The difference between “giving a rose” and “giving a kiss”: Sustained neural activity to the light verb construction

Highlights

• Event-related potentials measured to light verb constructions.

• Light verb constructions associated with complex semantic operations.

• Neural activity associated with semantic and syntactic argument structure mismatch.

• Late, sustained frontally-distributed negativity associated with complex events.

Abstract

We used event-related potentials (ERPs) to investigate the neurocognitive mechanisms associated with processing light verb constructions such as “give a kiss”. These constructions consist of a semantically underspecified light verb (“give”) and an event nominal that contributes most of the meaning and also activates an argument structure of its own (“kiss”). This creates a mismatch between the syntactic constituents and the semantic roles of a sentence. Native speakers read German verb-final sentences that contained light verb constructions (e.g., “Julius gave Anne a kiss”), non-light constructions (e.g., “Julius gave Anne a rose”), and semantically anomalous constructions (e.g., *“Julius gave Anne a conversation”). ERPs were measured at the critical verb, which appeared after all its arguments. Compared to non-light constructions, the light verb constructions evoked a widely distributed, frontally focused, sustained negative-going effect between 500 and 900 ms after verb onset. We interpret this effect as reflecting working memory costs associated with complex semantic processes that establish a shared argument structure in the light verb constructions.

Introduction

Most theories of argument structure assume a tight coupling between syntactic and semantic structure, such that each noun phrase maps onto a single semantic role. And, indeed, in most sentences, this is the case. For example, in a sentence like “Julius gave Anne a rose”, the giver (the Agent) is associated with the subject of the verb (“Julius”), the givee (the Recipient) is expressed as the indirect object (“Anne”), and the gift (the Theme) is expressed as the direct object (“rose”).

However, consider a sentence like “Julius gave Anne a kiss”—the so-called light verb construction. These constructions are complex predicates whose verbs are said to be semantically “light”, communicating only lexical and grammatical aspect, and the directionality of the action; the bulk of the predicative meaning stems from the event nominal within the construction (Butt, 2010, Wiese, 2006). While in a non-light construction such as “give someone a rose”, the verb “give” means “to hand over”, in “give someone a kiss”, the verb “give” only denotes a general sense of transfer and the event nominal “kiss” conveys the action itself. Thus, Julius acts not only as the Agent of the verb “give”, but also as the Agent of the direct object “kiss”, while Anne is both the Recipient of “give” and the Patient of “kiss”. This phenomenon is known as “argument sharing” (Baker, 1989, Butt, 2010, Durie, 1988, Jackendoff, 1974, Müller, 2010), and it violates the tight coupling of semantic and syntactic structure.

There have been several theoretical attempts to reconcile the lack of a direct correspondence between semantic and syntactic argument structure in the light verb construction (Hale and Keyser, 1993, Hale and Keyser, 2002, Goldberg, 2003). In this paper, we follow the Parallel Architecture framework (Culicover and Jackendoff, 2005, Wittenberg and Piñango, 2011), which allows both syntactic and semantic structure to be built independently, though the two are linked through a grammatical function tier (for further discussion, see Müller, in press, Wittenberg et al., in press). According to this theory, when a verb appears in a light verb construction with certain event nominals, the process of argument sharing is triggered: the arguments provided by the verb (in the case of “give”, the Agent, Patient, and Theme), and the arguments provided by the noun (in the case of “kiss”, Agent and Patient) need to be aligned.

As a result of the mechanisms that, according to the Parallel Architecture, are engaged during argument sharing (Culicover & Jackendoff, 2005, pp. 222–225), we predict that light verb constructions should incur processing costs during comprehension. Note that this hypothesis goes against what might be predicted on the basis of the frequency of light verb constructions, which, despite their complexity, are commonly encountered in everyday language. For example, according to the PropBank corpus (Palmer, Gildea, & Kingsbury, 2005), the most common English verbs appearing within light verb constructions, such as take, have, make, do, and give, are among the twenty most frequent verbs in English. More importantly, these verbs are more frequently encountered within light than non-light verb constructions (Wittenberg & Piñango, 2011). Thus, in the absence of other factors, these frequency data alone would predict reduced processing costs in association with the more frequent light verb construction than the less frequent non-light construction.

There have been only a few behavioral experiments examining light verb constructions. First, in a recent study, Wittenberg and Piñango (2011) asked participants to listen to German light verb constructions (e.g. “Weil der Student seiner Kommilitonin vor dem Seminar eine Zusammenfassung gab, spendierte sie ihm letzte Woche einen Kaffee”; English literal translation: “Because the student to his fellow student before class a summary gave, she bought him coffee last week.”). These light verb constructions were compared with non-light constructions using the same verbs (e.g. German: “Weil der Student seiner Kommilitonin vor dem Seminar einen Kugelschreiber gab, spendierte sie ihm letzte Woche einen Kaffee”; English literal translation: “Because the student to his fellow student before class a pen gave, she bought him coffee last week.”), or the same nouns (e.g. German: “Weil der Student seiner Kommilitonin vor dem Seminar eine Zusammenfassung kopierte, spendierte sie ihm letzte Woche einen Kaffee”; English literal translation: “Because the student for his fellow student before class a summary copied, she bought him coffee last week.”). The default Subject–Object–Verb (SOV) word order in German allowed the authors to probe processing costs at the critical verb where they predicted the effects of argument sharing would be most prominent. After these critical verbs, letter-string probes appeared on a screen and participants made a lexical decision about these probes. Participants were slower to respond to probes appearing 300 ms after the offset of the verbs in the light verb constructions, compared to the two non-light constructions. The authors interpreted this as evidence for an increased processing load in computing light verb constructions (see also Piñango, Mack, & Jackendoff, in press, for similar findings in English).

In another recent behavioral study, Wittenberg and Snedeker (in press) used a conceptual sorting task to explore the argument structure of light verb constructions in English. During a training phase, participants were trained to sort pictorial depictions of events by the number of thematic roles they encoded (e.g. a picture of man giving a woman some flowers would be classified as a ‘three role’ event: man, woman, flowers). They were then asked to sort a mix of pictures and written sentences into these different types of event structures (with different numbers of thematic roles). Despite the fact that they have three syntactic arguments, events described by light verb constructions (e.g. “The teenager is giving his rival a kick”) were most frequently grouped with event structures with two semantic roles (Agent–Patient events, e.g. “The cowboy is taming the pony”). This suggests that light verb constructions do, indeed, typically involve a non-canonical mapping between semantic and syntactic event structure. However, in about a quarter of cases, the light verb constructions were grouped with three-role event structures (e.g. Source–Theme–Goal Events, like “The businessman is passing pamphlets to the pedestrians”). This in-between pattern provided indirect evidence for argument sharing; that is, light verb constructions may be intrinsically associated with two different argument structures that can be active at the same time: an Agent–Patient non-canonical argument structure in which the number of semantic and syntactic arguments mismatch, and a Source–Theme–Goal canonical structure in which the number of semantic and syntactic arguments match.

Together, these behavioral studies provide some evidence that both the processing and final interpretation of light verb constructions involve argument sharing. Nevertheless, there are some limitations in the interpretation of the results. First, Wittenberg and Piñango (2011) used a cross-modal lexical decision task, which imposes dual task demands, potentially altering participants’ processing of the sentences (see Pickering, McElree, Frisson, Chen, & Traxler, 2006, for a critique of this method). Second, Wittenberg and Snedeker (in press) probed participants’ final interpretation of these constructions, rather the time course of their online neural processing.

There has only been one study investigating neural activity associated with light verb constructions. In an MEG study, Briem et al. (2010) carried out three experiments in German. They contrasted light verbs like “geben” (“give”) with non-light verbs like “erwarten” (“expect”), either by themselves (Experiment 1), presented together with a subject pronoun (Experiment 2), or in object-verb-subject order (Experiment 3). In all experiments, light verbs (e.g. “geben”/“give”) evoked less activity than non-light verbs (e.g. “erwarten”/“expect”).

The authors interpreted these findings as reflecting reduced lexical processing due to the semantic underspecification of light verbs. At first glance, these findings appear to contradict the behavioral findings described above, namely more cost associated with light verb constructions than non-light constructions. However, there are several confounds that limit the interpretation of Briem and colleagues’ study. First, the non-light verbs were both longer and more morphologically complex than the light verbs (the stimuli can be found in the appendix of Briem, 2009); both of these differences can explain the early effects in visual regions observed in all three experiments. Also, in the second experiment, several of the pairings between the non-light verbs and subject pronouns resulted in ungrammatical phrases, which is likely to have contributed to increased neural activity (see Friederici & Frisch, 2000). Finally, the object–verb–subject word order in the third experiment is very uncommon in German, so it is difficult to generalize the results to more naturalistic language comprehension. Moreover, because the authors did not analyze activity after the verb, any effects of argument sharing associated with light verb constructions may have been missed altogether – for example, under the Parallel Architecture, one would expect effects from argument sharing on the post-verbal subject, which would receive a semantic role from the object.

In the present study, we used a different technique – event-related potentials (ERPs) – to explore the time-course and nature of processing light verb constructions. Of most relevance to this study are three groups of ERP components, summarized below.

First, the N400 is a negative-going potential peaking at approximately 400 ms post stimulus onset, and which is thought to reflect the retrieval or access to semantic features associated with an incoming word (Kutas & Federmeier, 2011). In sentence and discourse contexts, the amplitude of the N400 reflects the match or mismatch between the semantic features associated with this word and those activated by context. The N400 therefore tends to be smaller to words that are lexically predictable versus unpredictable in relation to their context, with predictability usually operationalized using cloze probability (Federmeier et al., 2007, Kutas and Hillyard, 1984).

Second, the P600 is a positive-going potential, which usually appears between 500 and 900 ms post stimulus onset. While this waveform was originally characterized as being most closely linked to syntactic violations and ambiguities (Hagoort et al., 1993, Osterhout and Holcomb, 1992), it is also seen when the linguistic input is highly semantically implausible or incoherent (Kuperberg et al., 2006, Kuperberg et al., 2007, Kuperberg et al., 2003; for reviews, see Kuperberg, 2007, Van de Meerendonk et al., 2009 and Bornkessel-Schlesewsky & Schlesewsky, 2008, and see Kuperberg, 2013, for a recent discussion).

Finally, a set of anteriorly-distributed negativities have been associated with working memory operations required to maintain, link and select sentential and discourse constituents during online processing (e.g. King & Kutas, 1995; Mueller, King, & Kutas, 1997; Nieuwland & Van Berkum, 2008a). Of most relevance to this study is emerging evidence that a set of prolonged negativity effects, starting at approximately 500 ms post stimulus onset and lasting for several hundred milliseconds, may be associated with the computation of complex but plausible event representations. For example, Baggio, Van Lambalgen, and Hagoort (2008) reported a sustained anteriorly distributed negativity beginning at approximately 450 ms after the onset of sentence-final verbs that signaled that an anticipated ongoing event (writing a letter) was prematurely terminated (by spilling coffee on the paper). A similar pattern was observed by Bott (2010) who reported a sustained anterior negativity starting at 500 ms after the onset of critical words (e.g. discovered) that signaled an additional event had taken place (a searching event). And, in a recent study, we observed a late-onset sustained negativity effect in association with iterative aspectual coercion, in which a punctive action (e.g. pounce) must be repeated for a period of time, specified by the preceding context (Paczynski, Jackendoff, & Kuperberg, in press). We will return to a more complete analysis of these studies in the Discussion section. For now, we note that in all three studies, the meaning of the events described cannot be derived solely from combining the meanings of the individual words with the surface syntactic structure; rather, the elements in the sentence need to be (re)-combined in a non-canonical way in order to arrive at a meaningful final sentence interpretation.

In the present study, we measured ERPs as participants read three types of sentences: light verb constructions, non-light constructions, and anomalous constructions (see Table 2). Similar to the behavioral study by Wittenberg and Piñango (2011), we used German sentences with verb-final ordering, which allowed us to examine processing on the verb, after the presentation of all arguments. In the light verb constructions, a light verb was combined with an eventive noun (“eine Ansage machte”, i.e., “made an announcement”). The non-light sentences used the same light verb, but paired it with a non-eventive noun (“einen Kaffee machte”, i.e., “made a coffee”), which resulted in a non-light construction. Finally, the anomalous constructions used the same verb, but paired it with an abstract noun that, in most cases, could be combined with a different light verb, but that rendered the overall construction ungrammatical (^*“ein Gespräch machte”, i.e., “made a conversation”; note that this construction is unacceptable in German; an English example would be ^*“make a nap”). We examined ERPs evoked by the verb in each type of sentence. As the verb was identical across the three experimental conditions, any effects would necessarily arise from the combination of the verb with its arguments, rather than the lexical properties of the verb itself.

Our main focus was the contrast between the light and non-light constructions. As noted above, light verb constructions tend to be more frequent and thus more predictable than non-light constructions, which could potentially make them easier to process. The critical question was whether, despite their higher frequency and predictability, we would see evidence of argument sharing in the light verb constructions, as predicted by the Parallel Architecture (Culicover & Jackendoff, 2005). This might manifest as a late sustained negativity effect, similar to that previously observed in association with complex semantic operations involving the maintenance, computation and/or selection of non-canonical event structure representations within working memory (e.g. Baggio et al., 2008, Bott, 2010, Paczynski et al., in press).

With regards to the contrast between light verb constructions and anomalous constructions, we predicted that the anomalous verbs would produce a P600 effect, similar to that previously observed in association with the detection of other violations of overall propositional coherence (see Kuperberg, 2007, Kuperberg, 2013 for reviews), which may or may not be accompanied by an N400 effect.