verb should play a role in both types of ambiguity that we considered, namely, object/subject and MC/RC ambiguities. For object/subject ambiguities, as in (1), if processing is determined by verb biases, then the ambiguous NP will be preferentially attached as object for transitively biased adverbial clause verbs (e.g., read or drink), whereas it will be preferentially attached as subject if the adverbial clause verb is intransitively biased (e.g., dance). For MC/RC ambiguities, as in (3) or (4), the preference for analyzing the ambiguous NP as the object of the optionally transitive verb inside the relative clause should increase for more transitively biased relative clause verbs.

Evidence regarding the effect of verb bias in object/subject ambiguities is limited. In an eye-tracking study, Pickering et al. (2000, Experiment 3) found that participants adopted the object analysis of the ambiguous NP in object/subject ambiguities, even when both prior context and subcategorization frequencies supported the subject analysis (as in While the pilot was flying the plane/the horse that had arrived stood over by the fence). Relevant findings are also reported in Itzhak et al. (2010). In this event-related potentials study, participants listened to object/subject ambiguous sentences with no prosodic breaks, or to unambiguous sentences with such breaks. The experiment also manipulated the transitivity bias of the subordinate verb: while all the verbs were optionally transitive, half of the verbs were transitively biased, and the other half were intransitively biased. Itzhak and colleagues found a larger P600 garden path effect at the disambiguation region for transitively biased, compared to intransitively biased, verbs, suggesting a higher proportion of object attachment in the former. Crucially, though, there was still a P600 effect in sentences with intransitively biased verbs (compared to the unambiguous baseline), which means that the verb’s bias did not fully prevent the attachment of the ambiguous NP as an object to the adverbial clause verb. More recently, Huang et al. (2024), in a large-scale self-paced reading study that tested seven structures and included 2000 participants, found no correlation between verb bias and the size of the garden path effect in sentences with the object/subject ambiguity. In particular, in object/subject sentences, every item resulted in a garden path effect that ranged from 59 to 258 ms, but, crucially, this item-level variability was not explained by verb subcategorization bias (cloze-based or corpus-based). These findings all seem to suggest, in line with Pickering et al.’s (2000) conclusion, that while verb bias may have an effect on processing, it is not strong enough to guide initial parsing choices, which are determined by other principles, whether structural (e.g., Late Closure) or interpretational (maximizing thematic assignment).

In contrast to object/subject ambiguities, there are no experimental studies of the MC/RC ambiguity that we know of. Pritchett (1992, p. 114) only reports informal off-line investigations, asking English-speaking participants for their judgments on processing difficulty in sentences such as (3) and (4). For both types of sentences, Pritchett notes that “…hearers appear to judge these sentences either as unprocessable or unproblematic. In other words, some individuals are led down the GP, some aren’t”. This result aligns with his predictions: when the ambiguous NP is attached to the low verb, a sentence such as (4) will be judged as unproblematic, and a sentence such as (3) will be judged as difficult. When the NP is attached to the high verb, the opposite pattern will emerge. Thus, for each sentence type, both types of judgements should be attested.

In the current study, we examine the MC/RC ambiguity, which enables us to tease apart structural and thematic considerations in attachment choices. We compare the MC/RC ambiguity to the object/subject ambiguity in the same participants and with the same verbs, to see whether there is a dissociation in attachment decisions between the two structures.

2. Experiment 1

The aim of Experiment 1 was to examine whether thematic assignment maximization guides processing, or whether attachment choices are guided by structural considerations (i.e., Late Closure) or by verb bias. To do this, we manipulated ambiguity type, contrasting object/subject (obj/subj) ambiguities (6) with main clause/relative clause (MC/RC) ambiguities (7). Two possible completions, each a two-word string in Hebrew, were provided for each fragment. One of the completions indicated local attachment of the ambiguous NP (as object in obj/subj ambiguities, and as a complement to the relative clause verb in MC/RC ambiguities), and the other indicated non-local attachment (as subject in obj/subj ambiguities, and as a complement to the main clause verb in MC/RC ambiguities). We also varied the transitivity bias of the optionally transitive verb in both ambiguity types. In (6–7), the ambiguous NP is underlined, and the optionally transitive verb is doubly underlined.

As explained above, Late Closure predicts that completions indicating local attachment will be preferred in both obj/subj and MC/RC ambiguities. Processing guided by verb biases predicts that completion preferences in both ambiguity types will be correlated with the optionally transitive (OpT) verb’s transitivity bias. According to the Theta Attachment principle, however, performance will differ between the obj/subj and the MC/RC conditions. In the obj/subj condition, the need to satisfy the Theta Criterion given the maximal thematic grid of the verb will invariably lead to local attachment of the ambiguous NP, regardless of the verb’s transitivity bias, as only in this structure will the NP receive a thematic role. In contrast, in the MC/RC condition, either the local or the high attachment option can be chosen, as both satisfy the Theta Criterion to the same degree. Thus, both types of completions may be chosen. In addition, we hypothesized that when optionality in thematic role assignment arises (namely, in the case of MC/RC sentences), transitivity bias will become relevant, affecting processing choices, such that local attachment will be chosen at a higher rate for more transitively biased OpT verbs. The predictions of the different accounts are summarized in Table 1 below.

2.3 Results

The overall rates of completions (after exclusion of trials with RTs over 6 s) are presented in Figure 1a for control conditions, and Figure 1b for experimental conditions (for data without exclusion, see Table A1 in Appendix A).

The control conditions behaved as expected, exhibiting uniform attachment choices in line with the grammatical structure of the sentence, i.e., local attachment in the MC/RC control condition, and non-local attachment in the obj/subj control condition. Mean accuracy in the control conditions for the participants included in the analysis was 95.3%.

Figure 1a: Rate of completions in control conditions, Experiment 1.

As for the experimental conditions, the obj/subj condition overwhelmingly exhibited local attachment, whereas in the MC/RC experimental condition, local and non-local attachments were chosen at similar rates.

Figure 1b: Rates of completions in experimental conditions, Experiment 1.

The by-item average rates of local attachment in the experimental conditions are presented in Figure 2, representing the effect of verb transitivity bias on the rate of choosing local attachment in each sentence type.

Figure 2: The effect of transitivity bias on the rate of choosing local attachment, Experiment 1.

Table 3 and Table 4 summarize the results of the frequentist and Bayesian analyses, respectively. The frequentist analysis showed an effect for ambiguity type, such that the rates of local attachment in obj/subj sentences were significantly higher than in MC/RC sentences (p < .001). The interaction between ambiguity type and verb bias was not significant. However, the follow-up model with nested slopes showed that in the obj/subj sentences, there was no effect of transitivity bias on local attachment rate (p = .627), whereas in the MC/RC condition, the effect of transitivity bias on local attachment rate was significant (p = .007), indicating a higher local attachment rate with more transitively biased OpT verbs.

Table 3: Results of the frequentist analysis, Experiment 1: Estimates, standard-error, t-values and p-values.

	Estimate	SE	t	p
Grand Mean (Intercept)	1.145	0.182	6.308
Effect of ambiguity type on completion choice obj/subj > MC/RC	1.994	0.333	5.972	<0.001***
Overall effect of verb transitivity bias on completion choice	0.012	0.006	1.856	0.064
Interaction between verb bias and ambiguity type	–0.014	0.011	–1.299	0.193
Effect of verb bias in the obj/subj condition	0.005	0.010	0.485	0.627
Effect of verb bias in the MC/RC condition	0.019	0.007	2.712	0.007 **

The results of the Bayesian analysis were consistent with those of the frequentist analysis. There was an effect of ambiguity type, reflecting a decrease in the proportion of local attachment in the MC/RC condition, relative to the obj/subj condition: the mean of the posterior distribution was 88%, with a CrI of [83, 93], in the obj/subj condition, and 50%, with a CrI of [41, 62], in the MC/RC condition. We consider the effect reliable, as the 95% credible interval for this contrast does not cross zero. In this analysis, the overall effect of verb bias was not reliable, nor was the interaction. However, similarly to the frequentist analysis, the nested analyses showed a reliable effect of verb bias in the MC/RC condition, such that more transitively biased verbs led to more local attachment, but no reliable effect in the obj/subj condition.

Table 4: Results of the Bayesian analysis of question 1, Experiment 1: Posterior means on the log-odds scale (with 95% Bayesian CrI).

	Log-odds scale posterior mean	[CrI]	Posterior beyond zero
Grand Mean (Intercept)	1.197	[0.803, 1.622]
Effect of ambiguity type obj/subj > MC/RC	2.052	[1.399, 2.721]	100%
Overall effect of verb transitivity bias	0.014	[–0.005, 0.035]	94%
Interaction between verb bias and ambiguity type	–0.02	[–0.056, 0.015]	90%
Effect of verb bias in the obj/subj condition	0.005	[–0.016, 0.028]	70%
Effect of verb bias in the MC/RC condition	0.022	[0.004, 0.043]	99%

As for the difference from chance in each ambiguity type, the binomial tests showed that in the MC/RC condition, the by-participant average choice of attachment was not significantly different from the null hypothesis of a chance distribution (t = .7, df = 57, p = .48), whereas in the obj/subj condition, the by-participant average choice of attachment was significantly different from the null hypothesis, signifying that the results are inconsistent with a chance distribution (t = 16, df = 57, p < .0001).

3. Experiment 2

The main goal of Experiment 2 was to test whether the overwhelming local attachment found in Experiment 1 for sentences with object/subject attachment ambiguity can, indeed, be attributed to Theta Attachment alone, or whether some type of locality consideration is taken into account as well. In particular, we asked whether local attachment will be chosen at a lower rate when there is distance between the verb and the subsequent NP, compared to when they are adjacent. Possibly, when the verb is not adjacent to its argument, the role of Theta Attachment decreases. Additionally, if Theta Attachment is, indeed, less decisive in such cases, it is also possible that, in these cases, we will see an effect of transitivity bias, such that local attachment will be chosen at a higher rate for more transitively biased OpT verbs.

To test the effect of distance on thematic assignment maximization, we used Hebrew counterparts of obj/subj ambiguity sentences such as (9), in which an adverb is present between the verb and the NP.

In English, the presence of the adverb at home after the verb leads to early closure of the adverbial clause, before the NP pizza, because in English, a direct object has to be adjacent to the verb (Chomsky, 1981). Indeed, previous experiments in English that examined sentences with the obj/subj ambiguity used an adverb in baseline sentences on a par with a comma (e.g., Adams et al., 1998; Mitchell, 1987). Unlike English, however, Hebrew allows a word order where an adverb appears between a verb and its direct object. This enables us to explore whether and to what extent the distance between the optionally transitive verb and the ambiguous NP modulates attachment choices.

Distance between verbs and arguments is a relatively robust cause of sentence processing difficulty. This has been proposed in different models, starting with Frazier’s (1987a) Late Closure (see also Visibility in Frazier & Clifton, 1998) and was variously attributed to activation decay (e.g., Lewis & Vasishth, 2005), interference by similar items (cue-based retrieval; Lewis & Vasishth, 2005, and subsequent work), or the introduction of new discourse referents (Gibson, 2000; see also Van Dyke & Lewis, 2003). This makes it possible that the parser avoids combining a verb and an argument over a distance, even at the cost of not assigning a thematic role which could otherwise be assigned. Namely, these theories predict lower rates of local attachment when an adverb is present, compared to when it is absent. A similar prediction is derived if parsing decisions are based on statistical information regarding the distribution of adverbs. Although adverbs may appear between the verb and its object in Hebrew, they are rare in this position, and more frequently appear at the end of the verb phrase. Thus, statistically, the probability that the ambiguous NP opens a new clause is larger when it is preceded by an adverb than when it follows the verb directly.³

In contrast, according to thematic assignment maximization approaches, the object attachment is preferable, as it entails assignment of a thematic role, regardless of distance (in Hebrew, where this is grammatically possible). This would result in robust local attachment whether an adverb is present or not. Theories which assume guidance by verb bias predict that the transitivity bias of OpT verbs should play a role both when an adverb appears in the sentence and when it does not, with more local attachment for more transitively biased verbs, as predicted for Experiment 1.

As we noted in Experiment 1, one theory which explicitly includes both thematic considerations and locality is Gibson (1991). According to this theory, in obj/subj sentences, when there is no distance between the verb and the ambiguous NP, the structure with local attachment will be strongly preferred (as already explained in Experiment 1). However, in sentences with increased distance between the verb and the NP, recency will decrease slightly and, therefore, the structure with local attachment will receive a slightly weaker preference, although it will still be preferred.

In the experiment, we manipulated the distance of the optional direct object from the preceding verb by adding two adverbial phrases between them. We also manipulated the transitivity bias of the verbs, as in Experiment 1. Two possible completions were provided, one indicating local attachment (where local means attachment as an object, as in Experiment 1), and the other, nonlocal attachment (namely, attachment as the subject of the upcoming clause). Thus, the experiment included two conditions, both with the same structure (obj/subj ambiguity), differing only in the presence or absence of the adverb (see Table 5 for a translated example, and Table 6 for a Hebrew example).

3.3 Results

Eight out of 936 responses (less than 1%) were late responses and, therefore, were not available for analysis. Completion choices and correlation with verb bias are presented in Figures 3 and 4 (for additional data, see Appendix C).

Figure 3: Rate of completions by condition, Experiment 2.

The results of the frequentist and Bayesian analyses are summarized in Tables 7 and 8, respectively. The results show that in both the Adverb and No Adverb conditions, participants principally chose local attachment. However, the difference between Adverb and No Adverb sentences in the rate of selecting a local attachment was significant, with higher rates in the No Adverb sentences, as indicated by both the frequentist and the Bayesian models. In the Bayesian analysis, the mean of the posterior distribution was 90%, with a CrI of [85, 94], in the No Adverb condition, and 80%, with a CrI of [71, 88], in the Adverb condition. The frequentist model also showed a significant interaction between adverbial intervention and transitivity bias (p = .04), such that the insertion of an adverb resulted in a slight, non-significant, positive relation between the rates of local attachment and the verb’s transitivity, whereas in sentences without an adverb, the (non-significant) relation was in the opposite direction. However, this interaction did not turn out to be reliable in the Bayesian analysis, as its 95% CrI crossed 0.

Figure 4: By-item proportion of choices of local attachment in each condition as a function of verb transitivity bias, Experiment 2.

Table 7: Results (on the log-odds scale) of the frequentist analysis of Experiment 2: Estimates, standard error, t-values and p-values.

	Estimate	SE	t	p
Grand Mean (Intercept)	1.715	0.195	8.791
Main effect of adverbial intervention Adverb < No-Adverb	–0.815	0.233	–3.504	<0.001***
Overall effect of verb transitivity bias	–0.003	0.008	–0.382	0.702
Interaction between adverbial intervention and transitivity bias	0.020	0.010	2.067	0.039*
Effect of transitivity bias in the No Adverb condition	–0.012	0.009	–1.290	0.197
Effect of transitivity bias in the Adverb condition (from the same model after releveling conditions)	0.008	0.008	0.939	0.348

Table 8: Results of the Bayesian analysis of Experiment 2: Posterior means on the log-odds scale (with 95% Bayesian CrI).

	Posterior mean	[CrI]	Posterior beyond zero
Grand Mean (Intercept)	1.81	[1.37, 2.26]
Effect of adverbial intervention Adverb < No Adverb	–0.86	[–0.29, –0.14]	99%
Overall effect of verb transitivity bias	–0.003	[–0.02, 0.02]	60%
Interaction between adverbial intervention and transitivity	0.02	[–0.05, 0.01]	90%
Effect of transitivity in the No Adverb condition	–0.01	[–0.03, 0.01]	83%
Effect of transitivity in the Adverb condition (from the same model after releveling conditions)	0.01	[–0.02, 0.03]	73%

4. Experiment 3

The main aim of Experiment 3 was to test a specific prediction made by Pritchett (1988, 1992) that is not necessarily shared by other thematic approaches, namely, that when Theta Attachment allows two attachment options (as in the MC/RC structure), the global obligatoriness of thematic roles is ignored, because thematic assignment maximization operates locally and incrementally.

To test this, in Experiment 3 we focused on the MC/RC ambiguity. The experiment included three conditions, differing only in their embedded verb type. The first two conditions were the same as in Experiment 1, namely, an intransitive relative clause verb with a PP (which served as a control in Experiment 1), and an optionally transitive embedded verb. In this experiment, we added a condition with an obligatorily transitive (OblT) embedded verb⁴ (see Table 9 for a translated example, and Table 10 for a Hebrew example). In the latter condition, both the main verb (‘brought’ in (9)) and the embedded verb (‘received’ in (9)) require a direct object. For the sentence to be grammatical, participants must attach ‘cold water’ locally, as a complement to ‘received’, and choose the completion ‘orange juice’, so that this NP can receive the obligatory Theme role of ‘brought’. If participants choose ‘last night’, the resulting sentence is ungrammatical, as one of the verbs remains with an obligatory thematic role unassigned.

The predictions for the IN and OpT conditions in Experiment 3 are identical to those for Experiment 1. Structural considerations (specifically, Late Closure) predict that the OblT condition should behave very similarly to the OpT condition, with local attachment predicted in both, since the first post-verbal NP will always be attached as an argument of the embedded verb. As explained above, this will eventually lead to a globally grammatical sentence (even if this consideration was not applied locally). Theta Attachment, likewise, predicts similarity between the OpT and OblT conditions, but here, this will be manifested as both local and non-local completions in both conditions, since both present optionality in the attachment of the NP. As explained above, in Pritchett’s theory, the obligatoriness of the thematic role should not influence processing at the point of building the initial structure; only local maximization of thematic assignment is considered. Thus, based on Pritchett’s proposal, initial attachment should follow a chance distribution of local vs. non-local attachment choices in both the OpT and OblT conditions. Further, if we consider OblT verbs as OpT verbs which are extremely biased to the transitive end of the scale, then our proposal regarding OpT verbs in Experiment 1 is also relevant for OblT verbs, namely, that when Theta Attachment does not decide between the two attachment options, the decision will be influenced by verb bias. Since the average verb bias is larger in the OblT condition than in the OpT condition, rates of local attachment are predicted to be higher in the OblT condition. Crucially, if we do find a substantial proportion of non-local attachments, leading to ungrammatical sentences, this provides strong evidence that participants opt for local thematic assignment maximization, rather than always considering global grammaticality.

4.3 Results

Overall completion rates are illustrated in Figure 5 (for data without exclusion, see Table C1 in Appendix C).

Figure 5: Rate of completions by condition, Experiment 3.

The by-item average rates of local attachment in the Optionally Transitive condition are presented in Figure 6, representing the effect of verb transitivity bias on the rate of choosing local attachment. For by-item average rates of local attachment in the Obligatory Transitive condition, see Figure C in Appendix C.

Figure 6: The effect of transitivity bias on the rate of choosing local attachment in the Optionally Transitive condition, Experiment 3.

The results of the frequentist and Bayesian analyses are summarized in Tables 11 and 12, respectively. In the frequentist analysis, we observed a main effect of verb type, with significant differences between all three conditions (all p’s < .001), reflecting higher rates of local attachment in the Intransitive condition, compared to both transitive conditions, as well as higher rates of local attachment in the Obligatorily Transitive condition than in the Optionally Transitive condition. The results of the Bayesian analysis were consistent with this, reflecting an increase in proportion of local attachment in the Intransitive condition, relative to both transitive conditions: the mean of the posterior distribution was 63.5%, with CI [54.7%, 72.6%], in the Optionally Transitive condition, 84.1%, with CI [75.6%, 91.3%], in the Obligatorily Transitive condition, and 95.2%, with CI [92.2%, 97.5%], in the Intransitive condition, with none of the credible intervals for the contrasts crossing zero. As for the correlation between the transitivity bias of OpT verbs and attachment preferences in the OpT condition, the results showed that the effect of transitivity bias on local attachment rate was significant in the generalized mixed-effects model (p = .015). The result of the Bayesian analysis was consistent with this.

Table 11: Results of the frequentist analysis, Experiment 3: Estimates, standard error, t-values and p-values.

	Estimate	SE	t	p
OblT condition > OpT	1.166	0.279	4.172	3.02e-05
IN > OpT	2.649	0.388	6.821	9.03e-12
IN > OblT	1.619	0.273	5.930	3.03e-09
Effect of verb bias in the OpT condition	0.016	0.006	2.438	0.015

Table 12: Results of the Bayesian analysis, Experiment 3: Posterior means on the log-odds scale (with 95% Bayesian CrI).

	Posterior mean	[CrI]	Posterior Beyond zero
OblT > OpT	1.136	[0.577, 1.740]	100%
IN > OpT	2.454	[1.869, 3.088]	100%
IN > OblT	1.555	[0.925, 2.232]	100%
Effect of verb bias in the OpT condition	0.016	[0.006, 0.027]	99.8%

Further, the binomial tests showed that in both the OpT and OblT conditions, the by-participant average choice of attachment was significantly different from the null hypothesis of a chance distribution (t = 3.260, p = .002 and t = 10.675, p < .001, correspondingly).

5. General discussion

The present study investigated the degree to which attachment decisions are guided by thematic considerations. We specifically tested one implementation of this idea, namely, Pritchett’s (1988, 1992) principle of Theta Attachment: At every point during comprehension, the parser attempts to maximally satisfy the Theta Criterion. We compared the predictions of this principle with the structural principle of Late Closure, as well as with guidance by verb bias.

Experiment 1 showed that in sentences with object/subject ambiguity, where there is only one attachment choice which enables thematic assignment (10), the ambiguous NP invariably attaches to the preceding verb, maximizing thematic assignment at this point. This result is in line with both Theta Attachment and Late Closure. However, in MC/RC sentences, when both attachment options enable thematic assignment to the ambiguous NP, as in (11), the results show optionality with regard to attachment of this NP, contrary to the predictions of Late Closure, and in line with Pritchett’s proposal.

Additionally, we found that the transitivity bias of the optionally transitive verb was not correlated with attachment choices in sentences like (10), but only in those like (11). We proposed that this is so since in (10), when only one structure enables thematic assignment, this consideration dictates parsing choices. However, in (11), when thematic considerations cannot decide, other factors – i.e., transitivity bias – affect processing choices.

Experiment 2 found that when adverbs were added to sentences with object/subject ambiguity, such that the ambiguous NP was no longer adjacent to the preceding verb, participants still largely opted for local attachment, aiming to maximize thematic assignment, and there was no effect of transitivity bias on attachment choices, replicating Experiment 1. However, the effect of distance was statistically significant, with a higher proportion of local attachment when the NP was adjacent to the preceding verb. Thus, locality and thematic assignment maximization seem to both be at work during structure building.

Experiment 3 further found that when sentences with MC/RC ambiguity, such as (11), included an obligatory transitive subordinate verb, both local and non-local attachments were attested, providing some evidence for Pritchett’s assumption that optionality will arise, since the Theta Criterion can be locally satisfied in two different ways, even when there is only one globally grammatical option. However, verb bias (possibly coupled with global grammaticality) influenced parsing choices, such that participants showed a stronger preference for local attachment than in the OpT condition.

Comparing the processing strategies considered in the present study, we can first observe that the structural principle of Late Closure mostly does not account for the results, as it predicts uniform choice of local attachment, regardless of the number of options for thematic assignment, distance, or obligatoriness of thematic roles. Specifically, the variability found in attachment choices in the MC/RC condition, both within and between items, provides evidence against the purely structural approach.

Further, the results show that verb biases do not always guide processing. In the first experiment, the effect of verb bias was qualitatively different between the case with one thematic assignment option and the case with two thematic assignment options, with no effect in the former. This result is hard to explain if verb preferences determine processing choices. In addition, there was no effect of verb bias on attachment choices in Experiment 2.

The experiments suggest a central role for thematic considerations during processing. All three experiments point to the importance of the satisfaction of the Theta Criterion in making parsing decisions, such that structures are selected that maximize thematic assignment. When there is only one structure that maximizes thematic assignment, this structure will be selected in a large majority of the cases, without consideration of the bias of the verb (Experiments 1 and 2), although distance can somewhat reduce this tendency (Experiment 2). In contrast, when both structures satisfy the Theta Criterion to the same degree, as in the MC/RC ambiguity structures in Experiments 1 and 3, other considerations, namely, verb bias, have a much more pronounced effect. Thus, if we compare the effect of transitivity bias when thematic considerations are strong (obj/subj ambiguity, Experiments 1 and 2) vs. weak (MC/RC ambiguity, Experiment 1 and 3), we find a strong effect of bias in the latter case, but no effect in the former. Taken together, the experiments provide evidence for the precedence of thematic assignment and, consequently, of interpretation maximization during processing. Though Pritchett’s Theta Attachment does not take into account the influence of distance and verb bias information (or other types of frequency information), its contribution to processing theory is in its implementation of the maximal interpretation hypothesis.

While the current study is the first to investigate MC/RC sentences, much previous work has targeted object/subject ambiguities, and our results largely align with theirs. Most relevant are Pickering et al. (2000, Experiment 3), Itzhak et al. (2010) and Huang et al. (2024), which examined the effect of verb bias in these ambiguities. Notably, Itzhak and colleagues found a correlation of the P600 effect with verb bias in these sentences, though there was processing difficulty regardless of verb bias, whereas Pickering and colleagues and Huang and colleagues did not, similarly to our study. The difference from Itzhak and colleagues may be due to the different measure, or due to Itzhak and colleagues treating verb bias as a binary variable.

While verb bias fails to account for our findings, an open question is whether Surprisal (Hale, 2001; Levy, 2008), which is a more recent probabilistic theory, can account for them. Surprisal assumes a fully parallel parser which uses the statistics of the language to construct parallel parses of the input, with their weights reflecting the probability of the parse in the language. The focus of Surprisal is on the prediction of localized processing difficulty as reflected in reading times, with prolonged reading times for less predictable input. This is irrelevant to the current study, as our experiments did not measure reading times, but rather attachment choices (but see Huang et al., 2024, for testing the predictions of Surprisal in Garden Path sentences). Surprisal does not assume attachment “choices”, since multiple structures are built at any point. However, it is possible to assume that the preferred completion will be the one which is compatible with the most highly weighted structure to this point. To test whether Surprisal can account for our results, it would be necessary to extract the probabilities of the different continuations in the different ambiguity types from a Hebrew corpus. We leave this for future research. We do note, however, that in the OblT condition in Experiment 3, readers opted in 20% of trials for completions resulting in globally ungrammatical structures, which have very low frequency in the language. It seems that at least in this case, completion options do not track corpus probabilities.

As the sentences we used, especially in the MC/RC conditions, were very long and complex, another model which might be considered to account for their processing is lossy-context surprisal (Futrell et al., 2020; Hahn, Futrell et al., 2022). This theory adopts Surprisal’s assumptions, but proposes that predictions are based not on a complete, accurate representation of the sentence up to this point, but rather on a lossy, noisy version of the sentence. According to such a model, it can be assumed that readers will sometimes fail to assign the thematic role of the main clause verb in MC/RC sentences not because of optionality, but rather because the memory representation of the first verb is degraded or weak, due to memory capacity and the low probability of this construction. However, we would like to point out that a “forgetting” scenario should lead to the same problem in the MC/RC control condition in Experiment 1 / Intransitive condition in Experiment 3, repeated in (12). Namely, participants would “forget” the ditransitive verb and select the adverb completion. However, participants overwhelmingly selected the NP completion in the experiments (95% in both experiments), showing that they did remember the beginning of the sentence. Future research can evaluate lossy-context surprisal against our data in more detail.

Based on the present experiments, we suggest that any processing model should capture the influence of thematic information and the goal of maximizing thematic assignment and interpretation. Other important factors influencing processing, such as thematic biases of verbs and their distance from their argument, have a secondary role, but do not eliminate the importance of Theta Attachment and maximal interpretation.

Appendix A. Analysis of all data from Experiment 1

Table A1 presents the rates of choosing local attachment in Experiment 1 when no trials are excluded based on long RTs.

Tables A2 and A3 present the results of the frequentist analysis and Bayesian analyses, respectively, of all data in Experiment 1 (without trial exclusion based on RT).

Appendix B. Distribution of direct objects with and without preceding adverbs

To verify whether in Hebrew obj/subj sentences, the probability that the ambiguous NP opens a new clause is larger when it is preceded by an adverb than when it follows the verb directly, we manually coded a random sample of 100 occurrences of four verbs in Sketch Engine (Hebrew Web 2014 (heTenTen14, Meni/Alon tagged + lempos)). We selected four representative experimental verbs, which are in different places on the transitivity scale. We counted occurrences of a direct object when an adverb was present after the verb, compared to when there was no adverb. The results showed that, indeed, occurrence of a post-verbal adverb indicates that the chance of a direct object is smaller, compared to when there is no adverb.

Table B1: Direct object NP occurrence when an adverb is present after the verb compared to when it is absent.

Verb	Total % occurrence with direct object	Total % of occurrence with adverb	% occurrence with direct object when adverb does not appear after verb	% occurrence with direct object when adverb appears after the verb
dance	16	36	21	8.6
cook	46	38	54	12
eat	69	14	74	58
operate/analyze	89	15	91	62

Appendix C. Analysis of all data from Experiment 3

Table C1 presents rates of choosing local attachment in Experiment 3 when no trials are excluded based on long RTs.

Tables C2 and C3 present the results of the frequentist analysis and Bayesian analysis, respectively, of all data in Experiment 3 (without trial exclusion based on RT).

Figure C below adds an information concerning the distribution of the by-item rates in the OblT condition, for trials whose RTs < 6800 ms.

Figure C: The rate of choosing local attachment as a function of the transitivity bias in the Optionally Transitive and Obligatorily Transitive conditions, Experiment 3.

Data accessibility statement

All data and R code are available in the following OSF online repository: https://osf.io/uxnmc/?view_only=199ede8bf030465aac9da1187b909b59.

Ethics and consent

Ethical approval for the experiments of this study was granted by the Institutional Ethics Committee at Tel Aviv University; Reference numbers 0000032-3, 0000032-4 and 0000032-5.

Acknowledgements

This research was funded by the Israel Science Foundation, grant 525/22 (PI: Aya Meltzer-Asscher). We thank audiences at AMLaP 27 and HSP 36, and at the workshop of the Sentence Processing and LISC labs at Tel-Aviv University, for their valuable feedback. Lastly, we thank the editor and three reviewers for insightful comments on earlier versions of this article, which have helped us improve it substantially.

Competing interests

The authors have no competing interests to declare.

Authors’ contributions

Lola Karsenti: conceptualisation, methodology, software, investigation, data curation, formal analysis, visualization, writing.

Aya Meltzer-Asscher: conceptualisation, methodology, investigation, funding acquisition, supervision, writing, review & editing.

ORCiD IDs

Lola Karsenti: orcid.org/0009-0004-7016-4665

Aya Meltzer-Asscher: orcid.org/0000-0003-4886-0029