Semantic similarity to high-frequency verbs affects syntactic frame selection

1 Introduction

Most English verbs can occur in more than one subcategorization frame, but not with equal frequency. Consider the verb give. In the Ditransitive or Double Object frame (1a), which give occurs most frequently in, the recipient argument is realized as an NP and precedes the gift argument realized also as NP. In the Prepositional Dative or Prepositional frame (1b), which give occurs much less frequently in, the recipient of the gift is described via a PP, typically following the NP that describes the gift.

This biased frequency distribution is often referred to as verb subcategorization frame preference or more simply as structural or syntactic preference. Structural preferences, as we will call the phenomenon from now on, are the norm for verbs (e.g. Ford et al. 1982) and are used by native speakers during online sentence comprehension and production (e.g. MacDonald et al. 1994; Jurafsky 1996; Hare et al. 2003, among many others). Quite often, as Hare et al. (2003) note, a verb has clearly distinct meanings in the different subcategorization frames it can occur in (e.g. admit students [into a major] vs. admitting the tax cut was a bad idea). Structural preferences, in this case, may reflect this important difference in meaning (including which meaning is more likely to be of relevance in different corpora, see Roland et al. 2007). Choice of verb sense, in this case, determines the choice of subcategorization or syntactic frame (we use both terms interchangeably). But quite often, the meaning of a verb does not change much across some of the subcategorization frames it occurs in; the verb has the same sense in the two syntactic frames. This is the case for the dative alternation exemplified in (1). Verbs that alternate between the prepositional and ditransitive frames describe similar kinds of events. Although there is a slight difference in meaning for many of them (in some cases, the difference is one of construal in Langacker’s 1987 sense), a single verb sense is involved in both frames, as give in (1) illustrates. For valence alternations like the ditransitive alternation, then, speakers or writers have a choice of syntactic structure to convey roughly the same message. Despite having the same sense describing very similar kinds of events in both syntactic frames, most alternating verbs still exhibit structural preferences. The goal of the present study is to examine the role of the meaning of high frequency verbs in the choice of syntactic frame for argument structure alternations where verb sense remains constant and speakers have a choice to make when conveying their message.

Many previous studies of the dative alternation have investigated properties of post-verbal arguments that modulate speakers’ choice of syntactic frame (Collins 1995; Thompson 1990; Wasow 2002; Bresnan et al. 2007; Theijssen et al. 2013). Summarizing this line of research, the ditransitive frame tends to be preferred if the recipient argument is semantically animate and definite and conveys pragmatically given information; it is also preferred if the syntactic expression of the recipient argument is pronominal and/or morphophonologically short. Bresnan and her colleagues (2007) showed that all those factors can help predict a speaker’s choice of the ditransitive or prepositional frame. It is important to note, though, that, to date, the majority of the factors shown to modulate the choice of syntactic frame are properties of the arguments that co-occur with the verb, not semantic properties of the individual verb chosen by the speaker.

The impetus for our research is the robust finding that a few verbs occur in a syntactic frame far more frequently than others and often a single verb occurs frequently enough to be considered semantically representative of that frame, e.g. go, do, get, take, put, give, etc. (Carroll et al. 1971; Gropen et al. 1989; Goldberg 1999, 2006; Stefanowitsch and Gries 2003; Ellis et al. 2014). This small number of verbs tend to be general-purpose verbs and, as a result, are used in a wide variety of contexts and situations. They usually are the first verbs to be learned by children (Clark 1978) and considered to play a significant role in second language acquisition as they serve as prototypical verbs for particular syntactic frames. They can thus be considered pathbreaking verbs (Ninio 1999; Ellis and Ferreira-Junior 2009; Ellis and O’Donnell 2011, 2012). Furthermore, Goldberg and her colleagues (Goldberg et al. 2004; Casenhiser and Goldberg 2005) provide evidence that the fact that the frequency of occurrence of verbs in a syntactic frame is highly skewed actually facilitates the acquisition of a construction. They report that participants acquire artificial grammars more easily and more accurately when they are exposed to syntactic structures with skewed verb frequencies than when they are exposed to syntactic structures with evenly-distributed verb frequencies.

In this study, we explore the possibility that the verb most frequently used in a syntactic frame attracts other verbs with similar semantic and syntactic profiles to that frame and thus serves as an anchor for these verbs. More specifically, we hypothesize that semantic similarity to the anchor verb modulates speakers’ choice of syntactic frame for those verbs. We call our hypothesis the Verb Anchor Hypothesis. We call a verb that occurs highly frequently in a frame and that modulates speakers’ choice of frame for other verbs an anchor for the frame. Using the dative alternation, we investigate whether some verbs can serve as anchor of either frame and whether semantic similarity between these anchors and other alternating verbs modulates speakers’ choice of frame, i.e. whether the more similar a verb is to an anchor, the more likely it is to occur in the same frame that the anchor is biased towards. Our ultimate goal is to show that semantic similarity to an anchor verb is one of the causes of structural preferences when the message conveyed remains relatively constant across syntactic frames.

The motivation behind our hypothesis is the assumption that the high frequency of occurrence of a verb in a syntactic frame leads to a strong cognitive association between that verb and that frame in the same way that a highly frequent exemplar of a category is strongly associated with the category (see Hintzman 1986; Medin and Schaffer 1978; Nosofsky 1988, among others on categorization in general and Lakoff 1986; Langacker 1987 for the role of categorization in language). Research on categorization shows that the more similar an entity is to the most frequent member of a category, the more likely it is to be considered a member of that category. Several previous studies support our assumption that the relationship between a verb and a syntactic frame is cognitively analogous to the relationship between an exemplar and a category (e.g. Johnson and Goldberg 2013; Snider 2008). First, syntactic frames are not explicitly taught but are abstracted from repeated exposure to sentences with various verbs in the same way categories are extracted from exemplars (Goldberg et al. 2007; Tomasello 1992). Second, syntactic frames often display typicality effects. For example, speakers tend to think of the most typical verb when asked to name a verb that can occur in a particular syntactic frame. If asked to provide a verb that would fit the word string “A man ___ a kid a toy,” people are most likely to come up with the verb give (Goldberg 1995; Ellis et al. 2014). Importantly for our concerns, these typicality effects seem to be a consequence of the fact that verbs occur in syntactic frames with different frequencies.

We assume two joint causes for the putative effect the meaning of verb anchors may have on production. First, verbs overlap in meaning and the concept that a speaker wishes to express activates the meaning of verbs it overlaps with (see, among others, Dell 1986). For example, McRae and Boisvert (1986), among many others, showed that words are primed to the degree to which two words are semantically similar to each other. Second, verbs whose meanings are activated activate in turn the syntactic frame(s) they occur in. So, when a verb is more similar to a frame’s anchor, the anchor’s preferred frame is more strongly activated. It is this stronger activation of the anchor’s preferred frame, we hypothesize, that explains the role of the anchor verb’s meaning on production. We talk for most of this paper as if there was only one anchor for a syntactic frame. However, since each verb is more or less semantically similar to all other verbs and each verb is more or less associated with a syntactic frame, there can be multiple putative anchors. But, some verbs play a particularly important role, as they occur particularly frequently in a frame and this paper focuses on the impact these verbs have on syntactic frame selection, both because the literature on argument structure has stressed the important role these verbs play and because it is initially most appropriate to test our Verb Anchor Hypothesis on verbs whose putative effect on syntactic frame selection is the largest.

Finally, note that the Verb Anchor Hypothesis only depends on one situation category (a verb meaning) activating similar situation categories (verbs with similar meanings) and the strength of association between verbs and the alternating frames they occur in. It does not depend on and is agnostic about whether syntactic frames themselves are associated with a meaning as scholars such as Pinker (1989), Goldberg (1995), or Ambridge et al. (2014) have argued. We return to this issue in the General Discussion.

The organization of this paper is as follows. Section 2 presents the results of an extensive corpus study that investigates the frequency distribution of verbs in the ditransitive and prepositional frames and explains how we estimated the strength of cognitive association between a verb and a syntactic frame. Section 3 provides details on how we measured semantic similarity between verbs and presents logistic regression models of our corpus data that support our Verb Anchor Hypothesis. Section 4 investigates whether semantic similarity to the anchor verb makes a unique contribution to the prediction of syntactic frame selection above and beyond other factors known to affect the choice of syntactic frame. Section 5 examines whether semantic similarity to an anchor verb affects syntactic frame selection within narrower semantic classes than the entire class of alternating verbs. Section 6 concludes the paper.

2 The dative alternation in the British National Corpus

Previous researchers have shown that the frequency distribution of verbs in the ditransitive and prepositional frames is highly skewed (e.g. Gropen et al. 1989). In this section, we corroborate previous research with a study of all verbs listed as alternating in Levin (1993). Our corpus was a version of the British National Corpus (BNC) syntactically annotated by an automatic parser (Charniak 1997). ^[1] The results we report provide an overview of the frequency distribution of verbs that occur in the two frames in that corpus. ^[2]

We first retrieved from the parsed BNC the verb phrases that instantiate the ditransitive or prepositional frame, i.e. [V NP NP] or [V NP PP]. We then discarded sentences whose main verb was not one of the 122 verbs that Levin (1993) listed as participating in the dative alternation. Levin’s original list includes 127 verbs. The number of verbs we started with, however, was 122, as five verbs are listed twice in Levin’s list but with two different senses and our parsed BNC cannot discriminate between verb senses. 13 verbs from Levin’s list occurred in neither frame in our corpus (schlep, tote, bus, truck, modem, netmail, satellite, semaphore, telecast, telex, wireless, bunt, and punt). We excluded render and vote by hand because most of their tokens did not instantiate the meaning “caused possession.” Finally, we had to remove pass and relay as well because they have two distinct senses, according to Levin, one of which entails caused possession irrespective of syntactic frame (or, more specifically, entails caused possession in a restricted set of possible worlds, as argued in Koenig and Davis 2001) and one of which does not (Rappaport and Levin 2008). ^[3] Since entailing (or not entailing) caused possession irrespective of syntactic frame is a semantic factor in some of our models and our automatic annotation of the BNC cannot distinguish between verb senses, we had to omit these two verbs from our study. This process left us with 105 distinct verbs and 62,713 sentences or tokens of either the ditransitive or prepositional frame.

As shown in Table 1, not only does our corpus study confirm the extremely high frequency of give in the ditransitive frame, it also shows that give differs from other alternating verbs in its relative frequency of occurrence in the ditransitive and prepositional frames. Give is overwhelmingly more frequent in the ditransitive frame than any other verb. Give accounts for 58% of all ditransitive tokens while the 104 other verbs account for only 42%. The second most frequent verb in the ditransitive frame is tell, which only accounts for 10% of the tokens of the ditransitive frame. Additionally, give occurs more frequently in the ditransitive frame than in the prepositional frame (D:P=65:35) whereas the vast majority of other verbs occur more frequently in the prepositional frame (D:P=28:72 is the mean distribution for the other 104 verbs). Give’s strong preference for the ditransitive frame stands out as most alternating verbs have a preference for the prepositional frame. These two distributional facts (relative and absolute frequency of give in the ditransitive frame) suggest that there may exist a strong cognitive association between give and the ditransitive frame and that the relation between the ditransitive frame and give may be analogous to the relation between a category and its most typical exemplar.

A few verbs (email, ask, and tell) show an even stronger bias for the ditransitive frame than give. But, none of them, we argue, constitutes a better exemplar of the ditransitive frame than give as they occur much less frequently in the ditransitive frame. What increases the strength of association between a verb and a syntactic frame is, we claim, not only the preference the verb exhibits for that frame (the relative frequency of occurrence of a verb in a syntactic frame), but also the token frequency of occurrence of the verb in that syntactic frame (its absolute frequency of occurrence in that frame). In other words, a verb is the best exemplar or anchor of a syntactic frame when (1) it occurs very frequently in that frame and (2) it strongly prefers that frame over the alternation’s competing frame. When both conditions are met, not only should the verb evoke the syntactic frame but the frame should conversely evoke the verb. Such is the case for the verb give and the ditransitive frame, whereas verbs like email and ask do activate the ditransitive frame, but the ditransitive frame does not strongly activate either ask or email as their total number of occurrences in the frame is very low.

There are various ways to quantify the strength of association between a verb and a frame. First, since our goal is to test factors that affect the choice of syntactic frame when much of the meaning of the verb remains constant between the two syntactic frames, we computed the association strength a verb has with the ditransitive frame, considering the prepositional frame as an only alternative. The meaning of verbs in other frames may have little to do with the meaning they have in these two frames (cf. This idea did not take). Second, we examined three possible measures of association strength: Gries and Stefanowitsch (2004) collocational strength, Ellis and Ferreira-Junior’s (2009) ΔP (more precisely, ΔP Attraction or ΔP Construction → Word), ^[4] and a measure based on Hebb’s (1949) learning rule (see Proulx and Hélie 2005; McClelland 2006 for recent discussion of the role Hebbian learning may play in human cognition). As each measure has advantages and drawbacks (see Schmid and Küchenhoff 2013 for a discussion and comparison of various verb-syntactic frame association measures), the use of multiple convergent measures helps ensure the robustness of our models of speakers’ choice of syntactic frame. Despite important conceptual differences between the three measures, they agree on all the anchors we make use of in this paper. All measures we considered confirm that the verb give is most strongly associated with the ditransitive frame and is far more so than any other verb. More details about the three measures and the full list of verbs and connection strengths estimates for each measure are provided in the Supplementary Material.

We selected our anchor verb from the rank-ordered list of association strengths. We surmise that the magnitude of association strength matters (both absolute strength and, possibly, the difference in strength between the top-ranked and the second ranked). Unfortunately, we do not know of any quantitative criteria by which we can decide what magnitude of association strength indicates a strong enough association for an anchor effect to occur. Furthermore, magnitude of strength of association may depend on the particular measure of association strength one chooses. We therefore only consider rank in selecting anchors in this paper.

3 Semantic similarity to give predicts the choice of syntactic frame

The Verb Anchor Hypothesis predicts that high semantic similarity to a frame’s anchor facilitates the occurrence of other alternating verbs in the same frame. Since give is the ditransitive frame’s anchor, high semantic similarity to give predicts a verb to be more likely to occur in the ditransitive frame than in the prepositional frame. We use logistic regression to test whether a verb’s semantic similarity to an anchor verb, e.g. give, is a significant predictor of the dative alternation, entering the semantic similarity between the anchor (give) and an alternating verb as predictor and the syntactic frame of any sentence that includes that alternating verb as an outcome variable.

To prepare the outcome or dependent variable for all logistic regression analyses, we coded as either ditransitive (1) or prepositional (0) frame every sentence collected from the parsed British National Corpus. We also identified the verb in every sentence so that we could measure its semantic similarity to give. Sentences whose main verb was give were excluded from the model. Since our prediction is that verbs that are more semantically similar to give will tend to occur in the dominant frame for give, including instances of give would inflate our results.

To estimate semantic similarity between alternating verbs and give (our predictor), we used Latent Semantic Analysis (LSA hereafter; Landauer et al. 2007), which is one of the distributional models of semantic memory developed in computational linguistics. The tenet of this approach is often described by Firth’s (1957) wording “you shall know a word by the company it keeps.” LSA computationally and statistically simulates the contextual usage or overlap of words and computes their similarities or relatedness using natural language corpora that are meant to reflect our experience with language (Landauer et al. 1998). The intuition behind LSA is that the similarity in meaning of two words or sets of words can be estimated by their co-occurrence and contextual overlap: Do the two words occur in the same documents? Do the two words occur with the same set of words even if they do not co-occur directly? (See Deerwester et al. 1990; Landauer and Dumais 1997; Kwantes 2005 for technical details.) Taking an example in Jones et al. (2015), robin and egg may be related because they often directly co-occur with each other. Robin and sparrow may also be related because they frequently occur in similar contexts or with the same set of words, although they rarely co-occur directly.

Crucially, LSA does not take into account syntactic differences in computing semantic similarity. That is, when LSA compiles the corpus term-by-document frequency matrix, the document is treated as a “bag of words” and does not include any transitional or syntactic information. For example, the LSA cosine of the ditransitive (she gave the boy the book) and prepositional (she gave the book to the boy) sentences is 1 (the highest similarity), a reflection of the irrelevance of syntactic frames in LSA estimates of semantic similarity. This insensitivity to syntactic differences is critical for our purposes as it means that LSA does not produce high estimates of the semantic similarity between two verbs just because they often occur in the same syntactic frame. The insensitivity of LSA to syntactic context is particularly appropriate for our study, as we investigate the role of verb similarity on syntax, not the role of syntax on verb similarity. We used LSA cosines (0 ~ 1) as our primary measure of verb similarity, but also tested the predictive role of a WordNet-based similarity metric (Pedersen et al. 2004) and GloVe vectors (Pennington et al. 2014) to ensure that our findings generalized across measures of semantic similarity. ^[5]

Our LSA was trained on the British National Corpus, used 400 dimensions and treated paragraphs as documents. In applying LSA, we chose to use past tense forms of verbs in our pairwise verb comparisons (gave vs. the past tense form of each of the 104 alternating verbs) in order to minimize inconsistencies across verbs. We chose past tense forms because close review of individual pairwise comparisons revealed that cosines tend to be slightly higher for verbs whose non-third singular present tense/bare infinitive forms are identical to their noun forms, e.g. to offer and an offer. The LSA cosines for our 104 pairwise comparisons ranged from 0.047 to 0.946. ^[6]

In reporting the models, we present the coefficients (b) and the z-statistic of individual predictors. We centered all predictor variables to minimize problems due to multicollinearity (Aiken and West 1991), but for any models where multicollinearity may be an issue (e.g. VIF>6), we provide their Variance Inflation Factor (VIF) (Cohen et al. 2003). When model selection is discussed, we will also report Akaike Information Criterion (AIC) estimates, i.e. ΔAIC and weight (Wagenmakers and Farrell 2004).

The results of our first model showed that, as predicted by the Verb Anchor Hypothesis, a verb’s semantic similarity to give contributes to predicting the use of the ditransitive frame (b=0.28, z=22.53, p<0.001). This result suggests that the frequency-driven association between an anchor verb like give and the ditransitive frame plays a significant role in predicting speakers’ choice of the ditransitive frame. To address the concern that the result was unexpectedly driven by our choice of LSA cosines as verb similarity measure, we tested our hypothesis with a different estimate of semantic similarity, the WordNet-based vector measure (Patwardhan 2003), to a subset of our data (22,787 sentences, D=8,031; P=14,756) that included the 49 alternating verbs that occur in four narrow classes in Pinker’s (1989) sense, namely give verbs (but excluding give), message transfer verbs, future having verbs, and send verbs. ^[7] We chose these four classes because they constitute the main narrow classes of alternating verbs and choosing the “right” senses proved difficult for many verbs in other classes (often, none of their senses in the WordNet dictionary was appropriate for their ditransitive use, see Footnote 5 for details of using WordNet similarity). We used the mean of the vectors between verb senses as an estimate of verb similarity. This model replicated the effects of similarity to give on syntactic frame selection (b=0.12, z=8.37, p<0.001) that we found when using LSA as our measure of semantic similarity. This study is the first, to our knowledge, to quantitatively investigate the role that the meaning of the most typical exemplar of a syntactic frame (what we call an anchor verb) plays in the process of syntactic frame selection.

The Verb Anchor Hypothesis can also be tested via a correlation analysis when only one factor is taken into account. It is confirmed if there is a significant positive correlation between an alternating verb’s semantic similarity to give and its proportion of uses in the ditransitive frame. And indeed, we found a significant correlation (Pearson’s r=0.295, p<0.001). As illustrated in Figure 1, however, many of the putative alternating verbs occur in the ditransitive frame very rarely. In fact, 47 verbs never occurred in the ditransitive frame in our dataset. The correlation between semantic similarity to give and proportion of occurrence in the ditransitive frame is stronger when only the 41 verbs that entail caused possession irrespective of syntactic frame (give type verbs, cf. throw-type verbs) are included in the analysis (Pearson’s r=0.639, p<0.001).

Figure 1:

Plots of the correlation between verbs’ semantic similarity to give (x-axis) and their proportions of occurrence in the ditransitive frame (y-axis) with regressional (dashed) and lowess (dotted) lines.

Although logistic regression and correlation analyses both support the Verb Anchor Hypothesis, logistic regression better serves the overall purposes of our study. Firstly, as each verb’s proportion of ditransitive uses constitutes a single data point in a correlation analysis, results are significantly affected by a few obvious outliers, an issue that is particularly problematic for narrow-class analyses we discuss in Section 5. Second, in contrast to logistic regressions, a correlation analysis does not allow us to simultaneously evaluate multiple predictors of choice of syntactic frame and compare our results with those of previous studies, in particular Bresnan et al.’s (2007).

Given that constancy of caused-possession “entailment” affects the strength of the correlation between semantic similarity to give and frequency of occurrence in the ditransitive frame, we conducted a second logistic regression analysis that included as additional predictor whether a verb entails caused-possession irrespective of syntactic frame. The results of this model show that both semantic predictors make distinct significant contributions to the prediction of speakers’ choice of syntactic frame in the same direction (always entailing caused possession, b=1.11, z=71.24, p<0.001; similarity to give, b=0.96, z=56.17, p<0.001), suggesting that the likelihood of a speaker choosing the ditransitive frame increases when the verb always entails caused possession and is semantically similar to give. In the next section, we test whether these two verb semantic properties remain significant predictors of syntactic frame selection when non-verb-semantic or verb-external factors are taken into account.

4 Combining verb-internal and verb-external factors

Most previous studies have paid attention to factors only tangentially related to verbs, for example, whether the recipient argument is pronominal or definite, whether the theme argument represents given information, or what the difference in length is between the two post-verbal arguments. Bresnan et al. (2007) conducted a comprehensive analysis of a natural language corpus and showed that these verb-external factors do predict, together and individually, the dative alternation. In this section, we are concerned with whether the verb-semantic predictors we examined in the previous section remain significant when these other factors are taken into account.

To that effect, we conducted regression analyses using Bresnan et al.’s (2007) dataset. ^[8] This dataset collected from the Switchboard and the Wall Street Journal corpus is smaller in size and number of verbs involved than the BNC data we used previously but it is manually annotated with nearly the full set of factors known to affect the dative alternation. ^[9] The dataset consists of 3,265 sentences and only 75 different verbs are observed in these sentences. As opposed to the BNC-based corpus we used in the previous analyses whose sentences contain only the verbs that Levin (1993) listed as alternating, 34 verbs from Bresnan et al.’s dataset are not on Levin’s list. These 34 verbs account for about 10% of the total tokens (329 sentences) and three of them (cost, charge, and do) make up three quarters of those tokens (242 ditransitive and 2 prepositional sentences). We excluded sentences that contained these 34 verbs from our analysis to allow for a better comparison of the results of this model with those of the previous models. As with all other models, sentences that contained the anchor verb give were excluded to prevent an artificial boost in the effect of semantic similarity to give. Table 2 presents an overview of the distribution of verbs in the Bresnan et al.’s (2007) dataset.

As in the BNC, the verb give is by far the most frequent verb in the ditransitive frame in Bresnan et al.’s corpus. It accounts for 63% of all ditransitive tokens (cf. 58% in Table 1). Also, give seems to occur even more frequently in the ditransitive frame (D:P=85:15) (cf. D:P=65:35 in Table 1) and, in contrast to our corpus study, other verbs also show a slight preference for the ditransitive frame (D:P=55:45, cf. D:P=28:72 in Table 1). These differences may be a consequence of the larger proportion of spoken language in Bresnan et al.’s (2007) corpus, which tends to contain more pronominal expressions for animate entities (you, me, him, them, etc.) than written language corpora. Pronominal recipients are known to favor the choice of ditransitive frame.

To assess the contribution of our verb-internal predictors, we first replicated Bresnan et al.’s (2007) model using only their predictors. The model (I) provides us with a baseline to which we can compare our model. A new model (II) additionally includes both of our lexical semantic predictors, i.e. semantic similarity to give and caused-possession entailment. Note that Bresnan et al. (2007) coded the ditransitive frame as 0 and the prepositional frame as 1, the opposite of what we did in our previous models. For ease of comparison we follow Bresnan et al.’s (2007) coding in this section and the predictor variables are not centered, just as in their study. When interpreting the results, a positive coefficient indicates that the predictor decreases the likelihood of the selection of the ditransitive frame (and increases the likelihood of prepositional frame selection), whereas a negative coefficient indicates that the predictor increases the likelihood of the selection of the ditransitive frame (and decreases the likelihood of prepositional frame selection). The coefficients and the significance levels of each predictor are summarized in Table 3.

The results show that semantic similarity to give and the entailment of caused possession (in shade in Table 3) make an independent contribution above and beyond the predictors Bresnan et al. discussed. Comparing Models (I) and (II), all the coefficients and significances of verb-external predictors remain essentially the same whether or not semantic similarity to give and an invariable caused-possession entailment are included as predictors. Model comparison showed that the new model with two verb-internal predictors is the better fitting model (ΔAIC=0, weight=1) than the replication of Bresnan et al. (2007) (ΔAIC=25.2, weight<0.001). ^[10] The results of our analysis confirm that the effect of semantic similarity to give, our hypothesized anchor for the ditransitive frame, makes a unique contribution to the prediction of choice of syntactic frame, a contribution that is independent of the effects of other properties reported in previous literature.

5 Testing the Verb Anchor Hypothesis within narrow semantic classes

Until now, we have investigated our Verb Anchor Hypothesis for the entire set of verbs listed in Levin (1993) as alternating between the ditransitive and prepositional frames. In doing so, we treated the verb give as an anchor for verbs that rather loosely share a semantic notion when they occur in the alternate frames. But previous research has shown that when trying to predict which verb can or cannot alternate between two syntactic frames, the “right” level of semantic abstraction may be smaller than the entire class of alternating verbs. Pinker (1989), in particular, has argued that a verb’s ability to alternatively occur in the ditransitive and prepositional frames is conditioned by whether its meaning instantiates a meaning common to a narrow class of verbs. It is within these semantically coherent narrow classes, Pinker argues, that the argument alternation is productive and it is the nature of those narrow classes that children must acquire. Similarly, Goldberg (1995) has argued that the ditransitive construction is associated with closely related meanings (or is polysemous) and that these various meanings arise out of the interaction between the construction and the verbs occurring in it. It has also been argued that some argument structure constructions, e.g. the conative construction (The horse pulled at the cart), cannot be defined at the highest possible level of abstraction but must be defined at relatively lower levels of schematization that are assigned more specific meanings (Perek and Lemmens 2010; Perek 2014). These studies suggest that narrower and more semantically coherent subclasses of alternating verbs play an important role in the mental representation of alternations.

If narrow classes play a role in the representation of alternations, our Verb Anchor Hypothesis should also apply to narrow classes. Verbs within each narrow class share more specific semantic information than verbs from different classes. For example, verbs in the message transfer class (tell, teach, write, etc.) describe events in which an abstract message is metaphorically transferred and possessed by the recipient. Verbs in the future having class (offer, promise, bequeath, etc.) share the semantic property that the transfer-of-possession will be completed in the future. Just as we argued give serves as the ditransitive anchor for the entire broad class of alternating verbs, there may be a highly frequent and typical verb within each narrow class that is strongly associated with either frame of the dative alternation. This possibility is particularly likely if the ditransitive frame is associated with a slightly different meaning or sense across narrow classes, as Goldberg (1995) argues. To draw an analogy from natural categories, there can be a typical exemplar of the bird category and another for the subordinate eagle category. We call such verbs, if they exist, narrow class verb anchors.

Our hypothesis is that, as was the case for give and the entire class of alternating verbs, the degree of semantic similarity to a narrow class anchor affects which syntactic frame other narrow class members occur in. However, putative narrow classes anchors share many semantic features with the broad class anchor give (after all, this is why they alternate just like give). A correlation analysis confirms the substantial semantic overlap between broad and narrow class anchors: a verb’s similarity to give is highly correlated with its similarity to tell (r=0.94, p<0.001) and also with its similarity to leave (r=0.97, p<0.001). As we are concerned with whether the portion of semantic similarity to narrow class anchors that is not at the same time semantic similarity to broad class anchors plays a role in speakers’ choice of syntactic frame, we must exclude from our estimate of semantic similarity the portion that is due to features verbs share with both narrow class anchors and give. We therefore assigned the variance shared by the two correlated predictors to the broad class anchor, partialed out from individual verbs’ semantic similarity to narrow class anchors the portion that is correlated with their semantic similarity to the broad class anchor, and used this residual semantic similarity to narrow class anchors as predictors, following the logic of Baayen et al. (2006).

We test our hypothesis on three narrow classes, the give class, the message transfer class, and the future having class. Note that verbs in these three narrow classes are all give-type verbs that invariably entail caused possession in both the ditransitive and prepositional frames (Rappaport Hovav and Levin 2008). Thus, for these verbs, speakers’ choice of syntactic frames is truly independent of any substantial semantic differences between the two constructions (although there still may some subtle construal differences in some cases).

The first step in testing the Verb Anchor Hypothesis on narrow verb classes is to select a narrow class anchor for each of the three narrow classes. We examined the frequency distribution of verbs within each narrow class, where membership in the class was based on Levin (1993). ^[11] The distributions are illustrated in Figure 2.

Figure 2:

Verb frequencies in three narrow classes.

Within the give class, as expected, give has the highest strength of association with the ditransitive frame and was selected as anchor. For this narrow class, we tested how a verb’s semantic similarity to give predicts the choice of ditransitive frame for a smaller set of verbs than in our previous models. In the message transfer class, the verb tell has the highest strength of association with the ditransitive frame and was selected as anchor. In the future having class, none of the frequent verbs favors the ditransitive frame and most are highly biased towards the prepositional frame. It is interesting to speculate on why verbs in the future having class favor the prepositional frame, particularly since they entail caused possession in both frames. Our tentative answer is that the “goal” meaning associated with the preposition to may better match the future component of the verbs’ meaning, i.e. the fact that the transfer-of-possession occurs in the future. Be that as it may, we chose as anchor the verb with the highest strength of association with the prepositional frame, namely the verb leave (see the Supplementary Material). Our prediction, here, is that high semantic similarity to leave increases the likelihood of a verb occurring in the prepositional frame.

Having selected anchor verbs, we fitted logistic regression models to each of the three narrow classes. We sorted the entire set of sentences extracted from the BNC into three separate datasets, one for each set of narrow class verbs. We then measured semantic similarity between narrow class members and their respective narrow class anchors using the same procedure as before (see Section 3 for details), i.e. give vs. give verbs, tell vs. message transfer verbs, and leave vs. future having verbs. In order to focus on the effect of the semantic similarity to narrow class anchors that is not due to features they share with give, we assigned shared variance to the semantic similarity to give predictor and added to the predictors discussed in previous sections semantic similarity to narrow class anchors after their residualization against semantic similarity to give. We coded each sentence with these three factors in our BNC-based dataset. We used our BNC-based dataset for narrow class analyses due to the small size of Bresnan et al.’s (2007). Sub-datasets sorted from Bresnan et al.’s (2007) for each narrow class contained only a small number of sentences and verbs. Given the size of the BNC-based dataset, we could not perform a manual annotation of all verb-external factors Bresnan et al. (2007) considered. But the three verb-external factors are known to be highly correlated with factors that require manual annotations. For example, the pronoun him in Jean gave him a book is definite, animate, constitutes given information, and is short. We therefore do not expect our omission of other verb-external factors to have much of an effect on our results. In all models we report below, the three verb-external predictors all turned out to be significant predictors, as predicted.

The goal of our analyses is to determine whether residual semantic similarity to narrow class anchors affects the choice of syntactic frame independently of the contribution of those other predictors. We fitted only one model for the give class, however, as the verb give is both the narrow and broad class anchor. But for the message transfer and future having classes, we fitted two logistic regression models to the data, namely, one with only the similarity to the broad-class anchor, and the other with the residual semantic similarity to the narrow class anchor as an additional predictor. We performed standard model selection procedures. The first and simplest model determines whether the broad class anchor give plays a significant role within a particular narrow class; the second can test whether a narrow class anchor has a significant effect in addition to the effect of the broad class anchor.

The give class data set consists of 5,731 sentences occurring with 13 distinct (pay, sell, hand, lend, feed, serve, lease, repay, loan, rent, refund, peddle, and trade). As before, sentences that contain the anchor verb give were excluded from the analysis. For this class, we can only test whether previous findings can be replicated within this smaller set of verbs as the narrow class anchor is the same as the anchor for the entire set of alternating verbs. This model showed that all four predictors make independent contributions to predicting the choice of syntactic frame, i.e. verb similarity to give (b=0.26, z=6.02, p<0.001). Not too surprisingly, and in line with our Narrow Verb Anchor Hypothesis, the results show that give serves as anchor to the narrow give class just as it did for the broad class of alternating verbs.

The data set for the message transfer class includes 2,960 sentences occurring with 9 non-tell verbs (show, ask, write, teach, read, pose, quote, preach, and cite). The simplest model showed similarity to the broad class anchor give is a significant predictor of syntactic choice for message transfer verbs (b=1.10, z=15.01, p<0.001). When both similarity to give and residual similarity to the narrow class anchor tell were entered in the model, the two similarity predictors both turned out to be significant and, notably, both in the positive direction as predicted (i.e. high similarity predicts the ditransitive frame, b=1.12, z=15.01, p<0.001, std error b=0.07 and b=0.12, z=2.04, p<0.05, std error b=0.06, respectively). The overall fit or predictive power of the model also increases when semantic similarity to the narrow class anchor tell is added as an additional predictor (ΔAIC=0.0, weight=0.75, compared to ΔAIC=2.2, weight=0.25), suggesting that residual semantic similarity to tell explains some portion of the variance that similarity to give is unable to explain. The model with both similarity predictors is also a better model than a model that only includes similarity to tell (ΔAIC=0.0, weight=1, compared to ΔAIC=18.2, weight<0.001).

The future having class data set consists of 5,149 sentences occurring with 17 different verbs, excluding sentences with the narrow class anchor leave (offer, owe, extend, grant, assign, award, allocate, issue, promise, guarantee, advance, concede, yield, bequeath, cede, allot, and will). The first model confirmed that high similarity to give increases the choice of the ditransitive frame for the set of future having verbs (b=0.60, z=11.01, p<0.05). When both semantic similarity to give and residual semantic similarity to the narrow class anchor leave were entered as predictors, they were both significant but in opposite directions, as predicted (b=0.54, z=9.55, p<0.001, std error b=0.06 and b=−0.28, z=−4.78, p<0.001, std error b=0.06, respectively). Model comparison shows that the model with both semantic similarity predictors (ΔAIC=0, weight=1) fits the data better than a model that only includes similarity to give as a semantic predictor (ΔAIC=20.8, weight<0.001) and then a model that only includes semantic similarity to leave as semantic predictor (ΔAIC=48.9, weight<0.001).

In the analyses we just presented, we used residual semantic similarity to narrow class anchors as predictors to account for featural overlap between broad and narrow class anchors, as our narrow class anchor hypothesis focuses on the effect on syntactic frame selection of features verbs of a narrow class share with their respective narrow class anchors but not with give. Wurm and Fisicaro (2014), though, suggest using unresidualized predictors for multiple regression analyses that include strongly correlated predictors. We therefore also tested a model where both similarity to give and unresidualized similarity to the narrow class anchor were predictors. In this model, the effect of individual semantic predictors could not be ascertained due to the high collinearity between the two predictors (for message transfer class, similarity to give, b=0.77, z=4.42, p<0.001, std error b=0.17, VIF=6.7 and similarity to tell, b=0.36, z=2.04, p<0.001, std error b=0.18, VIF=6.3; for future-having class, similarity to give, b=1.75, z=7.05, p<0.001, std error b=0.25, VIF=20.8 and similarity to leave, b=−1.16, z=−4.73, p<0.001, std error b=0.24, VIF=20.9). But since both models with unresidualized semantic similarity to narrow class anchors (ΔAIC=0, weight=1) are better than models with only similarity to give (for message transfer class, ΔAIC=2.2, weight=0.25; for future having class, ΔAIC=20.8, weight<0.001), we can still conclude, as the narrow class anchor hypothesis predicts, that narrow class anchors play a role in syntactic frame selection. ^[12]

To summarize, we showed that give serves as the broad class anchor of the ditransitive frame for each of the three narrow verb classes and that the residual portion of narrow class anchors may assist (e.g. tell) or counteract (e.g. leave) that of the broad class anchor give. The narrow class analyses should be taken with some caution, as the narrow class analyses are based on a relatively small number of verbs and we therefore cannot be sure, at this point, whether our results will generalize. Despite these limitations, our analysis of narrow class anchors suggests that they too play a role in syntactic frame selection independently of or cooperatively with broad class anchors.

6 General discussion

This paper quantitatively investigated one of the causes behind verb structural preferences when overall meaning is kept relatively constant across a pair of syntactic frames. Drawing insight from the categorization literature and the notion of exemplars, we treated syntactic frames as categories and sentences occurring in those frames, more specifically verbs that occur in sentences, as exemplars. We hypothesized that frequent experience with exemplar sentences where a particular verb occurs in a particular syntactic frame may lead to a strong association between the verb and the frame, analogous to the association between a category and its best exemplar (e.g. Medin and Schaffer 1978). Categorization research showed that similarity to the best exemplar is a critical factor for an entity to be considered a member of the category the best exemplar belongs to. We hypothesized that the “best exemplar” verb of a syntactic frame constitutes a lexical anchor for that frame, so that how semantically similar another verb is to the anchor affects speakers’ choice of syntactic frame. Our Verb Anchor Hypothesis claims that high semantic similarity to a syntactic frame’s anchor tends to lead speakers to choose the same syntactic frame.

We tested the Verb Anchor Hypothesis by investigating sentences that include verbs participating in the dative alternation. Since the choice of the ditransitive or prepositional frame is known to be influenced by a variety of factors, not only did we investigate the role of a highly frequent and typical verb in syntactic frame selection, we also made sure any effect of a syntactic frame’s anchor was independent of the effect of other known factors. We conducted a series of logistic regression analyses on ditransitive or prepositional sentences collected from the British National Corpus and those used in Bresnan et al. (2007). Our main predictor was semantic similarity between the ditransitive anchor verb give and other alternating verbs, as estimated by Latent Semantic Analysis. To summarize our results, semantic similarity to the anchor give is a significant predictor of syntactic frame selection and the effect of this predictor is not reducible to that of other factors such as the presence of a caused possession entailment in both frames or pronominality of postverbal arguments. Overall, our results confirm our hypothesis that semantic similarity to the verb most strongly associated with a syntactic frame modulates syntactic frame selection. We also tested our Verb Anchor Hypothesis on semantically more cohesive narrow classes like those discussed in Pinker (1989) and showed that narrow class anchors can also play a role in predicting syntactic frame selection. Of particular note is the fact that narrow class anchors may counteract the effect of broad class anchors: While give pulls future having verbs to the ditransitive frame, leave pulls them in the opposite direction, namely to the prepositional frame.

Many researchers have pointed out that verbs occurring particularly frequently in a certain syntactic frame tend to have a very general meaning (e.g. go, give, and put) that may be quite similar to the putative abstract meaning of the syntactic frame. Pinker (1989: 212), for example, suggests that “[give’s semantic] representations are virtually identical to the double-object thematic core” and Goldberg (1997: 386) claims that ”[give] codes an elaboration of the meaning of the [ditransitive] construction.” One may therefore argue that the results of the present study are due to the similarity between the meaning of give and that of the ditransitive frame rather than their high co-occurrence frequency. Since give occurs both highly frequently in the ditransitive frame and its meaning is indeed quite general and not much more than the notion of transfer of possession many assign to the ditransitive frame, teasing apart these two hypotheses is hard. The results of our analysis of the narrow class anchors, though, suggest that the mechanisms underlying the effect of verb anchors amount to more than the semantic generality of verbs like give and provide support for the effect of frequency of occurrence of verbs in particular frames. Namely, leave pull class members towards the prepositional frame, not towards the ditransitive frame as give does. ^[13] Since it is not the case that “the meaning of the prepositional frame” is roughly the same as that of leave, verb anchoring does not seem to reduce, in this case, to semantic similarity between the meaning of the anchors and the meaning of the frame or to the fact that verbs similar in meaning to the anchors would be verbs whose meanings are most compatible with that of the frame. Rather, we surmise, the effect of verb anchors arises from lexical similarity between alternating verbs and the verb most strongly associated with a syntactic frame because of the latter’s high frequency of occurrence in that frame.

The results of our narrow class models also suggest that our focus on a single anchor for the ditransitive frame might be an oversimplification of the role of semantic similarity in sentence production. Verbs overlap in meaning with many verbs, not just the verb that occurs most frequently in a syntactic frame: The verb that best expresses the situation category that a speaker wishes to communicate activates many other verbs. Each of these verbs activates the syntactic frames they occur in. The way in which semantic overlap increases the activation of a syntactic frame is thus probably the result of complex interaction between the weighted activation of other verb meanings and the weighted association of each of these verbs with the relevant syntactic frames (see Chang et al. 2006 for an approach to syntactic frame selection broadly compatible with the view we are articulating). Talking of a frame’s anchor as we have done throughout this paper is nonetheless justified, we believe, for two reasons. It is justified theoretically by previous claims about the importance of verbs that occur most frequently in syntactic frames in the acquisition (e.g. Goldberg et al. 2004; Ellis et al. 2014) or representation (e.g. Pinker 1989; Goldberg 1995) of argument structure constructions. Second, and more practically, the effect of most verbs other than anchors may be too small to be measurable, as the strength of their association with a syntactic frame may be quite weak. Starting the investigation of the influence of semantic similarity on syntactic frame selection with most frequently occurring verbs is, we believe, the wisest course of action.

The claim that the activation of a syntactic frame results from the activation of all verbs that overlap in meaning with the verb that best expresses the speaker’s message resembles the view put forth in Ambridge et al. (2014) that the meaning associated with the ditransitive construction incorporates the meaning of all the verbs that occur in the construction weighted by their frequency of occurrence in the construction (p.238; see also Ellis et al. 2014: 58 for a similar view). There are some important differences, however, between these claims and the view we just articulated: Both the behavior Ambridge et al. (2014) are trying to explain and the predictors of that behavior differ from the focus of our study. Ambridge et al.’s models are meant to study the effect of preemption on argument structure learning and explain the relative acceptability on a five-point scale of sentences that contain alternating and non-alternating verbs. Our study is meant to explain speakers’ choice of syntactic frame in naturally occurring English sentences. Not only is the behavior to be explained different, so are the predictors. In particular, our semantic predictor of interest is the semantic similarity between verb meanings whereas Ambridge et al.’s semantic predictors of interest are the semantic features Pinker (1989) thought important for the dative alternation. Also, our model, as mentioned in the introduction, does not assume that syntactic frames have meanings. Nor does it rely on a theory of meaning that probabilistically include features from the meaning of all verbs that occur in the syntactic frame. Our hypothesis relies on the more conservative view that lexical semantic overlap is what leads to the activation of other concepts, an assumption critical to explaining semantic priming effects (McRae and Boisvert 1986). Despite those differences, however, our two main claims are consonant with what we take to be Ambridge et al.’s main intuition: (1) Each verb meaning is associated with a syntactic frame as a monotonically increasing function of their frequency of occurrence in that frame, (2) the overlap in meaning between the situation category a speaker wishes to express and the meaning of other verbs occurring in that frame mediates the effect those associations have on the choice of syntactic frame.

Another issue we would like to tackle is whether the effect of anchor verbs on syntactic frame selection is a peculiarity of the dative alternation or can be generalized to other syntactic frames. As one reviewer pointed out, other verbs have been argued to be typical exemplars of argument structure constructions, e.g. put for the “caused-motion” construction. The Verb Anchor Hypothesis predicts that such verbs should attract semantically similar verbs to occur in the construction they are typical exemplars of. The results of a model we ran on the 45 verbs listed in Levin (1993) as participating in the locative alternation suggest that this prediction is borne out. After retrieving from the BNC all locative-PP (on[to], in[to], around, etc.) and with-PP tokens of the 45 alternating verbs, we ran a logistic regression similar to the first model we presented in Section 3 (e.g. John spayed the powder on the wall vs. John sprayed the wall with the powder). Semantic similarity to put was indeed a significant predictor of the occurrence of verbs in the locative-PP variant (b=−1.658; p<0.001). Despite this result, whether the Verb Anchor Hypothesis applies to most alternations or just a few depends on several factors whose effect we cannot at present ascertain and is thus a matter for further study. As Sun and Koenig (2017) and Sun (2018) point out, the verb frequency distribution observed in the dative alternation is rather unique among English verb alternations. It is the only alternation (or construction) where a single verb, give, accounts for such a large proportion of the syntactic frame’s tokens. Give might be uniquely strongly associated with the ditransitive frame and the weaker association of putative anchors to other frames might decrease the likelihood of finding an effect of semantic similarity to the anchor verb’s meaning.