1.1 The predicate decomposition approach to verb semantics

Predicate decomposition is a dominant approach to verb semantics, and it is widely adopted in generative theories. The predicate decomposition approach assumes that verb meaning can be decomposed into primitive predicates and constants. Primitive predicates are the structural part of verb meaning, and they represent the grammatically relevant aspects of verb meaning. Verbs that share the same primitive predicate are supposed to behave syntactically and morphologically alike. Constants, also known as roots (Pesetsky 1995), are the idiosyncratic part of verb meaning. They differentiate one verb from the others that share the same structural meaning. Under such an assumption of verb meaning, instantiations of predicate decomposition theory differ mainly in which aspect of events is emphasized in constructing the grammatically relevant part of meaning. The localist approach focuses on the motion and location in events (Jackendoff 1972; 1976; 1983; 1987; 1990); the causal approach highlights the causal relation and the transmission of force between event participants (Langacker 1987; 1991; 1994; Croft 1991; 1994; 1998; 2012); and the aspectual approach centers on the temporal properties of events (Vendler 1957; Tenny 1994; Van Valin & LaPolla 1997; Rappaport Hovav & Levin 1998; Van Valin 2005). Among these approaches, the localist approach is meant to provide an account of certain instances of systematic polysemy and preposition distribution, and it does not aim to account for argument realization per se. In contrast, the causal approach and the aspectual approach are designed to solve the problem of argument realization. Since the transfer of force and the temporal order can both be represented by the notion of precedence, these two approaches often end up using quite similar semantic representations for verbs. However, as will be argued shortly in Section 1.2, the aspectual approach makes explicit predictions of verb semantic complexity, so the rest of this section focuses on this specific approach.

The aspectual approach to verb semantics can be dated back to the long tradition of categorizing verbs based on lexical aspect (also known as Aktionsart) in both linguistics and philosophy. The current study does not try to cover the full richness and depth of the vast literature on lexical aspect, but rather discusses the aspectual approach with a special emphasis on its applications in argument realization. The aspectual predicate decomposition approach to verb semantics generalizes verb meaning by the aspectual properties that were developed primarily based on Vendler’s (1957) four verb aspectual classes – states, activities, achievements, and accomplishments. Generally, states describe events that do not involve change (e.g., know, believe); activities describe events that have no inherent temporal endpoints (e.g., run, laugh); achievements describe events that are punctual and have inherent temporal endpoints (e.g., realize, discover); and accomplishments describe events that are durative and have inherent temporal endpoints (e.g., build, fix). According to this classification, verbs are represented by lexical semantic templates, as shown in (2).

Primitive predicates are represented by nonitalics. The combinations of primitive predicates in a language constitute a fixed inventory of lexical semantic templates. Additionally, because these combinations often define the ontological types of events, they are also referred to as event structure templates. Constants are represented by the capitalized italics in angle brackets. The set of constants in a language is open-ended, but each constant can be categorized into a fixed set of ontological types (e.g., manner, state). Constants determine the basic association with event structure templates and the number of event participants of verbs. A variety of other aspectual classifications have been proposed since Vendler (1957). They either superimpose a different organization on these classes or further subdivide them into smaller categories. For example, Smith (1991) and Engelberg (1999) proposed a fifth class, the semelfactives, and Van Valin and LaPolla (1997) and Van Valin (2005) subdivided each of the four basic categories into a noncausative version and a causative version. However, none of these are radically different from Vendler (1957).

The aspectual representation of verb meaning has an important consequence in argument realization. Consider the following examples:

Sweep is an activity verb, and is represented as (5) under the aspectual approach.

This event structure has one subevent. There are two participants in this subevent: a structural participant, the sweeper (marked as ‘x’), which is defined as licensed by both the constant and the event structure, and a content participant, the swept (marked as ‘y’), which is defined as licensed by the constant only. To map this representation to the argument structure, the aspectual approach proposes two well-formedness conditions on syntactic realization: the Subevent Identification Condition (6) and the Argument Realization Condition (7).

The SIC aligns predicates with subevents, and the ARC aligns subevents with arguments. In example (3), the verb sweep identifies one subevent ensuring that the SIC is met. The structural participant is expressed in (3a) and (3b) so that both sentences satisfy the ARC. In addition, the constant participant, the floor, can be recovered with a prototypical understanding; thus, both (3a) and (3b) are acceptable. In this way, the predicate decomposition approach successfully predicts the argument structures in (3a) and (3b) through the aspectual representation of sweep.

In contrast, break is an accomplishment verb, and is represented as (8) under the aspectual approach.

This event structure has two subevents: a causative activity and a change of state. Each subevent further contains one structural participant: the breaker (marked as ‘x’) and the broken (marked as ‘y’), which are licensed by the constant and the event structure. The verb break identifies both subevents ensuring that the SIC is met. According to the ARC, the structural participants need to be obligatorily expressed. Thus, only (4b) is acceptable, not (4a). In this way, the predicate decomposition approach successfully predicts the argument structure in (4b) and rejects the argument structure in (4a) through the aspectual representation of break.

In sum, for predicate decomposition theory, argument realizations of verbs can be predicted by the aspectual event structures in the lexical semantic representations of verbs. Verbs with different aspectual representations, such as sweep and break, are supposed to project differently to argument structures.

1.2 Semantic complexity under the predicate decomposition approach

Under the aspectual predicate decomposition approach, an event is argued to be complex when the temporal relation between its subevents is not necessarily aligned (Levin & Rappaport Hovav 1999; 2005).¹ For example, John thawed the meat. It contains two subevents: John caused and the meat thawed. This is a complex event in the sense that these subevents do not have to continue for the same temporal duration. John could simply take the meat out of the fridge and do something else while it thaws.

By this criterion of event complexity, it is hypothesized that accomplishment verbs such as break are conceptually more complex than activity verbs such as sweep (Levin & Rappaport Hovav 1999). Note that these accomplishment verbs do not include incremental theme verbs (Dowty 1991), such as verbs of consumption (e.g., eat, drink), and verbs of creation (e.g., write, build).² Incremental theme verbs are analyzed as accomplishments in the sense that they involve an event with a duration and endpoint. However, these activities inherently coexist with the change of state. That is, the two subevents in these event structures are necessarily aligned. Therefore, they are predicted to have a simple, rather than a complex event (Levin & Rappaport Hovav 2004).

This hypothesis of event complexity is syntactically meaningful in, for example, explaining the sharp discrepancies between English resultatives as (9) and (10).

It is argued that sentences in (9) denote a simple event because the subevent of freezing and the subevent of becoming solid necessarily coexist with each other, while the sentences in (10) denote a complex event because the subevent of talking and the subevent of becoming hoarse do not necessarily coexist with each other. Therefore, according to the ARC, a “fake” reflexive is required in (10b) to qualify the sentence as grammatical but not in (9b) (Rappaport Hovav & Levin 2001).

The hypothesis of verbs as event-structure-taking functions has received supportive evidence in language processing studies. Gennari and Poeppel (2003) conducted a lexical decision task, a self-paced reading task, and found that eventive verbs (i.e., activities, achievements and accomplishments) take longer to respond to than stative verbs. McKoon and Macfarland (2000; 2002) performed two timed acceptability judgement tasks, two timed reading comprehension tasks and a lexical decision task. They reported that externally caused change-of-state verbs (accomplishments) take a longer time to respond to than internally caused change-of-state verbs (achievements). Moreover, McKoon and Love (2011) carried out a lexical decision task, a timed acceptability task and a stop making sense task. They concluded that externally caused change-of-state verbs (accomplishments) take longer to respond to than activity verbs, which directly supports the semantic complexity hypothesis discussed above.

1.3 The constructional approach to verb semantics

Construction grammar is a family of theories that take constructions as the fundamental building blocks in languages. There are various instantiations of construction grammar that share a similar assumption about the constructional nature of language knowledge but emphasize slightly different aspects in their goals. For example, Cognitive Construction Grammar (Goldberg 1995) looks into argument realization and its psycholinguistic implications; Radical Construction Grammar (Croft 2001) describes language with a strong typological orientation; Sign-Based Construction Grammar (Sag & Boas & Kay 2012) is centrally concerned with the precise formalization of languages; and Embodied Construction Grammar (Bergen & Chang 2013) and Fluid Construction Grammar (Steels 2017) aim at the computational formalism of constructions.

Instead of looking into the internal structure of verb meaning as in the predicate decomposition approach, construction grammar assumes that each sense of a verb is conventionally associated with the frame semantic meaning that specifies certain participant roles (Goldberg 1995: 43; 2002: 342; 2006: 39). Participant roles are frame-specific event participants. They differ from traditional semantic roles mainly in two aspects: first, participant roles are concrete concepts that are closely related to the scenes described by verbs rather than abstract generalizations across different verbs; second, participant roles, therefore, are open-ended, which are defined by individual verbs rather than a predetermined fixed set of roles. A subset of participant roles, namely the lexically profiled roles, are obligatorily expressed (Goldberg 1995: 45) or, if unexpressed, must receive a definite interpretation (Goldberg 2002: 342; 2005b: 225). For example, the participant roles of hand in the sense of cause-receive are the hander, the handee and the handed (11) in which all three roles are profiled (marked in bold).

Moreover, the most unique feature of construction grammar is that it takes construction as the primitive language unit. Constructions are generally defined as conventionalized pairings between form and meaning (or function) in a language (Goldberg 1995; 2006; 2013; 2019). Argument structure constructions are associated with argument roles, which roughly correspond to traditional semantic roles or thematic roles, such as agent, patient, instrument, theme, or location (Goldberg 1995: 43; 2002: 342; 2005b: 225; 2005a: 23; 2006: 39). Argument roles that are linked to a direct grammatical relation (SUBJ, OBJ, or OBJ2) are considered constructionally profiled (Goldberg 1995: 48; 2002: 343; 2005b: 225). The psychological reality of constructions has been supported by an increasing amount of evidence from language acquisition (Goldberg & Casenhiser & Sethuraman 2004; Casenhiser & Goldberg 2005; Goldberg & Casenhiser & White 2007; Boyd & Gottschalk & Goldberg 2009; Suttle & Goldberg 2011; Boyd & Goldberg 2012; Goldberg 2016; Perek 2016; Perek & Goldberg 2017; Robenalt & Goldberg 2015) and language processing (Ahrens 1995; Bencini & Goldberg 2000; Kaschak & Glenberg 2000; Goldberg & Casenhiser & Sethuraman 2005; Kako 2006; Goldwater & Markman 2009; Allen et al. 2012; Johnson & Goldberg 2013; Johnson & Turk-Browne & Goldberg 2015).

The representation of verb participant roles has a great impact on argument realization under construction grammar. Take the same set of example sentences discussed in Section 1.1, for instance (repeated below):

Sweep is a three-participant-role verb, and it is represented as (14) under the constructional approach. The participant roles are the sweeper, the swept and the sweep.tool, in which the first two are lexically profiled.

Fusion is the interaction between verbs and argument structure constructions. It captures the simultaneous constraints on participant roles and argument roles in argument realization. In sentences (12a) and (12b), the verb sweep fuses with an intransitive construction and a transitive construction, respectively. The fusion is restricted by the Semantic Coherence Principle (15) and the Correspondence Principle (16).

The SCP ensures that participant roles and argument roles are semantically compatible, and the CoP ensures that participant roles and argument roles are pragmatically compatible. Likewise, the Lexical Constructional Model (Ruiz de Mendoza Ibáñez & Mairal Usón 2008; Butler 2009; Mairal Usón & Ruiz de Mendoza 2009; Ruiz de Mendoza Ibáñez & Mairal usón 2011; Ruiz de Mendoza Ibáñez 2013; Butler & Gonzálvez-García 2014: 118–125) has developed a more refined set of constraints, both internal and external to the fusion process. The internal constraints take the form of licensing or blocking factors that depend on lexical class ascription, lexical-constructional compatibility, and either predicate or internal variable conditioning of external variables. The external constraints result from the possibility or impossibility of performing high-level metaphoric and metonymic operations on the lexical items involved in the subsumption (fusion) process.

In sentence (12b), the sweeper, John, fuses with the agent argument, and the swept, the floor, fuses with the patient argument so that the SCP is met. The lexically profiled roles, namely, John and the floor, realize as the constructionally profiled argument roles so that the CoP is satisfied. Therefore, sentence (12b) is acceptable. In sentence (12a), although the swept, the floor, is a profiled role, it can receive a definite interpretation in the context, and as long as it is not the focus of the sentence, then it is omissible. Therefore, sentence (12a) is also acceptable. In this way, the constructional approach successfully predicts the argument structures in (12a) and (12b) through the participant-role representation of sweep.

In contrast, break is a two-participant-role verb,³ and it is represented as (17) under the constructional approach. The two participant roles are the breaker and the broken, and both are lexically profiled.

In sentences (13a) and (13b), the verb break fuses with an intransitive construction and a transitive construction, respectively. In sentence (13b), the breaker, John, fuses with the agent argument, and the broken, the dishes, fuses with the patient argument so that the SCP is met. The lexically profiled roles, namely, John and the dishes, realize as the constructionally profiled argument roles so that the CoP is satisfied. Therefore, sentence (13b) is acceptable. In (13a), the lexically profiled role of the verb break, the broken, is not realized. The intransitive construction lacks an argument role for the verb break to realize its lexically profiled roles, which violates the CoP. Sentence (13a) is therefore unacceptable. In this way, the constructional approach successfully predicts the argument structure in (12a) and rejects the argument structure in (12b) through the participant-role representation of break.

In sum, for the constructional approach, the argument realizations of verbs are predicted by the fusion between the participant roles of verbs and the argument roles of argument structure constructions. Verbs with different numbers and profiling statuses of participant roles, such as sweep and break, are supposed to fuse differently with argument structure constructions.

1.4 Semantic complexity under the constructional approach

Under the constructional approach, a verb is hypothesized to be complex when it has more participant roles than others. That is, three-participant-role verbs such as sweep are conceptually more complex than two-participant-role verbs such as break. This hypothesis of event complexity is syntactically meaningful in argument realization because the number of participant roles is an important determinant of fusion. Take the case of the English resultatives again, for example.

Sentences in (18) are instances of the resultative construction with a three-participant-role verb wipe, and a one-participant-role verb talk. In (18a), the verb wipe inherently contributes a semantically compatible participant role, the wiped, to fuse with the patient object of the resultative construction. In contrast, in (18b), the verb talk lacks such a participant role; thus, according to the CoP and SCP, a “fake” reflexive herself is augmented to fill in the patient object position in the construction (Goldberg & Jackendoff 2004). This expression is motivated by external constraints that result from high-level metaphorical and metonymic operations (Ruiz de Mendoza Ibáñez & Mairal Usón 2008; Peña-Cervel 2016). The metaphor involved here is A VERBAL PROCESS/A VERBAL PROCESS AS A FORM OF BEHAVIOUR IS A MATERIAL CONTACT PROCESS. This implies that the verb talk is metaphorically construed as a material contact process in which the talker becomes the actor and the verbal process of talking corresponds to a material contact process such as wipe. Furthermore, the metaphor provides a goal, herself, and a result, hoarseness. In addition, the metonymy involved here is MANNER FOR VERBAL PROCESS/VERBAL PROCESS AS FORM OF BEHAVIOUR. This implies that the manner, which is conflated into the meaning of talk, stands for the whole matrix domain of the verbal process of talking.

The hypothesis of verb as participant-role-taking functions has received supportive evidence in language processing studies as well. Ahrens and Swinney (1995) conducted a cross-modal lexical decision showing a significant reaction time difference between verbs with different numbers of participant roles. Following this study, Ahrens (2003) performed a sentence-continuation lexical decision task indicating that verbs with more participant roles take a longer time to integrate into sentences. These findings have consistently supported the hypothesis that verbs with more participant roles are more complex in sentence processing.

In summary, the predicate decomposition approach and the constructional approach are two competing theories of verb semantic representation. Both approaches are syntactically meaningful in predicting argument structures and supported by language processing evidence. However, the two approaches make distinct hypotheses of verb semantic complexity. The predicate decomposition approach hypothesizes that accomplishment verbs (e.g., break) are more complex than activity verbs (e.g., sweep). On the contrary, the constructional approach hypothesizes that three-participant-role verbs (e.g., sweep) are more complex than two-participant-role verbs (e.g., break). The experiments reported below thus test these hypotheses in lexical access (Experiment 1) and sentence processing (Experiment 2) using reaction time as an indicator of semantic complexity. It is assumed that the representation of verbs is derived from the experience of real-world situations and language use. The semantic representation of verbs is a generalization of the events and scenarios associated with the verb. Thus, understanding a verb online recruits the representation of the corresponding events and scenarios. Following previous work on the semantic representation of verbs (McRae & Ferretti & Amyote 1997; McRae et al. 1997; McKoon & Macfarland 2000; 2002; e.g., Ferretti & McRae & Hatherell 2001; Gennari & Poeppel 2003; McKoon & Ratcliff 2008), it is hypothesized here that the more complex the events and scenarios, the greater the amount of information to be accessed in working memory, and thus the longer the processing time of the verb is likely to be.

2.2 Results

Reaction times shorter than 200 ms and longer than 3,000 ms were first excluded from the data analysis. For the rest of the data, any reaction time shorter than the mean (of the same participant and the same lexical decision target) minus three times its standard deviation or longer than the same mean plus three times its standard deviation was excluded. This has led to 2.64% loss of the original data and resulted in 1,069 observations. Participants responded faster and more accurately to accomplishment verbs (Mean Reaction Time = 631.38 ms; Standard Error = 9.56; Mean Accuracy = 0.96) than to activity verbs (Mean Reaction Time = 637.06 ms; Standard Error = 9.65; Mean Accuracy = 0.93). In addition, participants responded faster to two-role verbs (Mean Reaction Time = 626.64 ms; Standard Error = 9.81) than to three-role verbs (Mean Reaction Time = 641.99 ms; Standard Error = 9.37), but the accuracies for the two types of verbs were almost identical (Mean Accuracy = 0.95). The mean reaction times for the four types of verbs are shown in Figure 1.

Figure 1

Reaction time by verb types in Experiment 1 (Error bars indicate mean ± SE).

The reaction times and accuracies were further analyzed using linear mixed effects models (Baayen 2012) with the lmerTest package (Kuznetsova & Brockhoff & Christensen 2017) in R (R Core Team 2013). First, because the distribution of the reaction time was right skewed, the original data were log-transformed to improve the accuracy of statistical modeling. Second, the collinearity of the numerical factors summarized in Table 2 and Table 3 was checked. Factors measuring the same dimension (e.g., frequency and logarithmic frequency) were minimized to one to keep the correlation between predictors as small as possible. This left five predictors, which were logarithmic frequency, word length in terms of the number of syllables, phonological neighborhood size, orthographic neighborhood size and the number of senses (κ = 22.94).

A mixed effects model was fitted to predict the log-transformed reaction time. The fixed effects were the verb aspectual type, the number of participant roles, their interaction as well as the selected covariates and their binary interactions. These factors were coded as sum contrasts (1 vs –1 for accomplishment verbs and activity verbs, and for three-role verbs and two-role verbs). The random intercepts were the participant and the item. Then the model was minimized by the Akaike information criterion (AIC) using the likelihood ratio test. The results are summarized in Table 4.

It is shown that three-role verbs significantly increase (t(30.929) = 2.720, p = 0.011) the log-transformed reaction time by 0.035 ± 0.013 compared to two-role verbs. In contrast, the reaction time difference between activity verbs and accomplishment verbs is not statistically significant (t(31.534) = 0.775, p= 0.444). Additionally, the model indicates two pairs of significant interactions. The effects of orthographic neighborhood size (t(31.470) = 3.076, p = 0.004) and phonological neighborhood size (t(31.357) = –3.290, p = 0.002) on reaction time changed significantly based on the logarithmic word frequency.

Another mixed effects model was fitted to predict the accuracy. The fixed effect was the verb aspectual type. The factor was again coded as sum contrasts as in the models for reaction times. The random intercepts were the participant and the item. It is shown that there is no significant difference in accuracy between activity verbs and accomplishment verbs (t(32.098) = 0.948, p = 0.350).

2.3 A replication of McKoon and Love (2011)

The results from Experiment 1 do not support significantly longer lexical decision time of accomplishment verbs over activity verbs, which turns out to be inconsistent with the findings in McKoon and Love (2011). To better understand the discrepant results and to further examine the lexical decision time of accomplishment verbs and activity verbs, the lexical decision task in McKoon and Love (2011) was replicated here.

This replication was carried out with the same set of verbs as the original study. Lexical decision times of the 32 verbs used by McKoon and Love (Table 5) were again extracted from the English Lexicon Project. There were 34 to 40 observations for each verb, for a total of 1188 data points. The number of observations varies for each verb due to data loss. These data were contributed by 436 participants in the corpus, with each participant only seeing a subset of the 32 verbs.

Reaction times shorter than 200 ms and longer than 3,000 ms were first excluded from the data analysis. For the rest of the data, any reaction time longer than the mean (of the same participant and the same word target) plus three times the standard deviation or shorter than the same mean minus three times the standard deviation was excluded. This has led to 2.86% loss of the original data and resulted in 1,154 observations. Participants responded faster to activity verbs (Mean Reaction Time = 610.68 ms; Standard Error = 7.33) than accomplishment verbs (Mean Reaction Time = 641.39 ms; Standard Error = 9.51), but the accuracy was identical for both types of verbs (Mean Accuracy = 0.97). The mean reaction times for the two types of verbs are shown in Figure 2.

Figure 2

Mean reaction time by verb types in the replication of McKoon and Love (2011) (Error bars indicate mean ± SE).

The reaction time was first analyzed using an ANOVA test as in McKoon and Love (2011). The results show that accomplishment verbs take significantly longer reaction times (F(1, 1152) = 6.542, p = 0.011) than activity verbs, as concluded in McKoon and Love (2011).

An alternative way to analyze the data here is to use linear mixed effects models as in Section 2.2. Here, two mixed effects models were fitted. The outcome was the log-transformed reaction time. In the first model, the fixed effect was simply the verb aspectual type, and the factor was coded as sum contrasts (1 vs –1 for accomplishment verbs and activity verbs). The random intercepts were the participant and the item. This model indicates that the reaction time difference between activity verbs and accomplishment verbs is not significant (t(31.572) = 1.167, p = 0.252). The second model built on the first model by adding other covariates that are known to influence lexical decision times and adopted a similar structure as the mixed effects model in Section 2.2, except that it did not include the number of participant roles. This model was minimized by AIC using the likelihood ratio test. The result again indicates that the reaction time difference between activity verbs and accomplishment verbs is not significant (t(29.273) = 0.243, p = 0.810).

The two methods of data analysis have produced very different results. This can be attributed to the experimental design of McKoon and Love (2011), as well as the different assumptions between the two statistical methods. In the McKoon and Love’s study, the data contains multiple responses from the same participant and multiple responses to the same verb. In this case, the responses are not independent of each other, which in fact violates one of the most important assumptions of the ANOVA, the independence of observations (e.g., Howell 2014: 400). Violating independence may greatly increases the likelihood of obtaining spurious results and lead to meaningless p-values (Winter 2013). However, this problem can be addressed by adding random effects to the subjects and items and using mixed models. Mixed models can account for by-subject and by-item variation in a single model, solving the problem of non-independence.

The incongruence between the results of the ANOVA and the mixed effects models suggests that the significant reaction time difference reported by McKoon and Love (2011) may be better explained by the individual difference among participants and/or experimental verbs rather than verb aspectual types per se. This is probably why the current study failed to replicate their results and found a null effect of verb aspectual type on lexical decision time.

2.4 Conclusion

The results of Experiment 1 and the replication of McKoon and Love (2011) jointly show that verbs with more participant roles, but not more complex aspectual event structures, cause longer lexical decision time. This has supported the verb semantic complexity hypothesis under the constructional approach but not the predicate decomposition approach. In other words, according to the word processing data reported above, the number of participant roles is a better indicator of verb semantic complexity than the aspectual event structure.

3.2 Results

The data were cleaned based on the accuracy of comprehension questions and the reading times of the words. First, any participants with a mean accuracy of comprehension questions lower than the mean accuracy of all participants minus three times the standard deviation were excluded from further analysis. This has led to the loss of one participant’s data. The rest of the data has reached a mean accuracy of 97.08%. Second, any reading time data that were shorter than 100 ms or longer than 3,000 ms were excluded. Third, any reading time data that were shorter than the mean of the word minus three times the standard deviation or longer than the mean reading time plus three times the standard deviation were excluded. This has led to a 4.89% loss of the original data. The remaining reading time data were all used in the analysis. This was done even if the corresponding comprehension question was not answered correctly.

Figure 3 and Figure 4 show the reading time of the region of interest: the critical verb, the noun before it and the determiner after it. Participants read activity verbs (Mean Reading Time = 413.04 ms; Standard Error = 8.45) faster than accomplishment verbs (Mean Reading Time = 428.11 ms; Standard Error = 8.65), and participants read two-role verbs (Mean Reading Time = 405.96 ms; Standard Error = 7.93) faster than three-role verbs (Mean Reading Time = 435.15 ms; Standard Error = 9.10).

Figure 3

Reading time of region of interest by aspectual verb types (Error bars indicate mean ± SE).

Figure 4

Reading time of region of interest by number of participant roles (Error bars indicate mean ± SE).

The reading times were further analyzed using linear mixed effects models. First, the distribution of the reading time was right skewed; thus, the raw data were log-transformed as the new outcome. Second, in addition to the verb semantic categories of interest here, a similar set of covariates as Experiment 1 were taken into consideration. They were word frequency, word length (in terms of numbers of characters, phonemes, syllables, and morphemes), orthographic neighborhood size, phonological neighborhood size, the number of senses, and the plausibility score from the second pretest. The number of argument structures was not considered here for the same reasons stated in Experiment 1. In addition, the previous-word reading time was not included, because this factor has been balanced by the Latin-square design.

The collinearity of the numerical factors was checked. Factors measuring the same dimension were minimized to one to keep the correlation between predictors as small as possible. This left five predictors, which were logarithmic frequency, word length in terms of the number of syllables, orthographic neighborhood size, the number of senses, and the plausibility score (κ = 15.32).

A mixed effects model was fitted. The fixed effects were the verb aspectual type, the number of participant roles and their interaction as well as the selected covariates and all their binary interactions. These factors were coded as sum contrasts (1 vs –1 for accomplishment verbs and activity verbs, and for three-role verbs and two-role verbs). The random intercepts were the participant and the item. This model was then minimized by AIC using the likelihood ratio test. The results are summarized in Table 6.

It is shown that three-role verbs significantly increase (t(30.716) = 2.115, p = 0.043) the log-transformed reading time by 0.026 ± 0.012 compared to two-role verbs. In contrast, the reading time difference between accomplishment verbs and activity verbs is not statistically significant (t(30.558) = 1.097, p = 0.281). In addition, there is a significant interaction between the plausibility score and word frequency. The effect plausibility significantly varies based on logarithmic word frequency (t(365.541) = –2.534, p = 0.012).

3.3 Conclusion

The results of Experiment 2 show that verbs with more participant roles but not more complex aspectual event structures, cause a longer reading time. This has again supported the verb semantic complexity hypothesis under the constructional approach. In other words, according to the sentence processing data reported above, the number of participant roles is a better indicator of verb semantic complexity than the aspectual event structure, which is consistent with Experiment 1.