The English dative alternation in vector space: how much does meaning matter?

2.1 Step 1: Prepare a dataset about the alternation under study, where observations are annotated for various contextual characteristics and what we will call “traditional” predictors, such as, e.g., constituent length

This paper reanalyzes the US-American section of the dative alternation dataset compiled by Bresnan and colleagues (2017). Dative observations were extracted from the Switchboard corpus of American English (Godfrey et al. 1992). The Switchboard corpus covers telephone conversations collected between 1990 and 1991. The variable context is defined as follows:

Attention was restricted to the dative verb give. The definition of interchangeable ditransitive and prepositional dative variants broadly follows Bresnan et al. (2007), which essentially means that all instances of give with two argument NPs minus non-interchangeable constructions were considered (Szmrecsanyi et al. 2017: 6).

The dataset has annotation for the following language-internal traditional predictors:

1. Recipient/Theme.type: pronominal (personal pronouns, demonstrative pronouns, impersonal pronouns) versus nonpronominal (noun phrase)

2. Recipient/Theme.definiteness: definite (definite, definite proper noun) versus indefinite

3. Recipient/Theme.animacy: animate (human and animal) versus inanimate (collective, temporal, locative, inanimate)

4. Semantics of the dative verb: (i) transfer; (ii) communication; (iii) abstract

5. Length difference: log10 difference of the length of the recipient and theme phrases in orthographically transcribed words (Bresnan and Ford 2010): log(Recipient.length) – log(Theme.length)

The dataset covers a total of N = 1,190 observations. After exclusion of observations with missing data, we end up with N = 1,164 observations, of which 153 (13.1 %) are prepositional dative constructions and 1,011 (86.9 %) are ditransitive dative constructions.

2.2 Step 2: From the dataset, extract information about the lexical material in the argument slots of variant constructions, called target words in vector space modeling. For the dative variants, we extract the theme and recipient constituent head lemmas, i.e., the single most important content word of the constituent slot

For each individual dative observation in the dataset, we consider the theme and recipient constituent head lemmas, which are preannotated in the dataset under analysis (columns Recipient.head and Theme.head). Take, for instance, example (2b) (the judge will usually give custody to the mother): in this example, the theme head lemma is custody, and the recipient head lemma is mother. We pay attention only to the head lemmas, and not to the entire constituents, because these are going to be the so-called target words in semantic vector space modeling. Note that we also include pronominal head lemmas in our analysis, as in (1a) (we’ll give him fifteen) – as Bresnan et al. (2007: 89) point out, pronominal recipients are extremely frequent in spoken English, and excluding those would be unrealistic.

2.3 Step 3: Choose a training corpus to train the distributional semantic models

Vector space models require balanced and large training corpora to generate solid and reliable distributional representations (Lenci and Sahlgren 2023). For this reason, we will use the Corpus of Contemporary American English (COCA; Davies 2019) as a training corpus. COCA is widely used, also in distributional semantics (Lenci and Sahlgren 2023:32). We specifically restrict attention to the spoken section of COCA (approx. 127 million words), given that our dative dataset likewise covers spoken English.

2.4 Step 4: Check if the target words also occur in the training corpus wordlist

Steps 4 and 5 identify the target words to be used in semantic vector space modeling. Importantly, each of the head lemmas in the dative should be attested at least once. If a head lemma does not appear in COCA, the source dative observation is excluded from the analysis. Accordingly, we had to exclude 23 dative observations, resulting in a new total of N = 1,164 valid and modellable dative observations (down from 1,187), of which 153 (13.1 %) are prepositional dative constructions and 1,011 (86.9 %) are ditransitive dative constructions.^[4] The target words thus obtained from the Bresnan et al. (2017) dataset (recipient and theme head lemmas combined) were tagged using the CLAWS7 PoS tag set (Garside and Smith 1997), the same PoS tag set used for COCA.

2.5 Step 5: Run a quality control check on the final target words list

To fix issues with PoS tagging errors, we hand-checked all the PoS-tagged lemmas. We removed from the wordlist the incorrectly tagged entries and corrected PoS tags where needed (for example, the lemma red can be both a proper noun (tagged as pn) and an adjective (tagged as adj)). The final target word list counts 590 PoS tagged head lemmas, of which 129 are recipient head lemmas, and 461 are theme head lemmas. It is important to note that these numbers reflect the high token frequency of pronouns and anaphorical markers in the recipient slot, in contrast to the more diversified lexical space in the theme slot.

2.6 Step 6: Define parameter settings for the distributional models

We are now getting ready for the actual distributional semantic vector space modeling. For the present analysis, we chose to employ a count-based type-level vector model, built with the python package Nephosem (QLVL 2021). This model allows us to build vectors as an aggregation over all the attestations of a given lemma in a given corpus, taking the form of an overall numerical profile of the lemma. The type-level vectors thus obtained abstract away from individual occurrences (i.e., tokens) that realize the (potentially) polysemous meanings of a lexeme, encoding instead “the patterns within the type-level matrices that are indicative of different senses” of the lemma (Heylen et al. 2015: 157).

We begin by fine-tuning the ability of the model to define and represent different configurations of the distributional context space by using different implementations of parameter settings. As demonstrated by Montes (2021), this is a critical process that severely affects the final output, as parameters are heavily corpus- and task-dependent. After several rounds of assessment and testing, we selected a set of parameters as a function of the specificities of our spoken dataset (see Lenci 2018 for a review). The combination of the different parameters listed below yields a total of 432 distributional models to be considered:

1. Window-size parameters, namely the selection of right and left context size: 3, 4, 7 tokens (applied in step 8)

2. Collocation matrix parameters, that is, the dimensionality of the type-level vector: 5,000, 10,000, 50,000 context words for each target word (applied in step 8)

3. Lexical context filters: no filtering, only nouns context (applied in step 8)

4. Association strength measures: Positive Pointwise Mutual Information (PPMI), and log likelihood (applied in step 9)

5. Similarity measures: cosine (applied in step 10)

6. Amount of reduction: none (no dimensionality reduction), reduction to 2 dimensions, reduction to 10 dimensions (applied in step 10)

7. Clustering: 3, 5, 8, 15 partitions (applied in step 11)

2.7 Step 7: Selecting the context words that will be used in subsequent steps, on the basis of a number of filtering criteria, most importantly their overall frequency in the training corpus

To start up the distributional semantics process, the first thing to be computed is a frequency list of all the lemmas in COCA, the training corpus. The lemmas contained in the list are the so-called context words, constituting the distributional lexical context from which our target lemma vectors will be modeled. Using the vocab class provided by the Nephosem package, the context words are extracted from COCA and stored in a python data structure called dictionary, where each lemma is paired with its frequency.^[5] In this way, the distributional context can be filtered by PoS tags. Specifically, we employed both a full (excluding punctuation and numbers) and a noun-only context words list for the sake of assessing if a nominal, hence less sparse, distributional context could help building more representative target word vectors.

2.8 Step 8: Using the target words list and the filtered context words lists, create co-occurrence matrices, in which rows correspond to the target words and the columns to the context words

At this point, we are ready to compile the large co-occurrence matrices at the core of our VSM-semantic predictors, as a function of the parameters defined in Step 6. These matrices are based on the target word lists and context word lists, which we previously computed. Each row represents a target word from the theme or recipient slot of each dative observation in the dataset, while each column represents a context word from the COCA corpus. Consider example (2b): this observation yields two rows with the target words custody and mother. The aggregation of the absolute co-occurrence frequencies between each target word and all the context words from COCA, a long array of numbers, constitutes a word-type vector, a geometrical representation of a lemma in the distributional semantic space. In order to obtain diversified and comparable distributional representations, we used a range of different co-occurrence configurations and so built a total of 18^[6] raw co-occurrence matrices as a function of the different parameters defined in Step 6 (context and window-size filters and vector dimensionalities).

2.9 Step 9: Weighting the raw co-occurrence matrices using association strength measures

Absolute co-occurrence measures (or raw frequencies) are, however, not a solid ground on which to build meaningful semantic representations: following Zipf’s law on skewed lexical frequency distributions, target words would only co-occur with a small number of context words, resulting in sparse, uninformative vectors. A customary practice in distributional semantics, therefore, is to turn raw co-occurrence values into significance weights using association strength measures. This lessens the effect of skewed frequency distributions and allows us to bring to the fore more informative semantic relationships between lemmas. Table 1 illustrates this based on a selection of recipient target lemmas from examples (1) and (2). The table shows the weighted co-occurrence values using the popular, reinforced version of Pointwise Mutual Information (PMI), called Positive PMI (PPMI; Bullinaria and Levy 2007): while maintaining the original goal of PMI, that is, to compare the probability of a target-context word pair to be distributionally closer with the expected probability if the two lemmas were independent, the PPMI relies on zero values instead of PMI negative values to better manage inaccuracies (Jurafsky and Martin 2023: 109).^[7] From the previous 18 raw co-occurrence matrices, we obtained 36 matrices by applying the two parameter settings on each raw co-occurrence matrix.

2.10 Step 10: Compute similarity matrices to represent the distributional similarity between the target words

From the matrices generated in step 9, we next compute similarity matrices to represent the distributional similarity between target words, using the classic pairwise similarity measure cosine. The more the two target words are similar, the closer the value will be to 1; the more the two target words are different, the more the cosine value will be close to zero. In Table 2, we can observe how guy and him are closer in vector space because of some context words shared in the co-occurrence matrix (Table 1), such as dad and it. We would expect, then, to find government in a different part of the vector space.

However, to assess if our data are meaningful and not skewed, we compared cosine-based similarity matrices to dimensionality-reduction-based similarity matrices. Nonlinear dimensionality reduction is a technique applied to reduce sparse, high-dimensional vector spaces into low-dimensional spaces that encapsulate the meaningful essence of the data. There are many ways to reduce dimensionality. We applied Uniform Manifold Approximation and Projection (UMAP; McInnes et al. 2020). UMAP essentially organizes high-dimensional data into a graph to find similarity patterns and then reorganizes these patterns in low-dimensional space using Riemannian geometry. Concretely, UMAP (a) first calculates how similar the target lemmas are in the original high-dimensional space using some distance measure (in this study, cosine) and then (b) reduces the data to a lower-dimensional space (e.g., 2 or 10 dimensions), where relationships between words are modeled using the UMAP-standard Euclidean distance measure.

In our case study, Step 10 yields 108 similarity matrices for each of the two dative construction slots, so 108 × 2 = 216 similarity matrices in total.^[8]

2.11 Step 11: On the basis of the similarity matrices, cluster the target words with separate clusters for theme and recipient

On the basis of the recipient and themes similarity matrices generated in Step 10, theme and recipient heads are now – always separately – going to be clustered. This process yields groupings of semantically related types which will serve as semantic predictors of dative variant choice. Specifically, using the Partition Around Medoids (PAM) clustering algorithm (Kaufman and Rousseeuw 1990) in its Python implementation provided by the scikit-learn-extra module for machine learning from the package scikit-learn (Pedregosa et al. 2011), the central members of the clusters, called medoids, are identified in the data. Then, the rest of the type-word vectors, i.e., the distributional representations of our dative head lemmas, are grouped around these medoids based on the cosine similarity measure. The partition values (i.e., the number of clusters to create) are set to 3, 5, 8, 15: for each of the 108 similarity matrices, we obtain four clustering models, for a final total of 108 × 4 = 432 clustering models per theme and recipient slot. We can now let the distributional machine rest and focus on the evaluation of the models and their predictors.

2.12 Step 12: Evaluate the vector space models obtained and their clusters, and choose the optimal model according to principled criteria

After obtaining the clustering models, the next step consists of selecting the model that clusters best. In other word, we need to choose the VSM-semantic predictors that best capture generalizations about semantic similarities between the recipients/themes in our dataset. To this end, the clustering models are evaluated by computing (i) the Concordance-C index (C-value)^[9] of their Conditional Random Forest (CRF), a multivariate statistical method that we utilize to predict dative variant choice based on only VSM-semantic predictors, and (ii) the negative token silhouette (Rousseeuw 1987) of the dative lemmas, i.e., the percentage of tokens of each dative lemma assigned to the “wrong” (or suboptimal) clusters, to gain a better understanding of data dispersion during clustering.

We considered only models with a negative token silhouette of less than 25 % in both recipient and theme clusters, and a CRF C-value higher than 0.75. After identifying the model with the lowest negative token silhouette and the highest CRF C-value,^[10] we also inspected the clustering consistency of the next best fifteen models to rule out possible biases. As the clusters of these next-best models share similar semantic patterns, we proceeded to pick the best model. This was a model with a window size of 4, without lexical context filtering, with a vector dimensionality of 5,000, using PPMI as association strength measure, and with a dimensionality reduction of 10 and 15 clusters for both dative slots (see above for details). Everything is now ready for using these clusters as VSM-semantic predictors of the dative alternation.

2.13 Step 13: Visualize the VSM-semantic predictors using t-SNE for a qualitative assessment

The last two steps in this workflow are dedicated to integrating VSM-semantic predictors into customary variationist alternation analysis. First, our methodology includes a module that qualitatively investigates VSM-semantic predictors using visualization tools. Inspired by the visualizations of the distributional clouds of tokens in Montes (2021), we map out VSM-semantic predictors using the stochastic algorithm t-SNE (Maaten and Geoffrey 2008) in its R implementation Rtsne (Krijthe 2015). From the same family of dimensionality reduction algorithms as UMAP, t-SNE is widely used to represent high-dimensional data in low-dimensional spaces (i.e., 2- or 3-dimensional spaces), maintaining and enhancing the meaningful features and semantic relations among the data. What we obtained are two plots, one for the VSM-semantic predictors relating to recipients, and the other relating to themes. The plots visually depict the distributional semantic geography of the predictors: how they are interconnected, which semantic patterns they show, and what their internal structure is. Additionally, the plots afford insights on misclassified data points (i.e., lemmas with negative token silhouette). In short, visualization is a tool that allows us to better understand the semantic structure and offers guidance for a responsible interpretation of the quantitative behavior of VSM-semantic predictors in regression analysis (see next step).

2.14 Step 14: Conduct two logistic regression modeling runs, creating models (a) with only traditional predictors as fixed effects and (b) with only VSM-semantic predictors as fixed effects. Determine the effect direction of VSM-semantic predictors and compare the goodness-of-fit of the two models

We utilize mixed effects logistic regression analysis, which is of course widely used in variationist/probabilistic linguistics (Tagliamonte and Baayen 2012). The model with only traditional predictors is structurally similar to the regression model in, e.g., Bresnan et al. (2007). The model with only VSM-semantic predictors uses binary categorical predictors in its fixed-effects structure. The binary structure of VSM-semantic predictors is simple: the recipient/theme lemma is a member of the cluster/predictor in question versus the recipient/theme lemma is not a member of the cluster/predictor in question.^[11] Hence, the regression model assesses whether a specific recipient or theme VSM-semantic predictor is a significant predictor of dative choices and evaluates effect directions (i.e., resulting preferences for the ditransitive or the prepositional variant).

Following best practice in variationist linguistics, we started out with a maximal model including all the predictors as fixed effects, and the language-external factor speaker ^[12] as well as the lexical effects Theme and Recipient heads as random effects (see, e.g., Röthlisberger et al. 2017). We include Theme and Recipient head lemmas as random effects even in the model with only VSM-semantic predictors because we want to tease apart the effect of purely lexical factors (via Theme and Recipient head lemma as random effects) from the effect of VSM-semantic predictors (via cluster memberships in the fixed-effects structure). After a round of model optimization to address multicollinearity and convergence issues, we pruned the models: first, we pooled infrequent random effect categories into “other” categories (thresholds: 5 for speaker and 2 for Theme and Recipient heads). Subsequently, those predictors that lacked explanatory power, namely the ones with the highest p-value, were removed one by one, until we ended up with minimal adequate models.

Data availability statement: the dataset(s) and analysis scripts are available at our OSF repository (https://osf.io/kfqsr/?view_only=71f776905c2645eb9dbb3b9f33aa42c5).

3.1 Clustering theme and recipient lemmas as a function of semantic similarity

The evaluation process carried out in Step 12 (described in Section 2.12) identified the optimal semantic vector space model. As previously stated, an optimal model should be associated with a very low negative token silhouette value for both recipients and themes (≤0.25), and with a conditional random forest C-value > 0.75. The optimal model is an excellent one, coming with a negative token silhouette of only 0.016 % for recipients and 0.06 % for themes, as well as with a C-value of 0.95 for a conditional random forest model predicting dative variant choices. We note that most of the parameter settings in the optimal model and its clustering are actually shared by the top fifteen models, indicating that the clustering output is fairly robust. The optimal model has the following parameters: 4-size context window, no vocabulary filter, dimensionality of 5,000, PPMI as weighting measure, dimensionality reduction to 10 dimensions, and 15 clustering partitions.^[13] This means that the optimal model yields fifteen clusters for each dative slot, for a total of thirty clusters of dative recipients and themes.

Before proceeding to the variationist analysis, we visually depict the clusters in step 13 (described in Section 2.13). The interpretation of the resulting plots is fairly straightforward: individual dots represent distributionally modeled recipient/theme lemmas, with color coding indicating cluster membership. Each cluster is labeled by its central medoid. The extra labels in black with the arrows locate the lemmas in examples (1) and (2).

Figure 1 plots the recipient lemmas. PAM clustering yields all in all 15 internally reasonably homogeneous clusters, which are listed in Table 3.

Figure 1:

Scatterplot of recipient clusters. Color coding indicates cluster memberships. Clusters are labeled by their central medoid. Items in white rectangles with arrows refer to examples (1) (we’ll give him fifteen, but they give the guy a job) and (2) (I gave it to the government, the judge will usually give custody to the mother).

Figure 2 plots the theme lemmas in our dataset. Again, PAM clustering yields in all 15 internally reasonably homogeneous clusters, as described in Table 4.

Figure 2:

Scatterplot of theme clusters. Color coding indicates cluster memberships. Clusters are labeled by their central medoid. The item in the white rectangle with the arrow refers to the pronoun it, a key lemma in the dative alternation dataset under analysis.

3.2 Variationist modeling: to what extent does distributional-semantic information predict dative choices?

In Step 14 (described in Section 2.14), we set a baseline by fitting a binary logistic regression model (Table 5) with only “traditional” predictors (i.e., those that old-school dative alternation research usually considers – recall that these come preannotated in the dataset). The response variable is dative variant choice (predicted odds are for the ditransitive dative). After model trimming (Zuu et al. 2009: 121–122), the minimal adequate model has a C score of 0.990; the condition number kappa^[14] is 6.541. Suffice it to say that the traditional predictors all have the expected effect direction, given the rich literature on the dative alternation (see, e.g., Bresnan et al. 2007; Szmrecsanyi et al. 2017 for models based on essentially the same dataset). We would like to emphasize that a C score of 0.990 indicates outstanding discrimination (Levshina 2015: 259) and underscores how well-understood the dative alternation is in principle.

A second binary logistic regression model was then fitted (Table 6), this time with only VSM-semantic predictors (i.e., those that we distributionally modeled). These predictors are basically binary dummy variables indicating if in a given dative observation, the recipient and theme head lemmas are or are not members of distributional-semantic clusters as discussed in Section 3.1. As in the “traditional” model, the response variable is dative variant choice (predicted odds are for the ditransitive dative). The maximal model contained 30 predictors (15 recipient clusters and 15 theme clusters, as discussed in Section 3.1). After model trimming, the minimal adequate contains four predictors, as shown in Table 6. The minimal adequate model has a C score of 0.991, and its condition number kappa is 6.804.

The model in Table 6 may be interpreted as follows. Predicted odds are for the ditransitive dative variant. The model, being minimally adequate, only shows significant VSM-semantic effects. If the recipient is in the advance cluster, which mostly contains lemmas about geopolitics (see previous subsection for discussion and exemplification), the odds for the ditransitive dative variant decrease by a factor of 0.142, i.e., by approximately 85 %. If the recipient is in the anything cluster (which mostly contains pronouns), the odds for the ditransitive dative increase by a factor of almost 7. If the recipient is in the myself cluster (which also mostly contains pronouns), the odds for the ditransitive dative increase substantially by a factor of almost 129. The anything and myself effects overlap with traditional modeling in that these clusters are highly pronominal in nature, and recipient pronominality has been shown to be a strong predictor of ditransitive dative usage in traditional dative alternation research (see, e.g., Bresnan et al. 2007). We also find one significant theme predictor: if the theme is in the one cluster (lemmas relating to general referents such as the pronouns something, anything and, but also nouns such as item and thing), the odds for the ditransitive dative decrease by a factor of 0.097. Figure 3 visually depicts the effect sizes of the predictors in the VSM-semantic model in Table 6.^[15]

Figure 3:

Odds ratios in the VSM-semantic model reported in Table 4.

Now that we have the traditional model (Table 5) and the innovative VSM-semantic model (Table 6), it is time to revisit the question of which model has more explanatory power. One clear way of assessment is grounded on the difference between full effects (fixed and random) and fixed effects only that are inherent in the two models. Table 7 presents the C-values for both kinds of effects in the two models.

Both models have the same random effect structure, which includes lexical lemma effects (as intercept adjustments) for both the recipient and theme. It has become an industry standard to include such lexical effects (so-called by-item effects) in regression models in alternation research (see, e.g., Röthlisberger et al. 2017 and many other recent studies on the dative alternation and beyond). We also include lexical effects in the VSM-semantic model because this allows us to better tease apart the power of semantic effects from the power of purely lexical collocation/collostructional effects (à la Gries and Stefanowitsch 2004), and we are primarily interested in the power of semantics. The point is that particular lexemes, such as the recipient mother in (2b), may favor particular dative variants, but the question that we are primarily interested in is whether such effects can be generalized to lexemes with similar meanings. According to our analysis, the lexical effect of mother is not generalizable because the cluster with the medoid mother is not selected as significant in regression analysis.

Table 7’s “C-value fixed effects only” scores are calculated using the somers2() function in the Hmisc R package: somers2(), in combination with the predict() function. This row essentially tells us how explanatory the fixed effects (traditional or VSM-semantic) are, counting out purely lexical effects.^[16] What Table 7 shows, then, is that the goodness-of-fit of both the traditional model and the VSM-semantic models is practically identical, scoring outstandingly high C-values (0.990 vs. 0.991). However, the two models work differently “under the hood” regarding the division of labor between fixed effects and random effects (including, crucially, purely lexical effects). In the traditional model, the fixed effects do the lion’s share of the explanatory work (0.957 out of 0.990). In the VSM-semantic model, the fixed effects (i.e., VSM-semantic effects) also do a lot of explanatory work (0.782 out of 0.991), but comparatively more variability is left over to be accounted for by the random effects structure (including lexical random effects) via intercept adjustments. In other words: traditional predictors do more heavy explanatory lifting than VSM-semantic predictors. The latter leave more unaccounted variability for the random effects structure to deal with.

4.1 Key findings

Vector space modeling involves various parameter settings. We tested a range of parameters and parameter combinations and in so doing generated a total of 432 models. We then assessed the models according to predefined, principled criteria, and eventually selected the model with the lowest negative token silhouette in both recipient and theme clusters and the highest C-value of a conditional random forest model predicting dative choices. This optimal model generates 15 semantic clusters for the recipient slot in dative constructions and 15 semantic clusters for the theme slot (and this is by the way also what the next 19 top-ranked models do). There appears to be a richer variety of meanings in the theme slot, ranging from economics, family, daily-life items, to more abstract referents like feelings and numbers. The recipient slot in dative constructions has a strong preference for animate subjects and pronouns, as well as lexical items related to general referents like matters, thing.

We then moved on to variationist modeling: to what extent can the semantic clusters predict dative choices? To set a baseline, we initially calculated a regression model containing in its fixed-effect structure only traditional predictors, and enriched with by-subject and purely lexical by-item effects in the random effects structure (Table 5), in the spirit of Bresnan et al. (2007). The predictors that were selected as significant were semantics of the verb give, recipient pronominality, recipient definiteness, and theme definiteness, recipient animacy, and end-weight effects. The model thus does contain some semantic information (semantics of the verb give, recipient animacy), which, however, is standardly included in state-of-the-art modeling.

Next, we calculated an alternative regression model (Table 6) including only distributional-semantic predictors in its fixed effects structure while retaining the exact same random effects structure as in the first model. The semantic predictors are specifically dummy variables indicating whether a given recipient or theme lemma is a member of one of the clusters identified in semantic vector space modeling. Four clusters were identified as significant predictors:

1. the advance cluster relating to recipients (which mostly contains lemmas about geopolitics)

2. the anything cluster relating to recipients (which mostly contains pronouns)

3. the myself cluster relating to recipients (which also mostly contains pronouns)

4. the one cluster relating to themes (lemmas relating to general referents such as the pronouns something, anything and, but also nouns such as item and thing)

The anything and myself clusters favor the ditransitive dative, while the advance and one clusters favor the prepositional dative. We note that the anything and myself clusters are strongly pronominal in nature and thus overlap with traditional pronominality effects as modeled in the traditional model. To some extent, pronominality also plays a role in the one cluster. We would like to remind the reader that in spoken English in particular it is hard to find dative constructions without pronominal slots, so excluding pronominal tokens from analysis was not an option.

How do the two variationist models compare to each other? The crucial criterion for our purposes is goodness-of-fit, which we measured by calculating the Index of Concordance C (Levshina 2015: 259). As shown in Table 7, both the traditional model and the alternative, VSM-semantic model, have outstanding and very comparable fits (0.990 vs. 0.991). But in the traditional model, the traditional fixed-effect predictors do considerably more work than the semantic fixed-effects predictors in the VSM-semantic model. In other words, there is more unaccounted variability in the VSM-semantic model, which is then “mopped up” as it were by the random-effects structure (including purely lexical effects). The conclusion is that traditional predictors are more explanatory than VSM-semantic predictors.

4.2 Theoretical implications

We note, first, that the semantic clusters identified by vector space modeling are overall lexically coherent, demonstrating that the method works in principle. But showing that vector space modeling works was not the aim of the paper. Rather, the research question we asked in the introduction section was the following: how adequately can we predict dative choices as a function of the lexical semantics of dative constituents (as opposed to the semantics of dative constructions)?

The upshot is that we can adequately model the dative alternation using information about the lexical semantics of dative constituents combined with by-subject and lexical by-item effects in the random effects structure. A model with this design accomplishes outstanding goodness-of-fit that is equal to goodness-of-fit of traditional models. The problem is that in this alternative model, only 4 out of 30 VSM-semantic clusters are selected as significant predictors of dative variant choice, and outstanding goodness-of-fit is achieved by letting purely lexical random effects do a lot of the work. The semantic predictors per se are less explanatory than traditional higher-order predictors. In a nutshell, the best of in all 432 semantic space models creates predictors that are outperformed by traditional, state-of-the-art predictors.

This verdict does not lend support to the idea that semantics is a strong factor fueling the dative alternation, not even through the detour via the constituent slots. Let us recapitulate: we knew before that dative variants are functionally equivalent and normally interchangeable (Bresnan et al. 2007). We now also know that variant choice is to some extent constrained by the semantics of the materials in the constituent slots, but this conditioning is weaker than conditioning by traditional predictors (note here that unlike some construction grammarians, we do not assume that probabilistic conditioning comes under the remit of the “meaning” of dative variants – see fn. 2). Had our exercise in dative alternation modeling shown that inclusion of VSM-semantic predictors improves dative alternation modeling (e.g., by reducing the share of explanatory work that random effects have to do), we would have been in a position to conclude that the idea of form-function symmetry has (some) empirical back up. However, inclusion of VSM-semantic did not buy us much. Therefore, the idea of form-function symmetry in the dative alternation retains the status of an empirically unconfirmed conjecture.

We also add that from the effect of VSM-semantic predictors in regression modeling, we do not see any indications whatsoever that change-of-state or change-of-possession contexts favor the ditransitive dative variant, or that change-of-place or movement-to-a-goal contexts favor the prepositional dative variant, as they should if advocates of the multiple-meaning approach had a point (see, e.g., Green 1974; Oehrle 1976). With that being said, we acknowledge that in the “traditional” model, we report in Table 5 there is a significant effect such that give-semantics of “transfer” decrease the odds for the ditransitive dative variant and increase the odds for the prepositional dative variant. Thus, one could interpret our results as indicating that, first, there is a semantic effect that bears some resemblance to the multiple-meaning claim, and that, second, the traditional model picks up traces of this effect, but the VSM-semantic-model does not. Hence, one might conclude that (a) there are traces of multiple-meaning effects, but they are less strong than the effects of other factors, and they certainly are not strong enough to fully explain or fully motivate the alternation, and that (b) the VSM-semantic predictors are simply not good at picking up these traces. The extent to which it is easy to predict the meaning of to give, in a particular observation, on the basis of the VSM-semantics of what is in the theme and recipient slots is an empirical question whose exploration is reserved for another occasion.

4.3 Directions for future research

The analysis we conducted is clearly just a first step and ought to be extended to other alternations in the grammar of English or of other languages. By choosing the dative alternation after the verb to give as a test case for the VSM-semantic approach, we sought to establish a gold standard: this alternation is, after all, extremely well understood, and variationist linguists have had decades to fine-tune traditional modeling. The issue is, however, that state-of-the-art traditional modeling of the dative alternation after give achieves outstanding goodness-of-fit scores that are hard to surpass by any model, so the English dative alternation after give is perhaps a particularly tough nut to crack for alternative approaches. But the point is that our methodology paves the way for reanalyzing the dative alternation after verbs other than give, and also for taking a fresh look at (perhaps less well understood) alternations in which semantics potentially plays a larger role. In this context, it would be particularly interesting to reinvestigate grammatical alternations that are less likely (compared to the dative alternation) to have pronominal constituents – pronouns, after all, do not straightforwardly carry semantic meaning, although of course Semantic Vector Space will assign pronouns to particular clusters. Be that as it may, what this paper would seem to have provided is a robust methodology for including distributional-semantic information in variationist modeling.