1 Introduction

1In Construction Grammar (Goldberg, 2006), the basic units of linguistic analysis are called constructions (Cxns), form-meaning pairings associated with autonomous, non-compositional abstract meanings, independently from the lexical items occurring in them. Examples of Cxns range from morphemes (e.g., pre-, -ing), to filled or partially-filled complex words (e.g., daredevil) to idioms (e.g., give the devil his dues) to more abstract patterns like the Ditransitive [Subj V Obj1 Obj2] (e.g., he gave her a book) (Goldberg, 2006).

2Cxns appear at any level of linguistic analysis, but the level at which the notion of constructional meaning represents a radical departure from other theories of grammar is argument structure. These Cxns, such as the English Ditransitive, are claimed to be associated with an abstract semantic content. In this case, constructional meaning can be paraphrased as X CAUSES Y TO RECEIVE Z. One of the main supporting arguments in favour of constructions as independent and primitive objects of grammar is the flexibility with which argument Cxns and verbs interact with each other, as in example (1) in which the original intransitive sense of “to sneeze” is overridden by the Caused Motion Cxn, and thus takes a transitive sense of “making something move by sneezing”.

(1) John sneezed the napkin off the table

3This flexibility in combining Cxns and verbs is known as valency coercion (Michaelis, 2004; Boas, 2011; Lauwers and Willems, 2011; Perek and Hilpert, 2014).

4This phenomenon, although vastly addressed for English, has not yet received a systematic investigation in other languages. For notable exceptions, see Boas and Gonzálvez-García (2014). In particular – to the best of our knowledge – no previous attempt to carry out an empirical investigation of valency coercion exists for Italian. However, even a simple corpus query reveals that the phenomenon is present in Italian, though it is not as pervasive as in English:

(2) Tossì una risata leggera tra i suoi capelli (He coughed a light laugh in her hair) [ItWac]

5This paper presents an analysis of Italian constructional flexibility that combines psycholinguistic and computational evidence: first, we present the results of a behavioral experiment on valency coercion. Then, we model Cxns with distributional semantics to investigate whether the semantic shape of Italian argument Cxns can affect the interpretation and processing of coerced sentences.

2 Studying valency coercion: an acceptability rating task

6MATERIALS AND SUBJECTS: The offline psycholinguistic experiment targets 9 Italian Cxns (see Table 1) that were selected using existing resources: LexIt (Lenci et al., 2012] and ValPal (Cennamo and Fabrizio, 2013). The resultant Cxns are of varying abstractness and schematicity levels (Barðdal, 2008).

7For each Cxn, we built 21 sentences, which were subdivided into 3 experimental conditions: GRAMMATICAL (3a), COERCION (3b), IMPOSSIBLE (3c) (7 sentences per condition). The total number of stimuli amounts to 189 sentences. The structure of the test was inspired by Perek and Hilpert (2014). Between conditions, sentences differ only for their main verb, to have as little variation as possible.

(3) a. Gianni ha detto che verrà domani (Gianni said that he will come tomorrow)
b. Gianni ha fischiettato che verrà domani (Gianni whistled that he will come tomorrow)
c. Gianni ha cucinato che verrà domani (Gianni cooked that he will come tomorrow)

8The coercion condition consists of verbs that display a partial semantic incompatibility with the constructional environment they are embedded in. They were selected by means of both native intuition and corpus query, selecting and refining cases that were either hapax or rare occurrences in the Italian corpus ItWac (Baroniet al.2009).

9120 Italian native speakers were tested: 39 adolescents (12-14 years old), 40 young adults (18-35 years old), and 41 adults (over 40). We tested subjects of different ages following extensive sociolinguistic literature that has shown that language use changes with age (Eckert, 2017; Labov, 2001; Wagner, 2012). Thus, it could be the case that grammaticality judgments on creative, non-standard sentences are also affected by age. Including different age groups in our analysis allows us to investigate a more representative sample of the population. To control for the possible influencing factor of education level, we only tested adult speakers either in possess of (at least) a bachelor degree or enrolled in a University course. Table 2 summarizes number, age groups and distribution of tested subjects.

10A within-subject design was used, in which each subject sees all stimuli. Participants were asked to judge the acceptability of the (randomized) stimuli on a Likert scale from 1 - “completely unnatural” - to 7 - “perfectly natural”. Presentation of the data varied across age groups: adolescents were given the test directly in their class. Young adults’ judgments were collected through the online platform Figure Eight. Older adults, instead, were presented with a simple Microsoft Word document, in order to include participants who did not have familiarity with online data gathering.

11RESULTS: We assessed statistical significance via linear mixed effect modelling, with by-subject and by-item intercepts.¹ Results show that coercion sentences (purple boxplot in Figure 1) are recognized as an intermediate condition between complete grammaticality and total ungrammaticality.² We consider this result to support the claim that coercion effects include a degree of semantic incompatibility that is nonetheless resolved in the interpretation process. Consistently with the main tenets of Construction Grammar, we argue that the resolution of such incompatibility is driven by a dynamic interaction between the main verb and the constructional context (Kemmer 2008; Kemmer and Yoon, 2013; Yoon, 2016). In a second analysis, we wanted to assess the effect of Cxn types on acceptability ratings. We used linear mixed effect modelling, adding an interaction between Cxn type and experimental condition.³ Results indicate high variability in Cxn ‘coercibility’ (see Figure 2 and table 3). That is, some Cxns in our dataset were consistently judged as more natural by speakers in the coercion condition.

Figure 1: distribution of judgments in the 3 conditions

Figure 2: line plot of judgments

12In particular, it appears that IM, VDE and VDI Cxns result to be more natural, while DT, PASS and (marginally) CO are the least naturally perceived ones in coercion sentences. Since coercion effects are said to be resolved by the general Cxn semantics overriding the lexical meaning of the verb, we hypothesize that the different flexibility degrees of the Cxns in the first experiment could be at least partially explained by distributional properties, such as type and token frequency, and semantic density of the Cxns in our dataset, the latter again estimated with distributional semantics.

13Different degrees of flexibility could derive either from cognitive processes that reflect on language use, or emerge from repeated exposure and thus entrench in speakers’ grammar. Both possible directions of this causal circle, however, ultimately allow us to fruitfully investigate construction flexibility using distributional semantics models. In other words, the higher ‘coercibility’ of novel instances of some Cxns could be due to speakers’ sensitivity to distributional semantic features of the constructions (Barddal, 2006; Bybee, 2013; Zeschel, 2012; Perek and Goldberg, 2017).

3 A Distributional Semantic Model for argument constructions

14PROCEDURE: Perek:2016 has shown that distributional semantics [] can be fruitfully used to model the semantic space covered by a Cxn. It has been argued in the literature that constructional meanings for argument Cxns arise from the meaning of high frequency verbs that co-occur with them (Goldberg, 1999; Casenhiser and Goldberg, 2005; and Barak and Goldberg, 2017). Therefore, we modelled the semantic content of Cxns with the semantics of their most typical verb, each represented as a distributional vector.

15We used the UDLex Pipeline⁴ (Rambelli et al., 2017) to obtain a mapping between the Cxns of our dataset and the most frequent verbs that occur in them (these were selected considering verbs that appear at least 5 times in the relevant subcategorization frames). Table 4 summarizes the number of verbs considered for each of the eight Cxns.⁵ Then, we built a Distributional Semantic Model (DSM) from the italian corpus itWac (Baroni et al., 2009) in order to represent verb meaning of the verbs obtained with UDLex. The 300-dimensional vectors (i.e., the embeddings) were created with the SGNS algorithm (Mikolov et al., 2013), using the most frequent 30,000 words as context, with a minimum frequency of 100.

16Following lebanimodelling2017, we represented each Cxn as the weighted centroid vector of its typical verbs, as follows:

where V the set of the top-associated verbs v with Cxn and frel(v, Cxn) is the co-occurrence frequency of a verb in a Cxn.

17We measured the pairwise cosine similarity among the weighted Cxn vectors: as shown in Figure 3, the distributional behaviour of the Cxn vectors suggests that some Cxns in our dataset show similar distributional behaviour.

Figure 3: Construction semantic similarity.

18Following Perek (2016), the semantic density of a Cxn is computed as the mean value of pairwise cosines between the verbs occurring in Cxn. Figure 4 plots the semantic densities of our Cxns.

Figure 4: Construction semantic density.

19Finally, to assess the effect of distributional properties on Cxns flexibility, we used semantic density, type frequency and token frequency (cf. Table 4) as predictors in linear mixed effect modelling. As dependent variable, we used the difference gramm – coer and coer – imp. We performed two separate analyses for type and token frequencies without interactions to avoid multicollinearity effects. Predictors values were centered.

20RESULTS: The estimates are reported in Tables 5 and 6 below. In the first two models frequency does not yield any effect. In the second models, instead, frequency appears to have an effect on the data. Hence, it appears that type and token frequency help discerning impossible from coercion instances of a Cxn, whereas only semantic density affects the higher naturalness of coercion phenomena. The more a Cxn is observed with semantically similar verbs (i.e., verbs that belong to the same classes or subclasses, which therefore increase the Cxn semantic density), the more the constructional meaning is easily coerced into novel instances.

4 Discussion

21These findings support our claim that coercion effects are resolved by a dynamic interrelation between verb and Cxn [, and Yoon2013]. Even though frequency effects are shown to affect Cxns extensibility to new items [], our results suggest that type and token frequency only facilitate the distinction between semantically incompatible and partially compatible formulations, whereas higher coercibility is only affected by semantic density.

22We interpret this finding in light of the upward strengthening hypothesis (Hilpert, 2015), according to which a novel occurrence of a linguistic unit strengthens a superior node (i.e., the abstract Cxn) only if the former is categorized ‘as an instance of a more abstract Cxn. If this categorization is not performed, or only superficially so, no upward strengthening will take place’ (Hilpert, 2015, p.38). Higher coercibility is hence not affected by frequency of the Cxn because of the ‘intermediate’ grammaticality level of coercion, which does not allow unambiguous categorization. Coercion sentences result more natural if the target Cxn is observed with verbs belonging to similar semantic classes or subclasses, which increases Cxn semantic density. We could therefore assume that coercion effects in Italian elicit a partial categorization. The effect of semantic density, however, only explains part of the data. In fact, visual inspection of the relation between semantic density and the estimates of table 3 reveals that this effect does not explain the high coercibility of IM, or the low values of CO Cxns (see Figure 5).

Figure 5: relation semantic density- estimates

23All things considered, semantic properties (modelled with distributional vectors) of Cxns (e.g., its density) are only one of the factors influencing speakers processing and recognition of coercion effects. In fact, it has been argued that Romance languages are more valency driven than English (and Germanic languages in general) (Perek and Hilpert, 2014). The results of both experiments provide substantial evidence for an integrated account of Italian coercion effects, which should consider not only the properties of the general abstract Cxn, but rather the interaction of the mismatching verb with Cxn meaning.

24These result also have interesting implications to understand the cognitive mechanisms underlying Cxn flexibility and productivity. In fact, these findings support the idea that Cxn meaning is abstracted from the semantics of prototypically occurring verbs. As we saw, several studies have argued in favour of this hypothesis for English, but the fact that we were able to adapt it to Italian suggests that the factors driving the acquisition of Cxns are - at least partially - not language-specific but rather general cognitive processes.

Acknowledgments:

25The authors thank Lucia Passaro and Florent Perek for their help and valuable suggestions.