Modelling the Meaning of Argument Constructions with Distributional Semantics

Current computational models of argument constructions typically represent their semantic content with hand-made formal structures. Here we present a distributional model implementing the idea that the meaning of a construction is intimately related to the semantics of its typical verbs. First, we identify the typical verbs occurring with a given syntactic construction and build their distributional vectors. We then calculate the weighted centroid of these vectors in order to derive the distributional signature of a construction. In order to assess the goodness of our approach, we replicated the priming effect described by Johnson and Golberg (2013) as a function of the semantic distance between a construction and its prototypical verbs. Additional support for our view comes from a regression analysis showing that our distributional information can be used to model behavioral data collected with a crowdsourced elicitation experiment. Argument constructions and verb meaning One of the main tenets of Constructionist approaches (Goldberg 2006; Hoffmann and Trousdale 2013) is that grammar contains argument constructions (Cxns) associated with abstract meanings independently of the verbs occurring in them. For instance, the Ditransitive Cxn “Subj V Obj1 Obj2” (She gave him a book) is associated with a semantic content that can be paraphrased as X CAUSES Y TO RECEIVE Z. This represents a major breaking point between Constructionist models and so-called “projectionist” approaches stemming from the formal generative tradition (Chomsky 2000) and arguing for the the pivotal role played by the main verb in the interpretation of the sentence. According to this latter view, grammar includes only formal combinatorial principles, while the semantic properties of a sentence are determined by the syntactic and semantic properties projected from the main verb. For instance, the syntactic configurations in the sentences in (1) are projections of the main verb to slice: (1) a. He sliced the bread. b. Pat sliced the carrots into the salad. c. Pat sliced Chris a piece of pie. d. Pat sliced the box open. (examples from Goldberg (2006)) Copyright c © 2017, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. This view has been criticized in the Constructionist tradition, following theoretical considerations as well as psycholinguistics arguments (Goldberg 1995; 2006). From the theoretical side, it has been pointed out that, if argument structure is projected solely by the meaning of the verb, than it is necessary to stipulate a different meaning for each occurrence of a given verb in various argument structures. As for the sentences in (1), this view would lead to postulate special senses of this verb roughly meaning (1a) to cut something with a sharp instrument; (1b) to cut something with a sharp instrument so as to move it; (1c) to cut something for someone else with a sharp instrument; (1d) to cut something with a sharp instrument so as to change its state. In a Constructionist perspective, the verb “to slice” is always used with the intuitive meaning of to cut with a sharp instrument, the additional meaning coming from the construction in which it occurs, whose semantics can be paraphrased as (1a) something acting on something else; (1b) something causing something else to move; (1c) someone intending to cause someone to receive something; (1d) someone causing something to change state (Goldberg 1995). Psycholinguistic evidence, on the other side, mostly originates from research on language comprehension and language acquisition. As for the former, studies like Bencini and Goldberg (2000), Kaschak and Glenberg (2000), Kako (2006), Goldwater and Markman (2009) and Johnson and Goldberg (2013) support the idea that the construction of a sentence (rather than the verb only) plays a role in its interpretation. In a sorting experiment, Bencini and Goldberg (2000) showed that, when asked to sort sentences on the basis of their overall meaning, subjects were as likely to rely on the verb as on the construction. Kaschak and Glenberg (2000) and Goldwater and Markman (2009) tapped into the semantic content of different syntactic frames by using novel denominal verbs in a comprehension task. Likewise, Kako (2006) investigated the meaning of six syntactic frames by collecting linguistic judgments over phrases whose content words were replaced by nonsense words (a.k.a. “Jabberwocky”sentences like The grack mecked the zarg). While all these works exploited off-line tasks or explicit judgments, Johnson and Goldberg (2013) demonstrated that the constructional meaning is accessed quickly by asking their participants to perform a speeded lexical decision task on a target verb, after being exposed to JabThe AAAI 2017 Spring Symposium on Computational Construction Grammar and Natural Language Understanding Technical Report SS-17-02