Hebrew copulative perception verbs

3.1 Corpus study

In this section I present distributional data on Hebrew CPVs extracted from the Hebrew Web 2014 corpus (heTenTen14, with automated part-of-speech tagging developed by Meni Adler) available on www.sketchengine.eu (Kilgarriff et al. 2014). This is a multiple domain corpus of 890 million words, made up of texts collected from the Internet, including newspaper materials, Wikipedia articles, blog posts, and both personal and commercial web pages. In presenting the data, I focus on differences between the copulative and verbal constructions, examining their distributions with each CPV, their co-occurrence with a dative argument, and the subjects and complements they attract. The corpus data and the scripts used to analyze them are available online at https://osf.io/g7kw6/.

To make it possible to distinguish between copulative and verbal constructions, data on them are restricted to feminine and plural inflections of the CPVs. This is because adjectival and adverbial complements in masculine singular form are indistinguishable, with very few exceptions. To eliminate false positives where margish ‘feel’ occurs as an experiencer-based verb rather than a CPV, data on it are restricted to instances with a dative argument. This is because margish takes the same types of complements whether it occurs as a CPV or as an experiencer-based verb, and the latter are far more frequent.^[8]

The total occurrences of each CPV in copulative, verbal and impersonal constructions, and their co-occurrence with a dative argument in each construction, are presented in Table 1. Overall, nire ‘look’ is more frequent than nishma ‘sound’ by an order of magnitude, as is nishma in comparison to meriax ‘smell’. The frequency of margish ‘feel’ falls between nishma and meriax.

The distribution of each CPV’s occurrences in copulative and verbal constructions, with and without a dative argument, are visualized in Figure 1. Copulative constructions are considerably more frequent than verbal constructions for nire ‘look’ (∼3:1), nishma ‘sound’ (∼3:1) and margish ‘feel’ (∼4:1). The opposite is true for meriax ‘smell’ (∼1:24). A series of Fisher’s exact tests, with Bonferroni corrections for multiple comparisons, reveals that the distribution of constructions differs significantly between meriax and each of nire (p < 0.001), nishma (p < 0.001), and margish (p < 0.001).

Figure 1:

Distribution of CPVs in copulative and verbal constructions. The darker shading represents instances with a dative argument.

Across verbs, co-occurrence with a dative argument is considerably higher in the copulative construction (between 14.4 % and 44.4 %) than in the verbal construction (between 2.5 % and 4.1 %). A series of Fisher’s exact tests, with Bonferroni corrections for multiple comparisons, reveals that co-occurrence with a dative argument differs significantly between copulative and verbal constructions for each of nire ‘look’ (p < 0.001), nishma ‘sound’ (p < 0.001), and meriax ‘smell’ (p = 0.002).

Impersonal constructions are most frequent for nire ‘look’ (58.5 %), followed by margish ‘feel’ (23.4 %) and nishma ‘sound’ (16.8 %), and finally meriax ‘smell’ (2.2 %). Co-occurrence with a dative argument in the impersonal construction is highest for nire (27.7 %), followed by nishma (10.3 %) and finally meriax (0.0 %).

Next, I employ the exploratory technique of Distinctive Collexeme Analysis (Gries and Stefanowitsch 2004) to identify further differences between copulative and verbal constructions. Distinctive Collexeme Analysis is a method of quantifying the attraction between a lexeme and a construction, in contrast to another construction. Unlike traditional collocational analyses, it focuses on lexemes occurring in particular slots within a construction, rather than all words within a given span.

Here I conduct Distinctive Collexeme Analyses on two slots in the copulative and verbal constructions: the complement and the subject. Due to the vast differences in frequency, I conduct separate analyses for each of the four CPVs. The analyses are performed using Gries’ (2014) R script. In each case, the results indicate which lexemes are attracted to the relevant slot in each of the two constructions, and to what extent, expressed in a measure termed collostructional strength.^[9]

The complete set of results of all eight Distinctive Collexeme Analyses, comprising a total of 6495 lexeme/verb combinations, is available online at https://osf.io/g7kw6/. Below I present side by side comparisons of the top 5 complements (Tables 2–5) and the top 5 subjects (Tables 6–9) attracted to each of the two constructions, for each of the four CPVs. These comparisons reveal striking differences between the copulative and verbal constructions, consistent across the four verbs. Each table shows observed and expected frequencies with both constructions, as well as collostructional strength with the preferred construction. For example, the top row of Table 2 shows that the complements of nire ‘look’ most strongly attracted to the copulative and verbal constructions are hegyoni ‘logical’ and nehedar ‘terrific’, respectively. The observed frequencies of hegyoni are 566 in the copulative construction and 9 in the verbal construction, whereas its expected frequencies (i.e., if it occurred in either construction randomly) are 444.27 and 130.73, respectively. Its collostructional strength with the copulative construction is 50.41 (representing a p-value under 10⁻⁵⁰).

As Table 2 shows, complements of nire ‘look’ attracted to the verbal construction are all subjective in some way (see Section 3.2), and are also virtually unrestricted in terms of what they can be predicated of; any object can be terrific, good, different, etc., whether it’s concrete or abstract, animate or inanimate, perceivable or non-perceivable. On the other hand, complements attracted to the copulative construction are not necessarily subjective, and some can only be predicated of specific classes of objects, e.g., hegyoni ‘logical’ for abstract ideas and raxok ‘far’ for physical locations. Complements that can only be predicated of specific classes of objects are shaded in gray in Tables 2–5.

In addition, Table 2 shows that the observed frequencies of complements attracted to the verbal construction are higher than those of complements attracted to the copulative construction, despite the verbal construction being overall less frequent (see Table 1). This indicates that a relatively small number of complements account for a considerable proportion of all verbal constructions. Indeed, tov ‘good’ alone accounts for 40.4 % of verbal constructions with nire ‘look’, and the top 5 complements taken together account for 77.8 %. In stark contrast, no single complement of nire accounts for even 2 % of copulative constructions.

Turning to complements of the other CPVs, the exact same patterns emerge; see Tables 3–5 for nishma ‘sound’, meriax ‘smell’ and margish ‘feel’, respectively. Complements attracted to the verbal construction are all subjective and, with the exception of naxon ‘correct’, can be predicated of anything. Moreover, these complements are extremely uniform, both within each verb – with very few complements accounting for considerable proportions of all verbal constructions; and across verbs – with the same 8 predicates recurring as the top 5 complements of all 4 CPVs. In comparison, complements attracted to the copulative construction are more varied both within and across verbs, with only 3 predicates recurring among the top 5 complements of the 4 CPVs. Again, many of these complements are not subjective, and many can only be predicated of specific classes of objects, e.g., mufrax ‘unfounded’ for abstract ideas and naki ‘clean’ for physical objects.

The top 5 subjects attracted to each of the copulative and verbal constructions, for each of the four CPVs, are presented in Tables 6–9. As Table 6 shows, all but one of the subjects of nire ‘look’ attracted to the verbal construction denote objects which can be directly perceived via vision, e.g., baxur ‘young person’. In contrast, subjects attracted to the copulative construction are just as likely to denote objects which cannot be perceived in the same way, e.g., she’ela ‘question’. Objects not perceivable through the relevant sensory modality are shaded in gray in Tables 6–9. It should be noted that davar ‘thing’, ish ‘person’ and xayim ‘life’ are attracted to the verbal construction despite having more occurrences in the copulative construction. As high as their observed frequencies in the copulative construction are, they are nevertheless lower than their expected frequencies, and vice versa for the verbal construction.

Data for the other CPVs again match the pattern observed with nire ‘look’; see Tables 7–9 for nishma ‘sound’, meriax ‘smell’ and margish ‘feel’, respectively. In fact, all of the subjects attracted to the verbal construction denote objects directly perceivable through the relevant sensory modality: hearing, smell, or bodily experience. In contrast, several subjects attracted to the copulative construction denote objects which cannot be perceived in the same way, e.g., tsura ‘shape’ via hearing, sipur ‘story’ via smell, and she’ela ‘question’ via touch. It may also be noted that, as Table 7 shows, subjects of nishma attracted to the verbal construction tend to denote musical instruments and audio devices, e.g., (kli) neshifa ‘wind instrument’ and oznit ‘earphone’, whereas subjects attracted to the copulative construction tend to denote linguistic objects, e.g., safa ‘language’.

3.2 Discussion

Summarizing the corpus findings detailed above, nire ‘look’, nishma ‘sound’ and margish ‘feel’ all occur in copulative constructions more frequently than in verbal constructions, whereas meriax ‘smell’ predominantly occurs in verbal constructions. Additionally, nire stands out for frequently occurring in impersonal constructions, which are relatively rare for the other CPVs. Across verbs, copulative and impersonal constructions take dative arguments more frequently than verbal constructions. Complements in verbal constructions are uniform, subjective and can be predicated over anything, whereas complements in copulative constructions are more varied, and more likely to be non-subjective and narrower in terms of what they can be predicated over. Subjects in verbal constructions tend to denote perceivable objects, whereas in copulative constructions they can denote non-perceivable objects.

All in all, the corpus data strongly suggest that the verbal and copulative constructions map to attributary and parenthetical meanings, respectively. Recall that on the attributary meaning, CPVs attribute a property to (a perceptual impression of) an object. This requires an object with a relevant perceptual impression, along with a property applicable to that impression (e.g., Petersen and Gamerschlag 2014). Verbal constructions have parallel requirements, in that their subjects denote perceivable stimuli, and their complements can be predicated of perceptual impressions (since they can be predicated of anything). Moreover, their complements are predominantly subjective, in accordance with previous observations that attributary uses tend to be value judgments (Poortvliet 2018; Rogers 1974; Usoniene 2000; Viberg 2019; Whitt 2009, 2011). The idea that verbal constructions map to attributary meanings also aligns with previous suggestions that complements in attributary uses behave like, have evolved from, or structurally are, adverbs (Gisborne 2010; Postal 1971; Taniguchi 1997).

In addition, we may consider Gisborne’s (2010) diagnostic for identifying the attributary meaning: the impossibility of the continuation but (it) isn’t really. Hebrew verbal constructions are indeed unacceptable with such a continuation, which can be explained by grammatical constraints on ellipsis. Thus, sentences like (9) are ruled out because the elided material can be neither an adverb – which would be missing a verb, nor an adjective – which would be missing an antecedent. Contrast this with the copulative construction in (10), which allows the continuation because the elided adjective has an antecedent.

On the parenthetical meaning, CPVs take a proposition as an argument. If we assume that this proposition may be formed by predicating the verb’s complement over its subject, then the two are only restricted in that the complement must be applicable to the subject, and, possibly, in that the complement must be inferable from the relevant sensory modality (Petersen and Gamerschlag 2014). Copulative constructions show a similar permissiveness, taking subjects which denote stimuli not perceivable through the relevant sense, or at all, as well as complements which cannot be predicated of physical objects or perceptual impressions thereof. The diversity of complements across CPVs may reflect how different properties are inferable from different senses, and the overall paucity of copulative constructions with meriax ‘smell’ may reflect how relatively few properties are inferable from smell, at least for Hebrew speakers.

At this point, we might consider existing analyses of the two constructions. A typical account of the attributary meaning is Petersen and Gamerschlag’s (2014) frame-theoretic analysis (for an analysis along the same lines in model-theoretic semantics, see Muñoz 2019). In their analysis, each CPV specifies an attribute (e.g., sound, taste) of its subject’s denotation, and the CPV’s complement assigns a value to that attribute (e.g., sound: high, taste: bitter). Thus, a felicitous attributary use must satisfy two constraints: the attribute specified by the verb must be appropriate for the subject’s denotation, and the value assigned by the complement must be appropriate for that attribute. The former blocks attributary uses with subjects denoting stimuli which are not perceivable via the relevant perceptual modality (e.g., #the music looks X, #the dish sounds Y), and the latter blocks attributary uses with complements which cannot be predicated of relevant perceptual impressions (e.g., #X looks logical, #Y sounds blue).

However, if we apply Petersen and Gamerschlag’s (2014) analysis to the Hebrew verbal construction, we over-generate which complements the construction takes. If the construction simply requires a complement which can be predicated of the relevant perceptual impression, why is it that complements are overwhelmingly subjective? Granted, some predicates might be ruled out because they have no adverbial form; this could explain why there are no color terms among the complements of the verbal construction, for instance. But there are predicates in Hebrew which can be predicated of perceptual impressions, and do occur as adverbs, yet never occur in the verbal construction; see (11)–(13).

Before we proceed, we should clarify what it means for a predicate to be subjective. This issue has received increasing attention in the past two decades, but remains thorny from both an empirical and a theoretical perspective (e.g., Kennedy 2013; Lasersohn 2005; McNally and Stojanoivc 2017). The most widespread diagnostic for identifying subjective predicates is whether they give rise to faultless disagreement, the intuition that speakers can disagree about certain propositions, e.g., whether something is tasty, without any one of them being objectively wrong (Kölbel 2004). However, experimental evidence shows that faultless disagreement arises to different extents for different predicates (Solt 2018) and depends on social consensus as much as on predicate choice (Kaiser and Rudin 2020). Moreover, faultless disagreement arises from at least three different sources: dimensional vagueness, multidimensionality, and judge-dependence (Kennedy 2013; McNally and Stojanovic 2017; Solt 2018).

Dimensional predicates such as strong and high may allow faultless disagreement due to vague standards. For instance, speakers might disagree about whether or not Rachel is strong, due to having different standards for what counts as strong in context. Notably, faultless disagreement disappears when these predicates occur in comparative and superlative constructions, e.g., Rachel is stronger than Adam.

Multidimensional predicates such as smart and healthy allow faultless disagreement by involving multiple criteria which can be used to determine whether, and to what degree, an object has the property in question. For instance, speakers might disagree about whether Rachel is healthier than Adam, due to having different considerations in determining how healthy someone is, e.g., their blood pressure, immune system, etc. Multidimensional predicates can be identified with phrases which explicitly target their dimensions, e.g., healthy with respect to X (Sassoon 2013).

Judge-dependent predicates such as fun and tasty give rise to faultless disagreement by involving a judgment regarding taste, value, etc. For instance, speakers might disagree about whether Rachel is more fun than Adam, because the degree to which something is fun is a matter of personal judgment rather than fact. Judge-dependent predicates can often – though not always – be identified by taking an optional argument representing the judge, e.g., fun for X (Lasersohn 2005; McNally and Stojanovic 2017).

For present purposes, it is clear that dimensional vagueness alone does not license complements in the verbal construction. As the examples in (11a)–(13a) above show, the dimensional complements xazak ‘strong’, xalash ‘weak’, gavoha ‘high’, namux ‘low’, and xalak ‘smooth’, are not acceptable in the verbal construction, even in non-comparative forms. As for multidimensionality and judge-dependence, most of the predicates attracted to the complement slot of the verbal construction are both multidimensional and judge-dependent (tov ‘good’, nehedar ‘terrific’, etc.). But are either of these a necessary or sufficient condition for complements in the verbal construction? Paradigmatic multidimensional predicates which are not judge-dependent, xaxam ‘smart’ and bari ‘healthy’, and likewise paradigmatic judge-dependent predicates which are not multidimensional, kef ‘fun’ and taim ‘tasty’, are ruled out for independent reasons – they have no adverbial form, or cannot be predicated of relevant perceptual impressions. There are, however, multidimensional predicates which are not judge-dependent, and which are attracted to the verbal construction, most prominently axeret ‘different’ but also other (dis)similarity predicates such as dome ‘similar’ and shone ‘different’. In contrast, only one judge-dependent predicate which is not multidimensional – naxon ‘correct’ – is attracted to the verbal construction to a comparable extent, and only with the verb margish ‘feel’.

It appears, then, that the verbal construction attracts multidimensional predicates to its complement slot. To account for this, I propose that the CPV in the verbal construction interacts directly with the dimensions of its complement. That is, rather than specifying an attribute of its subject’s denotation, as suggested by Petersen and Gamerschlag (2014), the CPV restricts the interpretation of its complement to a subset of dimensions: visual dimensions for nire ‘look’, auditory dimensions for nishma ‘sound’, etc. This functions similarly to phrases which directly target a multidimensional predicate’s dimensions, e.g., with respect to X (Sassoon 2013; see also Alrenga 2010), such that (14a) is analyzed as equivalent to (14b). If the complement is not a multidimensional predicate, the CPV has no dimensions to target and composition fails; see (15a) and (15b). Likewise, if the resulting predicate – restricted to a specific subset of dimensions – cannot apply to the subject’s denotation, composition fails; see (16a) and (16b).

We can see that the CPV restricts its complement to a subset of dimensions, and not to a single dimension, from the possibility of it co-occurring with another phrase targeting a predicate’s dimensions. Thus in (17a), the predicate niret muzar ‘look weird’ is restricted to visual dimensions but remains multidimensional, allowing the phrase mi-bxinat tseva ‘with respect to color’ to further restrict its interpretation to a single visual dimension, color. In contrast, (17b) with the phrase mi-bxinat ta’am ‘with respect to flavor’ is unacceptable, because the subset of dimensions which niret muzar ‘look weird’ is restricted to, does not include the dimension of flavor.

The present proposal explains why CPVs in the verbal construction so rarely co-occur with a dative argument in the corpus (between 2.5 % and 4.1 %). Unlike the copulative and impersonal constructions (see below), the verbal construction does not actually have a semantic role for the dative argument to fill (see Gisborne 2010). In the few cases where the CPV does take a dative argument, it likely fills the judge role required by a judge-dependent complement, e.g., tov ‘good’, ra ‘bad’, which licenses a dative argument on its own.

Moving on to the copulative construction, the CPV appears to function as a raising verb: its subject and complement combine to form a proposition, which the CPV takes as an argument and modifies (see also Avineri 2021, who analyzes the complement in this construction as a small clause). The present finding that the copulative construction attracts abstract subjects lends credence to analyses which don’t require the subject to be a source of evidence (Gisborne 2010; cf. Muñoz 2019). The remaining question is whether CPVs in the copulative and impersonal constructions are semantically equivalent. Particularly, does the CPV in each construction have an evidential meaning, an epistemic modal meaning, or both?

Note that I use the term evidentiality to refer to any encoding of information about the source or type of evidence for a proposition, whether or not it is grammaticized. I distinguish a priori between evidentiality and epistemic modality, which refers to the encoding of an individual’s beliefs or knowledge. It’s possible, however, for an expression to encode both evidential and epistemic meaning, and in fact, Matthewson (2012) explicitly claims that every expression which encodes one also encodes the other. Even if an expression only encodes evidential meaning, its use may convey epistemic information via pragmatic inference, and vice versa (Degen et al. 2019). This makes it difficult to tease apart evidential and epistemic meaning within corpus data. Thus, to test whether the copulative and impersonal constructions differ in encoding evidential and epistemic meaning, I conducted a preference experiment.

3.3 Preference experiment

The experiment was designed to test whether copulative and impersonal constructions differ in encoding evidential and epistemic meanings. I restricted the experiment to nire ‘look’ and nishma ‘sound’, since meriax ‘smell’ and margish ‘feel’ occurred in impersonal constructions very rarely in the corpus. The experiment relied on contradictions to the evidential and epistemic meanings presumably encoded by the CPV, specifically direct evidence of an object (e.g., Landau 2011) and compatibility with the speaker’s knowledge (e.g., Hansen and Markman 2005). I assumed that, if one construction encoded evidential meaning but the other did not, then the latter would be preferred in contexts contradicting the evidential meaning. Likewise, if one construction encoded epistemic meaning but the other did not, then the latter would be preferred in contexts contradicting the epistemic meaning.

The experiment consisted of categorical preference judgments. In each trial, participants were presented with a short context introducing two speakers discussing an object, and an incomplete sentence by one of the speakers. Participants were asked to indicate which of two possible continuations would be better, that is, would complete the speaker’s sentence in a way that sounds more natural, logical and reasonable. Participants could also choose to indicate that neither continuation was better than the other.

In every trial, the presented incomplete sentence contradicted either the evidential or the epistemic meaning presumably encoded by the CPV. For the evidential meaning, the incomplete sentence asserted that the speaker had no direct evidence of the object under discussion, of the type associated with the CPV (seeing for nire ‘look’, hearing for nishma ‘sound’). For the epistemic meaning, the incomplete sentence asserted that the speaker had prior knowledge contradicting the proposition argument of the CPV. In the critical trials, the two continuations presented to participants were one copulative and one impersonal construction, both with the same CPV. To ensure that a contradiction arises, all continuations included a 1st person dative argument. Examples of critical trials, translated from Hebrew, are given in (18) and (19) (in the original Hebrew, the dative argument immediately follows the CPV in both constructions, and there is no expletive subject in the impersonal construction). The full list of materials is available online at https://osf.io/g7kw6/.

A total of 32 critical trials were created, half with no direct evidence, and half with prior knowledge. Half of each included nire ‘look’ in their continuations, and the other half included nishma ‘sound’. 16 filler trials were also created, in which the two continuations presented to participants were instances of the same construction, with two different CPVs. The experiment was created in Google Forms, and participants were recruited via social media. 41 native speakers of Hebrew volunteered to participate. Before completing the experiment, participants saw two example trials which did not include CPVs, and were advised that there were no right or wrong answers. Each participant then saw all 48 trials, in a pseudo-random order.

The results of the preference experiment are presented in Figure 2. In trials where the speaker had prior knowledge contradicting the proposition argument of the CPV, participants consistently preferred the copulative construction, both with nire ‘look’ (76.8 %) and with nishma ‘sound’ (80.5 %). In trials where the speaker had no direct evidence of the object in question, speaker preferences depended on the CPV in the available continuations. With nire, speakers generally preferred the impersonal construction (67.8 %), but with nishma, there was no preference for either construction (35.1 % copulative, 30.5 % impersonal, 34.5 % no preference).

Figure 2:

Results of the preference experiment. Participants chose whether the copulative or impersonal construction was better in trials where the speaker had prior knowledge contradicting the proposition argument of the CPV, and in trials where the speaker had no direct evidence of the object in question.

These results strongly suggest that CPVs encode compatibility with the speaker’s beliefs – an epistemic modal meaning – in the impersonal construction, but not in the copulative construction. The impersonal construction is consistently dispreferred when the speaker’s knowledge contradicts the CPV’s proposition argument, similarly to how epistemic might or must are unacceptable when their prejacent contradicts the speaker’s knowledge.

In contrast, the results suggest that nire ‘look’ encodes direct evidence of the subject – evidential meaning – in the copulative construction, but not in the impersonal construction. The copulative construction is consistently dispreferred when the speaker has no direct visual evidence of the object in question, regardless of whether they believe that evidence. Note that direct evidence of the subject is a default assumption rather than a logical entailment, given the corpus finding that the copulative construction attracts abstract subjects which cannot be seen (e.g., she’ela ‘question’, matara ‘goal’). Additionally, the results do not preclude the impersonal construction from encoding a weaker evidential meaning, e.g., visual evidence from an unspecified source.

The picture is not as clear-cut for nishma ‘sound’. Since there was no preference for either construction in trials where the speaker had no auditory evidence of the object in question, it’s possible that both constructions were acceptable, or that both constructions were unacceptable. I hypothesize that participants found both constructions acceptable, not because nishma lacks evidential meaning, but because it allows indirect reported evidence in addition to direct auditory evidence (see also Gisborne 2010; Landau 2011; Viberg 2019).

These results raise a possible explanation for why the impersonal construction occurs in the corpus almost exclusively with nire ‘look’. If the meaning of the impersonal construction is primarily epistemic rather than evidential, it essentially doesn’t matter which CPV is used. Thus nire, as the most frequent CPV to begin with, might be used by default.