1.1. Effect of IS on nuclear stress and word order

There are two distinct threads in the literature on IS marking in Spanish. In work largely from a syntactic perspective, Spanish has traditionally been seen as a ‘word order’ language when it comes to the marking of focus (Büring & Gutiérrez-Bravo, 2001; Contreras, 1976; Zubizarreta, 1998). Prosody has only been considered relevant in that sentence-final placement of the focus has been argued to be prosodically motivated (‘p-movement’) by the need to have the nuclear stress final in the intonation phrase (Zubizarreta, 1998). This work has largely been based on introspective or grammaticality judgments. Work from a prosodic perspective within the Autosegmental-Metrical framework has looked at nuclear stress placement and choice of pitch accent type in relation to the IS of the sentence (see Astruc, Mora, & Rew, 2010; Beckman, Díaz-Campos, McGory, & Morgan, 2001; Hualde, 2002; Prieto & Roseano, 2010; Sosa, 1999). This work has largely used read sentences and therefore not looked at the interaction of nuclear stress placement with word ordering and choice of syntactic construction (cf. Face, 2000; Vanrell & Fernández-Soriano, to appear). It is only more recently that there has been experimental work looking at the interaction of prosody and word ordering in IS realization in Spanish (Face, 2000; Face & D’Imperio, 2005; Feldhausen & Vanrell, 2014, 2015; Gabriel, 2010; Gabriel, Feldhausen, & Pes̆ková, 2009; Hoot, 2012; Muntendam, 2009, 2013; Vanrell & Fernández Soriano, 2013; Vanrell & Fernández-Soriano, to appear). This work has revealed considerable variation in the marking of IS between varieties of Spanish. For the purposes of this paper, we summarize the main findings in relation to nuclear stress placement and word ordering/syntactic construction. We also mention findings on pitch accent type; however, this is somewhat orthogonal to explaining the interaction of nuclear stress placement and word ordering.

The literature on IS is wide, with much contradictory use of terminology for similar underlying phenomena (e.g., see Kruijff-Korbayová & Steedman, 2003). For the purposes of this article, we will simply present the framework to be adopted here, making links to existing descriptions of IS in Spanish. This framework recognizes that there are two orthogonal dimensions to IS which are often subsumed under the notion of ‘focus’: the ‘quantificational’ distinction between contrast (C) and background, and the ‘organizational’ distinction between theme and rheme (see Calhoun, 2010a; Halliday, 1967; Steedman, 2000, 2014; Vallduví & Vilkuna, 1998).

The rheme (ρ, also called the comment) is the part of the utterance which updates the common ground, or is new in relation to the current question-under-discussion (QUD) (rather than the discourse as a whole), while the theme (θ, also called topic) establishes a link between the current utterance and the common ground or what the current QUD is (see Vallduví, 2016). A broad focus utterance like the answer in (4) is all-rheme, as the context (the question) does not establish any part of the utterance as the QUD. In broad focus, the nuclear stress falls on the last word in the sentence, and can take a range of pitch accent types depending on the variety of Spanish and the pragmatic function of the sentence (Hualde & Prieto, 2015). Content words earlier in the sentence will usually also be pitch accented; these are less prominent than the nuclear stress accent and have a different pitch accent type (usually L*+H) (Hualde & Prieto, 2015).

In (5) and (6) the question sets up the object and the subject, respectively, as the rheme of the sentence, while the rest of the utterance links back to the context and is hence thematic. We will call the rheme here the QUD-focus (also called presentational or new information focus). As can be seen in (6), it is traditionally claimed that the QUD-focus is always sentence-final in Spanish, so that when the subject is the QUD-focus, it is placed in sentence-final position. According to Zubizarreta (1998), subject final placement is prosodically motivated by the need to have the nuclear stress final in the intonational phrase (‘p-movement’). Note that in the English translation subject focus is marked by stress shift.¹

This claimed pattern of focus-related movement draws mainly from introspective or grammaticality judgments. In more recent experimental work, it has been shown that there are considerable differences in focus marking between varieties of Spanish (Face, 2000; Face & D’Imperio, 2005; Feldhausen & Vanrell, 2014, 2015; Gabriel, 2010; Gabriel et al., 2009; Hoot, 2012; Muntendam, 2009, 2013; Vanrell & Fernández Soriano, 2013; Vanrell & Fernández-Soriano, to appear). For Peninsular Spanish, Face (2000) found that while QUD-focused words were usually sentence final, speakers use nuclear stress in situ to mark QUD-focus 30% of the time, as in the English translation of (6), although Feldhausen and Vanrell (2014) found nuclear stress in situ was used to mark QUD-focus in Madrid Spanish only 15% of the time. In Latin American varieties of Spanish, on the other hand, nuclear stress in situ may be the primary means to mark QUD-focus. Gabriel et al. (2009) and Gabriel (2010) show nuclear stress in situ was used for QUD-focused subjects in 75–95% of cases in Argentinian Spanish. Similar results were found for Andean Spanish (Muntendam, 2009, 2013) and Mexican Spanish (Hoot, 2012). Nuclear stress in situ has also been found to be more common than final subject placement for other regional varieties of Peninsular Spanish (Vanrell & Fernández Soriano, 2013). There are also differences in the realization of the nuclear stress accent between QUD-focus and broad focus (whether in situ or in phrase-final position): In QUD-focus the nuclear accent is generally produced with greater pitch range and an earlier pitch peak (Hualde & Prieto, 2015). There seems to be variation as to whether post-focal accents are deleted entirely in the case of focus marked by nuclear stress in situ early in the sentence, or whether they are realized with a lower pitch range (Hualde & Prieto, 2015; Vanrell & Fernández-Soriano, to appear); however, in all varieties the nuclear stress shifts to the focal word.

In (7) and (8) the theme-rheme structure is the same as in (5) and (6) respectively. However, here we see the operation of the second dimension of IS: Contrast (C)/Background (following Steedman, 2014); Contrast is marked with C, background is unmarked. The notion of Contrast draws on Rooth’s (1992) alternative-semantics definition of focus, i.e., Contrast-marking introduces a presupposition of alternatives to the contrastive element, which needs to be resolved in the context, in this case an explicit contrast between la banana and la manzana in (7) and el chico and la chica in (8). We will call this contrastive focus (which includes corrective focus). It is debated whether all rhemes necessarily contain a Contrast (see discussion in Calhoun, 2010a; Vallduví, 2016), e.g., it could be argued that there is an implicit alternative set to the rheme in (5) and (6).

It has traditionally been claimed that contrastive focus can be marked with stress shift in Spanish, as in (8), while QUD-focus cannot (Domínguez, 2004; Zubizarreta, 1998). As outlined above, however, in many varieties, stress shift can also mark QUD-focus. No consistent differences have been found in pitch accent type or realization between QUD-focus and contrastive focus (Hualde & Prieto, 2015; Vanrell & Fernández-Soriano, to appear).

Production studies have further found that, in both Peninsular and Latin American varieties of Spanish, contrastive focus is usually marked with clefts, with nuclear stress on the cleft (Feldhausen & Vanrell, 2014; Gabriel, 2010), for example (for contrastive focus on the subject):

Clefts had previously been noted as a means of marking contrastive focus in Spanish and other languages (see Guitart, 2013; Lambrecht, 2001; Sedano, 1990). However, they were not usually discussed as the primary means of focus marking in Spanish. Both Feldhausen and Vanrell (2014) for Peninsular Spanish, and Gabriel (2010) for Argentinian Spanish, found that both simple clefts, as in (9) and pseudo-clefts, as in (10) were used to mark contrastive focus, again with nuclear stress on the focus.

Feldhausen and Vanrell (2015) further found that for Peninsular Spanish, simple clefts and pseudo-clefts were also frequently used to mark QUD-focus (cf. Moreno Cabrera, 1999), with pseudo-clefts more frequent than simple clefts for QUD-focus, although this was not found by Gabriel (2010) for Argentinian Spanish.

In all of the examples above, the theme is all background, and it is often not possible to distinguish the background of the rheme from the theme in terms of prosody or word ordering, e.g., (4), (5), and (7) all have the same nuclear stress placement and word order. There is often no Contrast in the theme, as the question-under-discussion is usually established, and so alternatives are not evoked. However, as (11) shows, where there are competing questions-under-discussion, there can be a Contrast within the theme as well as the rheme. We will call this contrastive theme (also called contrastive topic, see Büring, 2003; Calhoun, 2010a; Steedman, 2000, 2014; Vallduví, 2016). In this case, el chico and la chica may be more prosodically prominent than if they were not contrastive. Hualde (2005) reports that the pre-nuclear accent on such initial themes may be produced with a later peak than standard pre-nuclear accents, and the initial topic may be separated into a separate intermediate phrase with a H- phrase tone (see also Hualde & Prieto, 2015).

There is one further aspect of IS relevant to the present study. Spanish is generally held not to apply ‘anaphoric deaccenting,’ as languages like English do, i.e., avoiding putting the nuclear stress on a word which is given in the immediate context (Hualde & Prieto, 2015; Ladd, 2008; Nava & Zubizarreta, 2010; Zubizarreta, 1998). This can be seen in (12):

The most natural way to say the reply in English would be The girl ate the apple, i.e., avoiding putting the nuclear stress on the given apple. Likewise, Spanish is said not to shift stress to mark the most informative or least predictable word within the rheme (Ladd, 2008; Nava & Zubizarreta, 2010). However, it is claimed that there is a preference to put given information before new in a sentence (e.g., see Zubizarreta, 1994). For example, a perhaps more natural response to the question in (12) would be La chica se la comió ‘The girl ate it’ (at least in Venezuelan Spanish). Note here the given pronoun la comes before the verb.

In summary, it is traditionally claimed that QUD-focus on the subject in Spanish is marked by final subject placement, while contrastive focus is marked by stress shift. Experimental studies on varieties of Latin American Spanish suggest that stress shift is most common to mark QUD-focus, and clefting contrastive focus. However, these studies also show considerable variation in focus marking within and between varieties of Spanish. To our knowledge, there has been less experimental work on marking of contrastive themes, and anaphoric deaccenting (or lack thereof).

1.2. Nuclear stress placement and word ordering in intransitives

In intransitives, there is only one noun, the subject. If these follow the same patterns as for transitive sentences in Spanish, as outlined above, we would expect broad focus to be realized sV (where capitals show the nuclear stress), QUD-focus to be realized through final subject placement, i.e., vS, and contrastive focus by stress shift, i.e., Sv. As outlined above, for Latin American varieties like Venezuelan Spanish, stress shift seems to be more likely for QUD-focus, i.e., Sv, and a simple cleft for contrastive focus, i.e., fue Sv (fue indicates the cleft). However, the situation with intransitives is not so straight-forward: Syntactic and semantic features of the verb and subject also affect nuclear stress placement and word ordering. How then do these factors interact with each other? One possibility is that IS and semantic factors are competing influences on word ordering and nuclear stress placement. However, it has been often claimed that the semantic factors said to underlie split intransitivity actually do so because they change the default IS organization of utterances (Allerton & Cruttenden, 1979; Bolinger, 1954, 1972; Contreras, 1976; Faber, 1987; Féry, 2011; Gussenhoven, 1983; Hatcher, 1956; Hirsch & Wagner, 2011; Irwin, 2012; Kratzer & Selkirk, 2007; Schmerling, 1976). We discuss that literature here, and relate it to existing experimental research that has manipulated both unaccusativity and IS in Spanish.

1.2.1. Syntactic and semantic approaches to split intransitivity

According to the unaccusativity hypothesis, it has long been argued that intransitive verbs in many languages can be split into two classes depending on the syntactic properties of their single argument (Levin & Rappaport-Hovav, 1995; Perlmutter, 1978; Sorace, 2000; Zubizarreta & Nava, 2011). The key difference is whether the verb selects an argument which is external to the VP: These are unergative verbs; or an argument which is the complement of the verb within the VP: These are unaccusative verbs. The reflexes of this distinction vary across languages, and have been shown to include auxiliary selection (in languages including Italian, German, and Dutch) (Sorace, 2000), as well as word ordering and nuclear stress placement in broad focus (see further Verhoeven & Kügler, 2015). We will be concerned with the latter two reflexes here.

According to syntactic approaches, the underlying difference in the position of the verb’s argument can explain the observed tendency in unaccusatives to show final subject placement (in Spanish) and stress shift (in English), while unergatives show neither. The exact mechanism by which the syntax-to-prosody mapping occurs differs across frameworks; however, the basic idea is that the nuclear stress is assigned to the most deeply embedded node in the underlying structure (Chomsky & Halle, 1968; Zubizarreta, 1998) (for details of how this can be worked out in a phase-theoretical approach, see Kratzer & Selkirk, 2007). This results in the nuclear stress on the verb in unergatives, and on the argument in unaccusatives, as follows with examples from (2) and (3):

(13)	Unergative:	Unaccusative:

In the case of final subject placement, the order of the words in the surface realization matches that in the underlying structure. In the case of stress shift, the argument moves out of the VP for case-checking or other reasons (before Spell Out, in phase-theoretical terms), but crucially, it carries the nuclear stress with it (see Kahnemuyipour, 2009; Kratzer & Selkirk, 2007).

In its original form, it was assumed within the unaccusativity hypothesis that there was a one-to-one correspondence between argument position (external or internal to the VP) and the thematic role of the argument (agent or patient respectively) (Baker, 1988; Perlmutter, 1978). However, it was soon shown that this correspondence breaks down in a significant number of cases (for a review see Sorace, 2000). This variability has been argued to show that the unaccusative-unergative distinction is semantically determined and syntactically encoded (Levin & Rappaport-Hovav, 1995; Sorace, 2000), with a number of factors being relevant to the semantic determination. In particular, Sorace (2000) has argued that there is a hierarchy of aspectual/thematic verb types, with those at either ends of the scale consistently showing unaccusative and unergative reflexes respectively, while those in the middle are more variable.

Table 1 shows the percentage of intransitive verbs in Spanish realized in VS order in a study by Nava (2007), classified according to Sorace’s (2000) verb type hierarchy. The data are from oral narrations of picture prompts by 13 monolingual speakers of Mexican Spanish.² The hierarchy of verb types generally follows that set out in Sorace (2000), with verbs at the unergative end of the scale (the top) largely realized in SV order, while those at the unaccusative end are more frequently VS. However, it should be noted that even the ‘most’ unaccusative verbs (anti-causatives) are still produced in SV order more than a quarter of the time, and the counts for all but the change of location verbs are low.

Table 1

Percentage of intransitive verbs produced in VS order by Sorace’s (2000) verb types, along with the total count (N) for each type; data from Nava (2007).

Type	Example	% VS Order (N)
Uncontrolled process	estornudar (‘to sneeze’)	0% (3)
Controlled process (non-motional)	reir (‘to laugh’)	0% (8)
Controlled process (motional)	brincar (‘to jump’)	22% (9)
Existence of state	existir (‘to exist’)	13% (8)
Continuation of pre-existing state	permanecer (‘to stay’)	27% (15)
Change of state	nacer (‘to be born’)	62% (21)
Change of location	salir (‘to leave’)	69% (117)
Anti-causative	partirse (‘to break’)	73% (11)

A number of interrelated semantic factors have been argued to explain the hierarchy of verb types, as shown in Table 1. As discussed by Sorace (2000), these include the extent to which the argument is dynamically affected by the action of the verb, or whether the verb involves a change of state or transition for the argument: For both process verbs and change of state/location the argument is affected, but only for the latter does the argument undergo a change of state or transition. Another factor is the degree of agentivity of the argument: For controlled processes the argument is usually also the agent, whereas for change of state/location verbs this is not implied, and for anti-causatives the argument is implied not to be the agent. For English, stress shift has been argued to be less likely for typically unaccusative verbs when they have agentive subjects (Faber, 1987; Hirsch & Wagner, 2011).

4.1. Syntactic coding of responses

The majority of the responses were simple intransitives sentences, SV or VS (n = 588), followed by cleft sentences (n = 37), and sentences with aspectual phrases (n = 26). Cleft sentences were those in which a subordinated clause had been used to focalize an element of the sentence (e.g., el que nació fue el niño ‘The one that was born was the boy’). Aspectual phrases were those containing the constructions acaba de (‘have just’) or empezar a (‘started to’) which function as verb modifiers in Spanish. These were coded separately, as stress shift is noted to be less likely in longer sentences with material between the subject and verb (Ladd, 2008).

As can be seen in Table 3, almost all of the cleft sentences occurred in the contrast information condition, and almost all of the sentences with aspectual phrases occurred in the theme condition, although simple sentences were much more common than either type in these conditions. All but one of 26 sentences with aspectual phrases were SV, and 33 out of the 37 cleft sentences were VS. Since there were relatively few of each type, in the remainder of the analyses (apart from Section 4.6), the sentences with aspectual phrases and cleft sentences were grouped with the simple intransitive sentences according to the order of the subject and main verb. This was necessary as low cell counts are problematic for the statistical models used in Section 4.4. We will discuss these sentence types further in Section 4.6, where it will be shown that this grouping is justified in terms of the patterns found for these sentences.

The responses were also coded for tense, to see if this had an effect on nuclear stress or word order. Most responses were in the past tense (401 simple past, 5 past progressive) and 235 were in the present tense. No distinctive pattern was observed in the use of tense according to IS type or verb type.

4.2. Nuclear stress coding of responses

The responses were coded for the position of the nuclear stress by the three authors using Praat (Boersma & Weenink, 2016). As discussed in Section 1.1, the position of the nuclear stress is the main prosodic cue to the position of focus, and the main prosodic reflex of the unergative/unaccusative distinction, that has been found in previous studies of Spanish for transitive and intransitive sentences. We identified nuclear stress according to criteria compatible with the Spanish Tones and Break Indices (Sp_ToBI) annotation system (Aguilar, de la Mota, & Prieto, 2009; Hualde & Prieto, 2015). Tonal accent type was not annotated. As discussed in Section 1.1, tonal accent type may be relevant to distinguishing between broad focus and QUD-focus in Spanish; however, this is not directly relevant to the research questions in this study. Each response was coded for whether the subject or the verb sounded more prominent, based on auditory perception and the visual display of acoustic information. Annotators could code both the subject and verb as nuclear stressed, for cases where there was clearly a strong phrasal stress on both the subject and the verb, both of which were equally prominent; however, this was used as little as possible.

As they were short, the vast majority of the utterances were produced as a single intonation phrase with a L% boundary tone. In a minority of cases, there was a slight perception of a break after the initial constituent (subject or verb), consistent with a possible H- phrase tone at break index 3 within the Sp_ToBI framework (Aguilar et al., 2009). However, these were not frequent enough to separate out in the analysis.

There were a total of 252 sentences coded with nuclear stress on the initial constituent, 347 coded as final, and 52 with phrasal stress on both the initial and final. As the double stress cases were relatively infrequent, and it would significantly complicate the statistical analysis to have three stress levels (i.e., not binomial), the double stress sentences were dropped from the analysis.

Different participants were coded by different authors, and checked by at least one of the other authors. Two of the three authors are Spanish-speaking; the other author is an expert in prosodic annotation in different languages. Near the start of the annotation process, all three annotators independently coded a sample of 65 sentences, from four of the speakers. Of these, 13 were discarded as being disfluent (i.e., no verb present or a long pause between words) or transitive. Agreement between the annotators was measured for the remaining 53 sentences as to whether each sentence had initial or final nuclear stress (not including double accent cases, i.e., stress on the subject and verb, as these were not included in the main analysis). The agreement rate between the three coders was 82.7% (N = 53, κ = 0.697, z = 8.71, p < .001). Fleiss, Levin, and Myunghee (2004) consider kappa of 0.41 to 0.75 ‘fair to good.’ Cases of disagreement in this sample were discussed, and then the rest of the samples were independently coded, and checked by at least one of the other authors. When cases of disagreement occurred, these discussed by all three authors. In some cases, a fourth, independent Spanish-speaking listener was consulted.

4.3. Acoustic analysis of nuclear stress coding

In order to confirm that the nuclear stress coding was based on underlying differences in the phonetic prominence of the subject and verb in the responses, we analysed a number of acoustic correlates of prominence in the data, to see whether in sentences annotated as having initial nuclear stress, the initial content word was relatively more prominent than in sentences annotated as having final nuclear stress. Using a Praat script, for the subject and verb in each sentence, the values for word and syllable duration, intensity, maximum F0, mean F0, as well as the time alignment of the peak F0 relative to the lexically stressed syllable, were extracted.

Tables 4, 5, 6, 7 show the values for key correlates of prominence in the initial and final content word (subject or verb) in the response sentences. As can be seen in Table 4, in sentences coded as having initial nuclear stress, the initial word has considerably higher maximum F0 than the final. In sentences coded as having final nuclear stress, the final word has slightly higher maximum F0 than the initial; this relatively smaller difference can be attributed to declination. Table 5 shows that in sentences coded as having initial nuclear stress the first lexically stressed syllable is relatively longer than in sentences coded as having final nuclear stress; although the second lexically stressed syllable is always longer in absolute terms (this can be attributed to final lengthening). Table 6 shows there is no clear difference in intensity between initial and final nuclear stress. Table 7 shows the peak in the first word is later if it is nuclear stressed. The F0 and duration measures are therefore consistent with the coded prominence difference.

In order to test whether sentences coded as initially versus finally nuclear stressed could be reliably distinguished based on the acoustic properties of the initial and final content words, we carried out a classification analysis using decision trees.

Decision tree models can be applied to categorical and continuous variables. They are non-parametric methods suitable for non-normal distributions and are not affected by outliers. This type of analysis requires less data cleaning, and does not require normalizing or transformation of continuous variables. A classification tree model recursively splits samples using a series of criteria. A value of a variable is selected and samples are divided into groups, by whether they are above or below this value. Subsequently, this process is repeated using a different variable or a different value of variables already used. Further splitting stops when the remaining samples are homogenous enough.

A classification tree was built to predict the classification of each sentence as initially or finally stressed using the R package rpart (Therneau, Atkinson, & Ripley, 2014). The position of the nuclear stress (initial or final) was predicted using the acoustic variables discussed above as well as the categorical variable Sex, the speaker sex. Note that we ran these models using raw F0 values, and transformed semitone units. We obtained better model fit with the raw F0 values, and hence report these. First a model was fit with a complexity parameter of 0.001. This model was then optimized (pruned) using a complexity parameter of 0.013 (for details on fitting regression trees see Therneau et al., 2014). Figure 2 presents the classification tree obtained in the final model. Nodes on this figure show the name of the variable and the value used to split the samples. Options to the left of a node indicate that the sample meets the criteria of the node, while options to the right indicate the sample does not. The terminal nodes (also called leaves) in the graph show the proportion of correct classifications in that node and the percentage of observations in that node.

Figure 2

Classification tree for stressed word position (initial versus final). Nodes represent the variables used for classification. The figures in each leaf represent the number of corrected classified cases over the number of observations in that node, and the percentage of the data represented by that node.

The percentage of correct classification obtained with this model was 75.5%. The model was more accurate classifying finally nuclear stressed responses (9.52% misclassified) than initially stressed ones (15.03% misclassified) These results indicate that there is a good relationship between the acoustic cues and the nuclear stress coding, and that therefore the stress annotation is reliable.

4.4. Responses by IS type and verb type

Table 8 shows the form of the responses by IS type and verb type condition. As above, SV and VS show the order of the subject and verb, and the capital letter shows the position of the nuclear stress (e.g., Sv is initial stress). As discussed above, double stressed sentences were omitted from the analysis. There were only 13 cases where the verb was nuclear stressed in verb-subject order, i.e., Vs pattern (out of 99 VS sentences). As cells with counts less than 10 are problematic for statistical modelling (see Courvoisier, Combescure, Agoritsas, Gayet-Ageron, & Perneger, 2011; Hosmer & Lemeshow, 2000), we grouped these with the other verb-subject order sentences in the analysis, so the vS column shows all verb-subject order sentences.

Table 8 shows that the vast majority of sentences were produced in subject-verb order. Of the two strategies for stressing the subject, stress shift (i.e., Sv) was more common than final subject placement (i.e., vS) over all conditions except for broad focus unaccusatives. Among the IS conditions, contrast appeared to have the largest effect on response form, i.e., nuclear stress placement and word order (compared to broad focus), with 21% sV responses (the assumed default). QUD-focus appears to have a similar effect to contrast, although less strong, with a greater proportion of sV responses. The theme condition was similar to broad focus, although with very few vS realizations. As there were a number of low or zero counts in the theme condition, which are problematic for statistical modelling, the broad focus and theme conditions are grouped in the models below. As discussed in Section 1.1, these are expected to be similar in realization.

Sentences with unaccusative verbs were more likely to be realized with subject stress than unergative verbs (see Table 8). Subject stress was more likely to be realized through stress shift than final subject placement. The two verb types were the most distinct in the broad focus condition, more similar in the QUD-focus and theme conditions, while there seems to be no difference between unaccusative and unergative verbs in the contrastive focus condition.

4.5. Modelling the effects of IS type and verb type

We used nested mixed-effect logistic regression modelling for the analysis of the effect of IS type and verb type on the response form (sV versus Sv versus vS). This seemed the best approach given a repeated measures experiment design and a multinomial dependent (Hilbe, 2009; Hosmer & Lemeshow, 2000). Two nested mixed-effects logistic regression models were fitted using the R package lme4 (Bates, Mächler, Bolker, & Walker, 2015). The first one regressed response form sV (n = 334) and other (a dummy variable for vS and Sv forms, n = 265), while the second model regressed response forms Sv and vS. Both models included response form as the dependent variable and an interaction between IS type and verb type as fixed factors. The models also included random effects for participant and stimulus item. This grouping of the responses directly reflects the research questions: The first model tested the factors that affect the likelihood of nuclear stress on the subject versus the verb (question i); while the second model the factors that affect the likelihood of stress shift versus final subject placement (question ii).

The overall effects of nested models was calculated obtaining significant effects for IS type (χ² = 114.079(2), p < .001), verb type (χ² = 18.166(2), p < .001), and the interaction between these factors (χ² = 10.901(4), p < .05). The analysis of the results for each nested model will be discussed separately.

The first model contrasted verb stress responses (sV) against subject stress (i.e., Sv and vS grouped together). An ANOVA confirmed that there were significant effects of IS type (χ² = 27.57(2), p < .001), verb type (χ² = 14.83(1), p < .001), and their interaction (χ² = 9.50 (2), p = .008) in this model. The fixed effects for this model are shown in Table 9. Figure 3 shows the estimated probabilities from the model of a nuclear stressed subject by IS type and verb type, produced using the R package effects (Fox, 2003). As can be seen, the likelihood of a nuclear stressed subject is highest in the contrast condition, lowest in the broad focus/theme condition, with QUD-focus in between. The two verb types are clearly distinct in the broad focus/theme condition, with the likelihood of a nuclear stressed subject lower for unergative verbs than unaccusative verbs. However, in the contrast condition, there is no difference between the verb types, with a nuclear stressed subject equally likely for both. The QUD-focus condition is again in between. Planned comparisons were run to test which levels of the interaction were significantly different from each other using the R package multcomp (Hothorn, Bretz, & Westfall, 2008), including only the comparisons relevant to the research questions. In order to run the comparison, the LME model was rerun with the interaction between IS type and verb type as a single factor. Results are shown in Table 10. The comparison shows that all three IS types are significantly different from one another for unergative verbs. For unaccusative verbs, only the broad/theme and contrast conditions are significantly different. For the comparison of IS types across verb types the estimates are significantly different in the broad focus/theme condition between unergative and unaccusative verbs, but not in the contrast condition, while in the QUD-focus condition verb type is just significant.

Figure 3

Probability of nuclear stress on subject by IS type and verb type.

The second nested model was fitted to the responses in which the subject word was nuclear stressed (169 Sv and 96 vS); that is, this model looked at the conditions that would produce final subject placement (as opposed to stress shift). An ANOVA showed that verb type was significant in this model (χ² = 5.94(1), p = .015), but not IS type (p = 0.30) nor their interaction (p = 0.21). The non-significant effects could not be excluded from this model, however, as IS type and the interaction were significant in the overall nested model, as reported above. The fixed effects of this model are shown in Table 11. As can be seen, although verb type was significant overall, in the model including the interaction, none of the fixed effects were significant. There was no significant effect of IS type and verb type, or the interaction of these, on the likelihood of final subject placement as opposed to stress shift. Planned comparisons are not reported for this model, as this was not appropriate given the fixed effects were not significant.

4.6. Clefts and sentences with aspectual phrases

As reported in Section 4.1, a small proportion of the responses were clefts or sentences with aspectual phrases (acaba de/empezar a), rather than simple intransitives. These were too infrequent to be separated out in the statistical analysis above, but are discussed further here.

As shown in Table 3, all but two of the cleft sentences were produced in the contrastive focus condition, as would be expected based on previous findings (see Section 1.1); although simple intransitives were much more common in this condition. Nearly all were vS, i.e., these were predominately pseudo-clefts (cf. Section 1.1). There appeared to be no effect of verb type. All but three of these cleft sentences were produced by two speakers. The choice of cleft versus other means of nuclear stressing the subject appears to be down to speaker choice.

As was also shown in Table 3, over half of the sentences with aspectual phrases were produced in the theme condition, with most of the rest in the broad focus condition, although simple intransitives were much more common in both conditions. They also seem to be more common with unergative verbs than unaccusative. Nearly all were sV. The majority of these (62%) were produced by one speaker (a different speaker to the two producing the most clefts). Therefore, although there is a clear association of the use of aspectual phrases with the theme condition, their use is largely idiosynractic to the speaker. To our knowledge, an association between marked themes and the use of aspectual phrases in Spanish has not been noted before. This could be worth following up in future research.