Content uploaded by Eva Wittenberg
Author content
All content in this area was uploaded by Eva Wittenberg on Nov 23, 2020
Content may be subject to copyright.
�
Volume xx, Issue xx xx xx
Productivity and Argument Sharing in
Hindi light verb constructions
Published by CSLI Publications
Contents
1 Productivity and Argument Sharing 3
Ashwini Vaidya & Eva Wittenberg
1
JSAL volume xx, issue xx xx xx
Productivity and Argument sharing in Hindi light verb
constructions
Ashwini Vaidya & Eva Wittenberg, IIT Delhi; University of California, San Diego
Received Date; Revised Date
Abstract
Light verb constructions (e.g. give a sigh, take a walk) are a linguistic puzzle, as they consist
of two predicating elements in a monoclausal structure. In the theoretical literature, there has
been much interest in the linguistic analysis of such constructions across a range of grammatical
frameworks. One such proposal is event co-composition, where the argument structures of noun
and light verb merge, resulting in a composite argument structure, which has been claimed to be
the source of increased processing costs in English and German. In contrast to these languages, in
Hindi a larger proportion of the predicates are light verb constructions. Hence, we may ask whether
a Hindi speaker’s experience with light verb constructions allows them to go through the same co-
composition operation faster than a speaker of English. Our results show that Hindi speakers are
adept at the process of using light verb constructions to ‘verbalize’ predicates, more so than speakers
of Germanic languages. We argue that these data provide evidence for a case of specic linguistic
experiences shaping cognition: cost disappears with practice.
1 Introduction
One fact that all languages have famously in common is the ability to convey any category of
meaning – things, ideas, events, or states. But what varies widely across the globe is how exactly
each language packages meaning into a syntactic structure. In this paper, we explore how the
prevalence of such a packaging strategy will lead to dierent observable behaviors in the speakers of
a language.
Our main concern is a predicational strategy where a verb such as do, make, or give combines
with an event noun, such as a jump or a call to form a phrasal structure with a single meaning
(jumping, calling). Such constructions are widespread and can be found in languages as diverse as
Hindi, Persian, and English. They belong to the class of complex verb constructions; in this paper,
we focus particularly on light verb constructions (Jespersen 1965).
Light verb constructions open up an interesting perspective on how we conceive of the interaction
between the usage patterns, grammatical allowances, and grammatical possibility spaces within a
particular language on the one hand, and the fundamentals of the human mind on the other hand:
While cognitive mechanisms, concepts in the mind, and expressive needs do not vary signicantly
across the world, each language has its own grammatical preferences, combinatoric possibilities, and
frequency biases. For instance, say you want to ask your child to hug you. In English, the language-
wide preferred way to encode an action concept like ‘hug’ is via a simple verb (‘to hug’). However,
3
JSAL Volume xx, Issue xx, xx xx.
Copyright © xx, CSLI Publications.
4 / JSAL volume xx, issue xx xx xx
you can also use a light verb construction (‘to give a hug’). Crucially, the language-wide preference
for simple verbs and the frequency of specic constructions can be independent: While overall,
English has few complex predicates like ‘to give a hug’ (low language-wide systemic frequency), the
ones that it has are highly frequent (high token frequency).
In other languages, the picture is dierent: In Hindi for instance, there is a language-wide prefer-
ence to encode actions as complex verbs, thanks to a very productive light verb construction schema.
In either language, comprehenders need to construct the path from a linguistic structure to an action
concept, relying on the same mental architecture across languages.
Here, we ask whether language-wide preferences, not only token frequency, interact with this
comprehension process: Light verb constructions have been analyzed as a general process of event
co-composition, one example of constructions that compose predicative meaning from both an event
noun as well as from a verb (Ahmed et al. 2012). Given that this predicational strategy is present
across many languages and language families, but also given each language’s systemic preferences
for one predicational strategy over the other, we can ask whether a speaker’s ample experience with
light verb constructions allows her to go through the same cognitive event co-composition operation
faster than a speaker of a language in which light verb constructions are less productive.1
There has been psycholinguistic research on the processing of light verb constructions, which has
found evidence for a cost of co-composition (Piñango et al. 2006, Wittenberg et al. 2014, Wittenberg
and Piñango 2011). However, the languages studied in these experiments (English and German)
use light verb constructions overall relatively infrequently. In contrast, this paper asks whether
the process of event co-composition is observable in Hindi, a language that uses complex verbs for
nearly a quarter of its predicates. That is, if we take the nature of the cognitive process of event
composition as identical across languages (and speakers’ minds), we would still expect the overall
frequency and productivity of co-composition in Hindi, and cognitive constraints, such as working
memory (Norclie et al. 2015), to interact with grammatical processes.
Our paper is organized as follows: We begin with a theoretical description of the event composition
process. We then examine evidence for this process in the form of existing psycholinguistic studies,
before turning to a discussion of the light verb construction in Hindi. Following this, we describe
four experiments that compare how light verb constructions and their non-light counterparts are
processed, using two dierent experimental techniques. We conclude with a summary and discussion
of the results.
1.1 A model of co-composition
Many theoretical studies across grammatical frameworks have proposed analyses of light verb con-
structions, because they are challenging for the interaction of semantic and syntactic information in
language (Sag et al. 2002, Culicover and Jackendo 2005, Wittenberg 2016). Specically, in most of
the literature, the theoretical questions are related to the formal representation of compositionality:
If a light verb and a noun together form a predicate, then how does the syntactic and semantic
representation of both these elements result in the particular syntactic and semantic properties of
the light verb construction?
Usually, a verb (for instance, to describe) denotes the meaning of an event (in that case, someone
uttering something) and, its object or objects (for instance, a dance) will ll in argument slots:
describe a dance denotes a recount of a dance. In light verb constructions (for instance, do a dance),
the verb (do) does not supply the event type – we know that we are talking about a dancing event
because the predicative meaning is supplied by the syntactic object (a dance).
One account that eectively handles this problem argues that the argument structures, demanded
by the lexical semantics of both noun and light verb, overlap with each other, and a ‘shared’, or
‘composite’ argument structure, emerges in the monoclausal light verb construction (Mohanan 1997,
1Note that this is neither a Whoran question – there is no claim that there are language-specic inuences on
perception – nor is it a question about token frequency of a particular construction. Rather, we ask how language-wide
structural preferences interact with cognitive processes.
Productivity and Argument Sharing / 5
ram-ne mohan-ko kitaab dii
ram-erg mohan-acc book.f give.m.prf
subject
object
ind. object
root
(a) Syntactic arguments of the simple verb
ram-ne mohan-ko kitaab dii
ram-erg mohan-acc book.f give.f.prf
agent
theme
recipient
root
(b) Thematic roles of the simple verb
ram-ne us baat-par zor diyaa
ram-erg that topic-loc emphasis give.m.prf
object
subject
root
(c) Syntactic arguments of the light verb
ram-ne us baat-par zor diyaa
ram-erg that topic-loc emphasis give.m.prf
agent
theme
theme
agent root
(d) Composite thematic roles of noun and light verb
FIGURE 1: A simple verb with its syntactic arguments is shown in Figure (1a). The same simple
verb maps the syntactic arguments to the semantic ones in (1b). The light verb has two surface
syntactic arguments in (1c). But these emerge after two sets of thematic roles from noun and light
verb combine in Figure (1d), resulting in a composite argument structure.
Alsina et al. 1997, Ahmed et al. 2012, Durie 1988, Piñango et al. 2006). For instance, the Hindi Ex.
(1) consists of the ditransitive verb give ‘de’ with three arguments, Raam,Mohan and kitaab ‘book’.
In this case, three syntactic arguments, subject, indirect object and object, are mapped directly to
the set of thematic roles for the verb, agent, recipient and theme (Figures (1a) and (1b)).
(1) raam=ne mohan=ko kit�ab d�-ii
Ram.M.Sg=Erg Mohan.M.Sg=Dat book.F.Sg give-Perf.F.Sg
‘Ram gave Mohan a book’
(2) raam=ne us baat�=par zor d�i-yaa
Ram.M.Sg=Erg that topic=loc pressure.M.Sg give-Perf.M.Sg
‘Ram put an emphasis on that topic’
On the other hand, Ex. (2) has the light verb de, which takes two syntactic arguments: Ram and
us baat (Figure (1c)). These emerge only after the two sets of thematic roles from the noun and
light verb are combined. Both noun and light verb contribute towards the argument structure of
the sentence. The light verb de licenses the nominal predicate argument zor and the agent Raam.
The noun zor also licenses the agent Raam and the theme us baat par (see Figure (1d)).
In language comprehension, the agent argument Raam must be identied as common to both
predicates zor and de. This is sometimes described as the argument identication step (Davison
2005). In Davison’s (2005) model, following argument identication, a semantic argument merger
step will take place such that a composite, but monoclausal structure is formed. This interaction
between the syntactic arguments on the surface and the two sets of semantic arguments is char-
acteristic of light verb constructions, and it has been dened as event co-composition: a process
where “two semantically predicative elements jointly determine the structure of a single syntactic
clause”(Mohanan 1997, p. 432).
Light verb constructions have additional properties that are crucial to the event co-composition
process. Across languages, light verbs consist of a small class of high-frequency, general-purpose
verbs that are form-identical with their non-light counterpart, e.g. make, do, give, take etc., a fact
6 / JSAL volume xx, issue xx xx xx
that has led Butt (2010) to consider the light verb as a unique category that shares a lexical entry
with its non-light form. That means that at one point during comprehension, the listener or reader
needs to interpret the light verb as light rather than non-light, in order for the the event composition
process to succeed. For example, in (1), we have the verb de ‘give’ in its non-light form, but in (2),
it is part of a light verb construction.
An eect of this form-identity with full verbs is that when a comprehender resolves a verb as light,
the predicating noun provides the eventive meaning, but the verb still supplies additional aspectual
or agentive information about the event and its structure (Butt et al. 2008, Wittenberg et al. 2017).
This also has implications for the structure and nature of the semantic arguments expressed in a
light verb construction. For instance, while the non-light use of give (e.g give an orange to someone)
has three semantic roles (Source, Theme, and Goal), the light use (e.g give a kiss to someone) has
only two (Agent and Patient), with the predicating noun kiss fusing its agent role with give. These
syntactic and semantic interactions with the predicating noun result in the structure of a complex
predicate. In the next section, we review the results from the psycholinguistic literature that have
shed light on how the process of light verb construction comprehension unfolds.
1.2 Measuring event co-composition
In light verb constructions, the challenge for comprehenders is to understand that the verb, usu-
ally the only projector of sentential argument roles, is sharing this power with the event nominal.
This means that the event co-composition process lies at the interface between syntactic and lexico-
semantic representation. During real-time interpretation of such constructions, how do comprehen-
ders resolve the mismatch between the syntactic and semantic arguments in the clause?
The literature oers several psycholinguistic studies on light verb construction comprehension in
English and German. These studies have focused on collecting behavioural and electro-physiological
data at the point when the verb is read or heard (in Subject-Object-Verb or Object-Verb-Subject
structures) or at the noun (in the case of Subject-Verb-Object structures), based on the theoretical
prediction that the composite argument structure of noun and light verb must be ‘resolved’ after both
verb and noun are processed, and the co-composition process would be observable as a behavioural
or electro-physiological signal.
Briem et al. (2009) carried out three experiments to study how light verbs are processed in
German. They contrasted light verbs like geben (“give”) with non-light verbs like erwarten(“expect”)
either in contrast with a pseudo-word, by themselves, or within a sentential context. Briem et al.
(2009)’s study used MEG in order to demonstrate that light verbs (e.g.give) when presented by
themselves or in comparison with pseudo-words showed less cortical activity as compared to non-
light verbs (e.g.expect). The authors interpreted this as a result of lexically underspecied features in
the light verb. In the third experiment, when presented in a minimal sentential context using object-
verb-subject structure, a verb like give had to get resolved as either light or non-light depending
upon the presence of the noun, e.g. a kiss gives he vs. a book gives he. Here, the pattern followed
that of the previous two experiments; a non-light context give a book resulted in greater left-temporal
activation as compared to light. Briem et al. (2009) interpreted these results as evidence for distinct
brain processing areas for distinct categories of verbs. However, the authors did not measure activity
after the verb, where subsequent behavioural studies have found dierences between light and non-
light conditions (for further discussion, see (Wittenberg et al. 2014)). We now turn to some of these
experiments which report such later eects.
A handful of behavioural experiments have studied late reections of computational costs asso-
ciated with light verb construction processing, motivated by the mismatch between the syntactic
argument structure and the semantic roles dened by the construction (see section 1.1). This mis-
match was predicted to surface as computational cost after the construction has been comprehended.
Piñango et al. (2006) predicted to show a cost of processing light verb constructions at around
250-300ms after the oset of the construction. This prediction was based on studies that had shown
that the (re-)assignment of semantic roles results in slower-developing eects that can be detected
Productivity and Argument Sharing / 7
at a later point during sentence processing (Boland 1997, McElree and Grith 1995). Based on
this idea, Piñango et al. (2006) used an interference paradigm in the form of a cross-modal lexical
decision task. Piñango et al. (2006) used three conditions for their study, where the light condition
was contrasted with non-light condition, and a third condition (‘heavy’) containing the same noun,
but paired with a non-light verb:
(light) Mr. Olson gave an order last night to the produce guy
(non-light) Mr. Olson gave an orange last night to the produce guy
(‘heavy’) Mr. Olson typed an order last night for the produce guy
Participants listened to sentences containing one of the three conditions as shown above. At a
point after the object (order or orange) was heard, participants had to make a lexical decision on
a letter string that ashed on the screen. The reaction time to make a decision was taken as a
reection of the demand placed on working memory by the construction that was just heard: Slower
reaction times would reect higher computational cost.
Piñango et al. (2006) also manipulated the timing of the probe placement: Either the probe
was placed immediately at the oset of the object, or 300ms after. In the former case, the non-
light condition was signicantly slower, an eect that Piñango et al. (2006) attributed to the higher
frequency of the light verb construction. However, when the probes were placed 300ms after the noun
was heard, this eect was numerically reversed (albeit with no statistically signicant dierence),
and the light condition elicited signicantly slower reaction times than the heavy condition. Piñango
et al. (2006) concluded that the computational costs of argument sharing become apparent when
measured later, i.e after the construction is ‘disambiguated’ as light.
Wittenberg and Piñango (2011) replicated this study in German. Like Piñango et al. (2006),
they also used three conditions (light, non-light, and heavy), but since German can be verb-nal
(e.g., a hug give), the probe was placed either immediately or 300ms after the verb (not noun) was
heard. The interference immediately after the verb resulted in no signicant dierences between
the conditions. But again, when the probe was placed 300ms after the verb, listening to light verb
constructions while making a lexical decision resulted in signicantly slower reaction times than
listening to either non-light or heavy constructions. This pattern of results was interpreted by
Wittenberg and Piñango (2011) as further support for the hypothesis that the event co-composition
process can be measured as a late, more gradually developing eect, after the lexical composition
process has taken place.
However, the cost of co-composition could not be replicated in a self-paced reading study in
German using the same stimuli (Wittenberg 2013). In this self-paced reading experiment, people read
light and non-light constructions at similar speed, with only semantically anomalous constructions
being processed slower. We return to the results of this experiment later in the paper.
Yet another study (Wittenberg et al. 2014) used Event-Related Potentials to understand the
processing of light verb constructions. This study had three conditions: light (give a kiss) and
non-light (give a book), like the interference-based paradigms, and then an anomalous condition,
consisting of a non-felicitous noun-verb pairing (*give a conversation). This was done mainly to
distinguish the processing of complex, but plausible, event structures from implausible ones. The
authors found evidence for a late, widely distributed, but frontally focused negativity after the onset
of the light verb, compared to the non-light counterpart, and the anomalous condition showed a larger
positive eect associated with semantic anomalies. Wittenberg et al. (2014) interpreted their results
as reecting a working memory cost caused by the process of event co-composition, particularly the
linking of the two syntactic argument structures that surface in a single monoclausal structure.
In sum, the results of these studies generally point towards two broad themes: rst, that the cost
of event co-composition can be measured after the verb and noun combination has been processed,
using both a cross-modal task and electro-physiological methodologies, but that the eect could not
be detected in self-paced reading. Second, this cost appears to be a late eect, developing several
hundred milliseconds after the light verb construction has been licensed. Its signature is distinct
8 / JSAL volume xx, issue xx xx xx
from semantically implausible constructions, but consistent with the processing of other types of
complex events. Taken together, three out of four studies provide evidence for the cost of event
co-composition in English and German.
1.3 Light verb constructions in Hindi
Light verb constructions are found across South Asian languages, including Hindi (Masica 1993).
Seiss (2009) identies the following properties that distinguish them from other types of verbal
multiwords. First, they are always form-identical with the main or ‘full’ verbs in the language.
Second, they are restricted in their combinatorial possibilities with the predicating noun, that is,
every light verb can only combine with a certain kind of noun; and nally, they contribute subtle
semantic information in the form of telicity and agentivity, among others (Hook 1974). In describing
these construction, we need to acknowledge terminological dierences: in the South Asian linguistics
literature, light verbs are sometimes subsumed under the term ‘complex predicates’. While this term
is arguably used more widely, we deliberately use the term light verb construction, because in Hindi
‘complex predicates’ may refer to noun and light verb combinations as well as verb and light verb,
or even adjective and light verb combinations. Here, we use ‘light verb constructions’ to only refer
to complex predicates consisting of a predicating noun and a light verb.
In a Hindi light verb construction, a verb combines with another pre-verbal noun, predicate,
adjective, adverb, borrowed English verb, or noun (Ahmed et al. 2012). In this paper, we focus only
on constructions with nouns, to keep comparability to previous studies. In the sentences below, the
verb de ‘give’ is used as simple predicate in Ex. 3, but in Ex. (4), the verb de is light and combines
with a predicating noun (both examples repeated from the previous section).
(3) raam=ne mohan=ko kit�ab d�-ii
Ram.M.Sg=Erg Mohan.M.Sg=Dat book.F.Sg give-Perf.F.Sg
‘Ram gave Mohan a book’
(4) raam=ne us baat�=par zor d�i-yaa
Ram.M.Sg=Erg that topic=loc pressure.M.Sg give-Perf.M.Sg
‘Ram put an emphasis on that topic’
In English, many light verb constructions have a denominal verb counterpart (e.g., take a walk
can also be expressed by walk; (Tu and Roth 2011)). In Hindi as well, some light verbs (like khoj
kar ‘search do’) will co-exist with their denominal verbs (khojnaa ‘to search’).
However, the distribution of Hindi light verb constructions diers from English and German
because the vast majority of the light verb constructions in Hindi do not have an denominal verb
counterpart. While both light and denominal verbs co-exist in English, the formation of denominal
verbs in Hindi has ceased to be freely productive (Davison 2005).
Butt (2010) notes that Hindi light verb constructions act as a verbalizers in order to create new
predicates and to incorporate borrowed items into the language (e.g. email kar ‘email do; email’).
Light verb constructions are highly productive and are sometimes described as “a preferred way of
augmenting the creative potential of the language” (Kachru 2006)[93]. This is reected in corpora:
If English has approximately 7000 simplex verbs, Hindi has only 700 (Vaidya et al. 2013).2
1.4 Frequency and co-composition
In all of the experiments that were discussed in section 1.2, the cost of argument structure composi-
tion was interpreted as being due to the real-time processing of the light verb construction, because
composite argument structures are built ‘on the y’.
An alternative view to this would be that light verb constructions are stored (in the manner of
non-compositional idioms) in the lexicon. In order retrieve the right syntax-to-semantics mapping,
native speakers would merely detect the construction as light, and retrieve the stored argument
structure associated with a given construction. If light verb constructions were stored and retrieved
2Based on counts from English PropBank and Hindi PropBank, respectively.
Productivity and Argument Sharing / 9
as non-compositional units like this, instead of assembled incrementally and compositionally, what
one would predict for real-time processing is that the higher the frequency of a given construction,
the faster the recognition; reaction times for light verbs should therefore be faster.
Crucially, in all of the experimental results reported in section 1.2, the token frequency of any
given light verb construction was higher than its non-light counterpart. For example, make,have, or
give are more likely to occur with a light noun (forming a light verb construction) than with a non-
light noun in English or German. That is, the collocational frequencies of light verb constructions
such as give someone a hug are higher than the collocational frequencies of non-light constructions
such as give someone a book.
Based on collocational frequency alone, then, one should expect speakers of English and German
to be able to process light verb constructions with more ease than their non-light counterparts.
However, the results of the psycholinguistic studies do not concord with this prediction (Piñango
et al. 2006, Wittenberg and Piñango 2011). In fact, what we see is the reverse: Reaction times are
slower despite higher frequency. This pattern of results was interpreted as the cost of co-composition
overriding any advantages of collocational frequency, when measured in reaction times at the verb.
These results demonstrate that at least in languages like English and German, where the language-
wide productivity of light verb constructions is relatively low, the higher collocational frequency of
an individual light verb construction does not facilitate processing.
At the same time, as mentioned above, there is a great deal of cross-linguistic variation in how
frequently light verb constructions are used. Using a corpus of 50 Wikipedia articles, Vincze et al.
(2011) estimated that in English, about 9.5% of the predicates are expressed by light verb construc-
tions. In Hindi, the proportion of light verb constructions is about 37% of roughly 37,600 predicates
in the Hindi Treebank (Vaidya et al. 2013). Hindi is not an exception when it comes to language-
wide frequency of complex predicates: In a language like Persian for instance, only about 115 simple
verbs are commonly used, whereas almost all the rest are light verb constructions (Sadeghi 1993).
These numbers highlight the dierences in systemic frequency of the light verb construction across
languages, and with it, the grammatical productivity of expressing a predicate with a complex verb
in a given language.
Thus, in a language like Hindi, where language-wide frequency of the light verb construction
is higher than English, we could expect to nd that the light verb constructions behave similarly
to English, i.e the overall token and language-wide frequencies do not facilitate processing when
measured at the verb. Alternatively, we may also nd that the overall systemic productivity of the
light verb construction results in greater exposure to the type of composite argument structures
associated with this construction. This sensitivity towards previously seen argument structures
makes them easier to process (Mitchell et al. 1995). If this is the case, then we would expect that
the systemic productivity of light verb constructions in Hindi would facilitate processing. Previous
exposure to light verb constructions could imply that such argument structures are stored, or that
Hindi native speakers are much more ecient at the process of event co-composition itself.
Some cross-linguistic studies support this idea. Structural preferences in dierent languages will
correspond to the frequency with which they appear in those languages. For instance, preferences in
relative clause attachment to the head noun in ambiguous sentences seem to dier cross-linguistically.
While French and Spanish prefer ‘higher’ attachment, i.e. to a noun higher in the structure, English
and Italian pattern ‘lower’ (Cuetos et al. 1996). These preferences may be tied to the frequency
with which these structures appear across languages (although see Grillo and Costa (2014)’s paper
which suggests that other factors may also be involved).
1.5 Frequency and predictability
Token frequency of a particular light verb is context-independent. In comparison, a word or lexical
item’s predictability depends upon its immediately preceding context, and the particular colloca-
tional frequency of a light verb and noun composition can play a role in facilitating its retrieval. Eye
tracking studies have shown that the eects of frequency and predictability on reading are distinct
10 / JSAL volume xx, issue xx xx xx
and additive in nature (Kennedy et al. 2013). This implies that if a word is both low frequency and
unpredictable, it will have greater cost than a word that is high frequency and predictable. There
is evidence for the eect of both frequency and predictability on reading times, and Staub (2011)
have also shown at these are distinct factors that do not necessarily interact with each other.
In the context of light verb constructions, it is close to impossible to control for the collocational
frequency of a noun and light verb (it is almost always likely to be greater than the non-light).
But we can control for the predictability of both light and non-light constructions, such that they
are matched. This will help us tease apart the eect of familiarity or frequent exposure to event
co-composition (as a result of language-wide frequency) vs. exposure to the individual noun-light
verb combination in its token frequency. Crucially, if both light and non-light constructions are low
in predictability, then any facilitation in processing can be attributed to systemic frequency, and not
to the individual collocation.
1.6 Experimental Predictions
In the previous sections, we have elaborated on the theoretical motivation for event co-composition,
the measurement of this phenomenon using behavioural and electrophysiological paradigms, and its
relationship with frequency and predictability. With respect to the processing of Hindi light verb
constructions, our experiments ask whether we can replicate the English and German data pattern
in Hindi, a language that uses light verbs much more frequently as a predicational strategy than
Germanic languages.
If comprehenders across the globe perform co-composition the same way, we should replicate the
previous results, with light verb constructions taking longer to process than non-light constructions.
However, if the systemic prevalence of light verb constructions in a language (and consequently
a greater exposure to those constructions) inuences the speed at which comprehenders perform
cognitive operations such as co-composition, then we would expect light verb constructions in Hindi
to be processed faster or equally fast as non-light constructions.
In order to account for the predictability of individual lexical collocations, we control for the
predictability of light verb constructions and their non-light counterparts. If light verb constructions
are processed faster than non-light constructions under this manipulation, we can conclude that any
dierence found in previous studies is not due to individual items’ predictability, but to adeptness
with complex verbs as a predicational strategy.
We test these predictions in four experiments, one of which is a self-paced reading study, and the
remaining three use the cross-modal lexical decision task paradigm.
2 Experiments
In this section, we report four experiments on the comprehension of light verb constructions in Hindi,
to understand whether the high frequency of complex predicates will lead to dierent processing
patterns from English and German.
2.1 Experiment 1: Self-paced reading
Experiment 1 was designed to ask whether light verb constructions incur a processing cost, compared
to their non-light counterparts, in a self-paced reading study. As Hindi is a verb-nal language like
German, the light verb will likewise appear at the end of the sentence. If we were to nd results
similar to those found in German, we would expect a dierence in the processing of the light condition
relative to the non-light condition in the verb region or right thereafter. In addition to both light and
non-light conditions, we also include an anomalous control condition that combines a light verb with
an incompatible noun (see Table 1), to distinguish eects that are due to semantic implausibility
from those that could be a result of event co-composition. A similar control condition was used for
German, both in a behavioural as well as in an ERP task (Wittenberg 2013, Wittenberg et al. 2014).
We predict that analagous to Wittenberg’s 2013 results, the anomalous condition will trigger
longer reaction times compared to light or non-light constructions, because of the semantic in-
Productivity and Argument Sharing / 11
Context phrase:
Light/Non-
Light/Anomalous:
Continuation
apne samay=ka prabandhan karnaa mushkil
hai isiliye...
own.obl time=Gen management.M do.inf dicult
be.Pres.sg therefore ..
‘It is dicult to manage one’s time,
therefore ...’
|
adhyapak=ne vidyarthi=ko calender/bhaashan/*silsilaa diyaa ...
teacher=Erg student=Acc calender/speech/*happening give.perf.3.M.Sg
‘the teacher gave the student a calen-
der/speech/*happening ...’
aur kuch aasaan upaay bhi bataaye
and some simple solution.pl also tell.Pl
‘.. and gave (him) some useful sugges-
tions’
TABLE 1: Example sentence showing all three conditions.
compatibility between noun and light verb in anomalous constructions. For light and non-light
constructions, we predict that if the frequency of complex verbs as a predicational strategy inu-
ences speed of co-composition, then light verb constructions will be processed faster or equally fast
as non-light constructions at the verb and thereafter. But if comprehenders across languages per-
form co-composition similarly, we would expect longer reading times for light verb constructions,
compared to non-light constructions.
2.1.1 Method
Participants read sentences in a masked word-by-word self-paced reading paradigm. We used Ibex
Farm for presentation (Drummond 2007). Participants were recruited using Amazon Mechanical
Turk. We included a participant screening task that included a series of 8 puzzle questions. Par-
ticipants were asked to choose between two Hindi sentences, where one was grammatical and the
other contained an agreement error. This ensured that the participants were able to make basic
grammaticality distinctions in Hindi. This test was introduced before the self-paced reading items
were shown as a way to prevent non-Hindi speaking Turkers from participating.
The experiment was preceded by four practice items, followed by 15 experimental items in a
Latin square design. We also included 20 llers, half of which were semantically anomalous. Each
experimental and ller item was followed by a comprehension question about the sentence, with two
choices (Y/N).
2.1.2 Materials
Fifteen experimental sentences were created for three conditions: light, non-light and anomalous, all
using the verb de ‘give’, which can appear both in light and non-light contexts.
Each sentence consisted of a short context phrase, followed by the main sentence ending with the
verb diyaa and a continuation. Table 1 shows an example sentence across three conditions: light,
non-light, and anomalous. All the stimuli sentences were minimal pairs with either a non-light,
light, or anomalous noun. In the example shown in Table 1, these are calendar/speech/*happening
respectively, where the noun *happening is semantically anomalous in combination with diyaa. A
list of all experimental items used in this experiment is given in the appendix.
Frequency Norming. As mentioned in section 1.3, Hindi light verb constructions are highly
productive. While it is impossible to control the productivity and frequency of a construction in
a speaker’s language system overall, we can control for individual frequency of a word. Thus, we
matched the frequency of pre-verbal nouns across conditions.
To obtain frequency data, we used a corpus consisting of 17 million tokens from BBC Hindi (6.5
12 / JSAL volume xx, issue xx xx xx
million) and the Hindi Wikipedia (10.5 million). This corpus was tokenized and tagged with parts of
speech to calculate the frequencies of the nouns as well as the collocational frequencies of the noun
and light verb (Reddy and Sharo 2011). As expected, the collocational frequency of noun and verb
was greater for light verb constructions (Mean: 13.62 pairs per million) than non-light constructions
(Mean: 1.04 pairs per million); t=3.68, p= 0.002 in a two-sample t-test. As expected, anomalous
constructions were signicantly lower in frequency (Mean: 0.027 words per million) as compared to
non-light (t=2.19 p=0.04 in a two-sample t-test). Anomalous were also much lower compared to light
verb constructions (t= 4.5, p < 0.0001 in a two-sample t-test) (Although the number for anomalous
should have been zero, two anomalous nouns pasand ‘like’ and ghoshanaa ‘declaration’ had counts
of 8 and 4 respectively, perhaps due to tagging errors in the corpus. However, the anomalous pairs
are indeed semantically anomalous).
We were able to control the frequency of the preverbal noun across all three conditions. On
average, nouns in the light condition appeared 91.3 times per million, in the non-light condition 92.9
times per million, and in the anomalous condition also 92.9 times per million tokens. There were
no signicant dierences in noun frequency across light and non-light conditions using a two-sample
t-test (t=-0.05, p=0.96) or anomalous and non-light conditions (t=-0.0004, p=1).
Acceptability Norming. We also conducted acceptability ratings across all three conditions with
16 native Hindi speakers, who were students of IIT Delhi (11 males, average age: 21.9). Participants
were asked to rate sentences on a 7-point scale, ranging from 1-Unacceptable to 7-Acceptable. The
average acceptability rating for the light sentences was similar to the non-light (6.22 light, SD=0.85;
5.64 non-light, SD=0.96), while the anomalous sentences had an average rating of 2.97 (SD=1.4).
There was no signicant dierence between the acceptability ratings of light and non-light in a two-
sample t-test (t= 1.74, p= 0.09). There were signicant dierences in the ratings between non-light
and anomalous in a two-sample t-test (t=6.06, p<0.001), which is to be expected.
Cloze probabilities norming. 44 Hindi native speakers, who were students of IIT Delhi (24
males, average age: 22.5), provided sentence continuations for the verb in all three conditions, such
as in 8, which had a missing sentence-nal verb. The participants were requested to complete the
sentence in the most natural way possible.
(8) rohan=ki daadi=ne use promotion milne par badhaii ...
rohan=Gen grandma=Erg him promotion get.Inf on congratulations ...
‘On getting promoted Rohan’s grandmother (gave) him congratulations’
The verb in the light verb condition was predicted 76% of the time, signicantly more often than
the verb in the non-light (49.88%) and the anomalous conditions (0.2%) A two-sample t-test showed
signicant dierences in the light and non-light cloze predictions t= 2.83, p= 0.008). For the light
and anomalous cloze predictions as well, a two-sample t-test showed signicant dierences t=6.68
and p<0.00001. Anomalous and non-light also showed a signcant dierence in a two-sample t-test
t= 3.83, p<0.0001. Thus, the verb in the light condition was highly predictable, compared to the
other two conditions. This is similar to data from German (Wittenberg et al. 2014).
2.1.3 Participants
154 participants completed the experiment on Amazon Mechanical Turk. We selected only those
participants who scored above 75% in the Hindi agreement puzzle questions (they had to get at least
6 out of the 8 questions correct). These participants on average had a comprehension score of 88%.
This resulted in a total of 101 participants with a mean age of 33.5 years (21 females).
2.1.4 Results
We t a linear mixed model to log-transformed reaction times with condition as xed eect; the con-
ditions were treatment-coded with the reference level as non-light, and items and participants were
random intercepts (including random slopes for items and participants resulted in non-convergence).
The t-values from the linear mixed model were approximated to p-values. The pnorm function in R
was used to compute the probability density of the region above the obtained t-values. Since this is
Productivity and Argument Sharing / 13
Region Noun Verb Post-verb-1 Post-verb-2 Sentence
end
Light vs.
Nonlight
t-value -0.76 0.45 -0.56 -0.92 -0.14
p-value 0.44 0.65 0.57 0.35 0.89
Anomalous
vs. Nonlight
t-value 1.0 4.35 2.73 0.48 0.75
p-value 0.32 <0.001* <0.01* 0.63 0.45
TABLE 2: T-values and p-values in the regions after the noun in Experiment 1. The critical region is
at the verb. Signicant eects in bold.
Region Noun Verb Post-verb-1 Post-verb-2 Sentence end
Light 753.38 724.82 591.28 580.27 593.32
Non-Light 709.97 676.71 566.36 573.79 569.59
Anomalous 783.61 813.32 637.5 587.79 584.82
TABLE 3: Reading times for comparison between the three conditions for the regions following the
noun until sentence end, in milliseconds. The critical region is at the verb, where the dierence in
the anomalous and non-light condition is signicant.
a two-tailed test, the obtained probability value is then multiplied by 2 to give us the approximated
p-values. Table 2 gives an overview of the signicance pattern for the regions following the noun.
The three conditions did not dier signicantly in the regions preceding the noun. At the noun
itself, we did not nd a dierence in reading times between conditions (see Table 3). Reading times
at the verb indicate that the verb in the anomalous condition was read signicantly slower than the
verb in the non-light condition (t= 4.34,p < 0.0001), but there was no signicant dierence at the
verb between the light and the non-light conditions (t= 0.45,p= 0.65). The slowdown incurred
by the anomalous condition also carried forward to the rst postverbal region, where the dierence
between non-light and anomalous was still signicant (t= 2.72,p= 0.006), but for the light vs.
non-light conditions, there were no signicant dierences in reaction times after the verb was read.
For t-values (and their approximated p-values) across all regions in the sentence, please refer to
Table 4 in the Appendix. Table 5 in the Appendix also provides the mean reaction times (and SDs)
for all regions in the Appendix.
2.1.5 Discussion of Experiment 1
This study compared reading times between light verb constructions, non-light constructions, and
anomalous constructions. We did not nd any dierences in reading times between the light and
non-light conditions, although the anomalous condition was read signicantly slower than the other
two at the verb, and in the region immediately following the verb.
This work is directly comparable to Wittenberg (2013)’s German self-paced reading study, which
also included our three conditions (light, non-light, and anomalous), and showed a similar pattern of
results as Experiment 1: a slower read anomalous condition and no detectable dierence in reading
times between light and non-light constructions.
Thus, our Experiment 1 serves as a conceptual replication of the study in German. At the
same time, as discussed in Wittenberg (2013) as well, self-paced reading paradigms may not be
able to detect more elusive semantic eects. While this method has been shown to reliably detect
semantically or syntactically implausible constructions (Mitchell 2004), it may be less eective at
capturing the plausible but subtle event co-composition processes in light verb constructions.
In section 1.2, we had reviewed the processing of light verb constructions in English and German
using interference tasks, specically the cross-modal lexical decision task (Piñango et al. 2006, Wit-
tenberg and Piñango 2011). Such a paradigm was used to detect the eects of processing costs by
placing an additional demand on working memory (e.g. Piñango et al. 2006, Kamienkowski et al.
14 / JSAL volume xx, issue xx xx xx
600
700
800
context subject object nom_mod noun verb postverb1 postverb2 sent_end
Region
Reaction Time
Condition
NonLight
Anomalous
Light
Reading times (N=101)
Experiment 1
FIGURE 2: Mean reaction times for anomalous, light and non-light conditions for nine regions of
interest in the sentence. The error bars show standard error. The critical region was at the verb,
where the anomalous condition is signicantly dierent from the non-light condition. Light and non-
light are not signicantly dierent at any region in the sentence. The error bars represent standard
error.
2011). In an interference paradigm, a deliberate interference with working memory following the
verb may slow down the processing of structures which are in fact semantically plausible and gram-
matical, and may otherwise be processed similar to any other grammatical sentence, but posit an
increased demand on working memory due to resolving the mismatch in syntax and semantics, like
light verb constructions. In the experiments that follow, we use the cross-modal lexical decision
task to test our predictions. Using the cross-modal decision task, we can manipulate the temporal
placement of the probe in order to capture these eects.
2.2 Experiment 2
In the previous experiment, we found no reliable dierences between reading the light and non-light
constructions using a self-paced reading study. In this set of experiments, we decided to investigate
the same questions using a dierent paradigm, particularly to understand whether an enhanced
demand on working memory at the verb would capture any ne-grained dierences between the
conditions.
Hence, this experiment was also designed to ask when dierences in processing between light
and non-light would be apparent. Just like previous German and English studies (Wittenberg
and Piñango 2011, Piñango et al. 2006), we placed the probe immediately after the verb (in this
experiment) and 300ms after the verb (in Experiment 3), and measured reaction times to the lexical
decision.
2.2.1 Method
Both experiments use a cross-modal lexical decision task paradigm. Participants heard sentences in
a light, non-light or anomalous condition. In Experiment 2, a string unrelated to the sentence was
visually presented immediately after the verb was heard; in Experiment 3, the probe was placed
300ms after verb oset. Participants had to decide whether the string was a word or a non-word
(lexical decision). The sentences were pseudo-randomized in a Latin square design, and the same
probe word was used in all three conditions.
Each participant also heard 25 ller sentences, and was asked 20 comprehension questions on both
Productivity and Argument Sharing / 15
the ller and experimental items. Out of the 25 ller sentences, 15 were semantically anomalous.
Thus, each participant heard a total of 40 sentences, of which half were semantically anomalous.
After each sentence, there was a pause of 1500ms and then the next sentence was heard. Each
experiment was preceded by a trial session where participants were familiarized with the task. The
cross-modal lexical decision task was coded using a browser-based presentation software, jsPsych,
version 5.0.3 (de Leeuw 2015), with a custom plugin for the cross-modal lexical decision task
paradigm.
2.2.2 Materials
The sentence materials for this experiment were identical to the ones created for Experiment 1.
Experimental and ller sentences were recorded by a female native Hindi speaker in randomized
order during a single setting. After every ten sentences, the recording was paused to ensure that
the tempo and volume was not inconsistent. All the sentences were then checked by another Hindi
native speaker to ensure that there was no variation in the volume and tempo for each item. For
each sentence, the oset up to the verb for each item and condition was noted. The sentence length
for each item prior to the verb did not vary signicantly across conditions. The mean length (in ms)
of the sentences prior to the critical region of the verb (i.e. sentence prexes) were 6380 ms for the
light condition, 6293 ms for the nonlight condition and 6391 ms for the anomalous condition.
In order to investigate if the prex (i.e., the region before the light verb) was signicantly dierent
across the 3 conditions, we t a linear regression model with the conditions as the independent
variable and the length of the prex (in ms) as the dependent variable. Treatment contrast coding
was used with the non-light condition acting as the reference level. The result showed no signicant
dierence between the baseline condition and the other conditions (Light p=0.8, t=0.24; Anomalous
p=0.3, t=0.7).
Lexical Probes For each experimental sentence, we created lexical probes that were semantically
unrelated to the items. We recorded individual reaction times for the probe words in a separate lex-
ical decision task. A total of 16 native speakers of Hindi (11 males, average age=22.31) participated
and carried out a lexical decision task, where a string in the Hindi Devanagari script was ashed on
the screen and participants had to decide whether it was a Hindi word or a non-Hindi word. A total
of 102 words (48 words and 54 non-words) were presented to the speakers in a randomized order.
The probes were presented visually on a screen in Devanagari. Out of the 102 words we chose 15
words for the experimental sentences. In isolation, these words had a mean reaction time of 717.53
ms (SD= 33). A one-sample t-test showed that they did not dier signicantly from each other
(t=0, p=1). The same probe was used across all three conditions of a single item in the experiment.
Each word was paired with one experimental item- the same across all conditions. As the items were
presented in a Latin square, participants saw only one of the three conditions, and consequently each
word probe was seen only once. We also chose 25 non-words that only appeared with the llers. For
non-words, the mean RT was 1020.87 ms (SD = 93).
2.2.3 Participants
83 native speakers of Hindi participated in Experiment 2, recruited through Amazon Mechanical
Turk. 39 participants (mean age=30.5) were included as part of the analysis based on performance
in comprehension questions (>70%) and (>60% )accuracy at the word-non-word task.
2.2.4 Results
We used a linear mixed eects model as before, using log reaction times with condition as xed eect
and item and subject as random intercepts. Light, Non-Light and Anomalous were treatment-coded
with Non-Light as the reference level. Mean reaction times for the three conditions were 1,172 ms for
the anomalous, 1,190.8 ms for the light, and 1,271 ms for the non-light condition as shown in Figure
3a. In line with results from English and German, there were no signicant dierences between light
and non-light (t=−1.53,p= 0.19) or non-light and anomalous (t=−1.53,p= 0.21).
16 / JSAL volume xx, issue xx xx xx
(a) Mean Reaction times for lexical deci-
sions in Experiment 2 (Cross modal with
probe immediately after the verb
(b) Mean Reaction Times for lexical deci-
sions in Experiment 3 (Cross-modal with
probe shown 300ms after the verb.
FIGURE 3: Results for Experiments 2 and 3.
Productivity and Argument Sharing / 17
2.3 Experiment 3
This experiment used the identical paradigm as Experiment 2, i.e., the cross modal lexical decision
task, and the same set of materials as before. The only dierence was that the lexical probe was
shown 300 ms after the verb was heard (in contrast to Experiment 2, where it was shown immediately
after the verb).
2.3.1 Participants
60 Hindi native speakers (36 male, average age=20.77), who were students at IIT Delhi, participated
in the experiment. Out of these, four participants were excluded due to less than 70% accuracy on the
word-non-word identication task, and two due to poor performance on comprehension questions.
We were left with a total of 54 Hindi speakers.
2.3.2 Results
The mean reaction times for all three conditions are shown in Figure 3b (anomalous: 1,194.3 ms,
light: 1,177.8 ms, and non-light: 1,162.7 ms). Again, light, non-light, and anomalous were treatment-
coded with non-light as the reference level in the analysis. The linear mixed model showed that there
was no signicant dierences between the light and non-light condition (t= 0.28,p= 0.77), and
also no dierence between non-light and anomalous conditions (t= 0.47,p= 0.63).
2.4 Discussion of Experiments 2 and 3
Experiments 2 and 3 showed that when reaction times to a word probe were measured either im-
mediately after the verb (Experiment 2), or 300ms after the verb (Experiment 3), there was no
dierence in reaction times to probes between the light and non-light conditions.
This lack of dierence in reaction time between the light and the non-light condition ts with the
experimental results of the self-paced reading study (Experiment 1), where there was no evidence for
a dierence in processing cost as measured by reading times between those two conditions. However,
the lack of dierence in reaction times to light versus non-light constructions diers from previous
results using the same paradigm in English and German, where dierences were found reliably with
late probe placement (Piñango et al. 2006, Wittenberg and Piñango 2011), and is a useful point of
cross-linguistic comparison.
Unlike in the self-paced reading task, the anomalous condition did not result in slower reaction
times in either of the probe placements, although the set of items used across all three experiments
was the same. The fact that the lexical decision task is not sensitive to the anomalous condition at
odds with the results of Experiment 1. We have little doubt in the adequateness of our sentences,
because Experiment 1 showed that the anomalous condition caused the expected slowdown. Instead,
the lack of eect for anomalous sentences may be explained with reference to Wittenberg et al. (2014),
which found a ‘semantic P600’ in response to anomalous constructions – a neural signature that has
been described as reecting a violation of overall propositional coherence, triggered by impossible,
unparseable combinations (see (Kuperberg 2013, Kuperberg et al. 2020) for reviews). We would
not expect such constructions that render event composition impossible to interfere with working
memory later on; thus, a lack of slowdown in our working memory interference task is not completely
surprising. We suggest to study this in the future.
Based on results from English and German, we trust the cross-modal decision task paradigm
itself, and its sensitivity to semantically plausible but complex constructions. With this premise,
the cross-linguistic dierences between the light and non-light condition provide an important point
of comparison for the construction in these languages (see section 3 for more discussion on this point),
and we argue that taken together, the results of these three experiments imply that in Hindi, the
event composition process does not result in a slowdown in reaction times at the light verb, unlike
(Piñango et al. 2006, Wittenberg and Piñango 2011). This suggests that the event composition
process does not seem to incur a cost for Hindi native speakers like it does for their German or
English counterparts.
We also note that light verb constructions have a greater collocational frequency and greater
18 / JSAL volume xx, issue xx xx xx
predictability as compared to their non-light counterparts (see section 2.1.2). But we cannot be
sure whether it was language-wide preference that facilitated processing, or whether the greater
predictability of the individual light verb construction was more helpful, compared to previous
work on German and English. In other words, it is possible that there may be context-dependent
predictability eects which are reducing the computational cost of processing light verbs.
In Experiment 4 that follows, we control for the predictability of the light verb construction
in order to tease apart both these frequency eects. We match the predictability between light
and non-light constructions, operationalized through cloze probability. If we were to control the
predictability of both light and non-light conditions, rather than keep them varied, we may be able
to show more clearly the eects of event composition when measured at the verb. To avoid a oor
eect in reaction times, we matched both conditions to be equally low in predictability.
2.5 Experiment 4
The aim of this experiment was to control both light and non-light conditions for token-based
predictability, using the same paradigm as in Experiments 2 and 3, i.e. the cross-modal lexical
decision task. Here, we can tease apart two factors that could have contributed to the null eects
between light and non-light constructions in Experiments 1-3.
First, these results could have been due to Hindi native speakers’ experience with the language-
wide systemic frequency of the light verb construction in Hindi as a means to express predicates.
If this is the case, we should nd faster reaction times to light verbs when they are matched in
cloze predictability: Hindi native speakers should be less surprised to hear a light verb construction
than a non-light construction. However, if experience with the individual construction was driving
reaction times, then we should nd slower reaction times in the light condition, compared to non-
light conditions, when cloze probabilities are matched. In this experiment, the anomalous sentences
as control condition were omitted, focusing only on the comparison between the light and non-light
constructions.
2.5.1 Method
This experiment also used a cross-modal lexical decision task paradigm. Participants heard sentences
in either the light or non-light condition. A lexical probe was presented 300ms after the verb oset
and participants had to decide whether the string was a word or a non-word. The sentences were
randomized in a Latin square design and the same probe word was used in both conditions. Each
participant also heard 15 (grammatical) ller sentences. The participants were asked 10 compre-
hension questions. Each participant heard a total of 30 sentences, half of which were experimental
and the other half were llers. In a manner similar to Experiment 2 and 3, there was a trial session
that familiarized participants with the task. The presentation software used was also the same as
Experiment 2 and 4 (jsPsych, version 5.0.3 (de Leeuw 2015)).
2.5.2 Materials
40 items were constructed, in the format described in Table 1, with a context sentence, a sentence
containing either the light or non-light verb and a continuation. All items included the same light
verb (de ‘give’), as in Experiments 1-3.
A sentence completion task was used to calculate the cloze probabilities of the items. 16 native
Hindi speakers (Mean age 33.06, female=12) were shown an incomplete sentence leading up to the
light verb and were asked to complete it as naturally as possible. The items were presented using
Ibex Farm (Drummond 2007). The responses were coded with respect to how often the light verb
(de ‘give’) was predicted. Those items that were at chance (i.e. with a cloze probability of 50-60%)
were not considered. 10 such items were removed. From the remaining 30 items, a subset of 15
low-cloze items were chosen. In the low-cloze group, both light and non-light conditions were almost
equally predictable (light: 45% and non-light: 42%) with no signicant dierence between the two
in a two-sample t-test (t=0.42, p=0.67). The remaining items were discarded.
The experimental items were recorded by a female native Hindi speaker in randomized order
Productivity and Argument Sharing / 19
during a single setting. After every ten sentences, the recording was paused to ensure that the
tempo and volume remained consistent. All sentences were then checked by another Hindi native
speaker to ensure that there was no variation in the volume and tempo for each item. For each
sentence, the oset up to the verb for each item and condition was noted. The sentence length for
each item prior to the verb did not vary signicantly across conditions (mean sentence length for
light prior to the verb was 13333.33 ms and mean sentence length for non-light was 13226.27 ms.
In order to investigate if the prex (i.e., the region before the verb) was signicantly dierent in
the two conditions, we t a linear regression model with the conditions as the independent variable
and the length of the prex (in ms) as the dependent variable. Treatment contrast coding was used
with the non-light condition acting as the reference level. The result showed no signicant dierence
between the baseline condition and the light condition p=0.89; t= 0.14.
2.5.3 Norming of sentences
The pre-verbal nouns in the light and non-light conditions were matched for frequency using a large
corpus of 60 million tokens (Kilgarri et al. 2010). The corpus was tokenized and tagged with parts
of speech to calculate the noun frequencies and the collocational frequencies of noun and light verb.
The pre-verbal nouns were matched for frequency in the light and non-light conditions, with
nouns in the light condition appearing 55.33 times per million and nouns in non-light appearing
58.88 times per million. A two-sample t-test showed no signicant dierence between the two
conditions (t=−0.18, p = 0.8).
As seen in Experiment 1, the mean collocational frequency of the predicating noun and light verb
was greater than the non-predicating noun and verb, which is to be expected (Mean light: 16.83 per
million words, Mean non-light: 2.09 per million words). Despite the higher collocational frequency
of the light condition, note that light and non-light collocation were not signicantly dierent in
terms of predictability (see Section 2.5.2).
2.5.4 Lexical Probes
The lexical probes used in this task were identical to those used in Experiments 2 and 3. A list of
all probes and items can be found in the Appendix.
2.5.5 Participants
120 native Hindi speakers (Mean age 31 years) participated in the experiment on Amazon Mechanical
Turk, out of which 59 remained after ltering based on performance on comprehension questions
(more than 70 % correct) and accuracy in the lexical decision task performance (more than 70 %
correct).
2.5.6 Results
Reaction times to probes in the light condition were on average 63 ms faster than probes in the
the non-light condition (Mean light= 1265 ms, Mean non-light=1328 ms). This pattern is shown in
Figure 4. As in the previous studies, we t a linear mixed model predicting log reaction times from
construction (light vs. non-light) as xed eect, and item and subject as random intercepts. This
model showed that the dierence between reaction times to probes presented while hearing the light
vs. non-light construction were signicant (t=−2.12,p= 0.01).
2.5.7 Discussion of Experiment 4
The results of Experiment 4 show that when the predictability of light and non-light construction
is equally low, listening to light verb constructions incurred signicantly faster reaction times while
making a lexical decision to unrelated probes than listening to non-light constructions.
This eect should not be due to predictability or frequency of the individual constructions. The
matched cloze probability ensured that the number of possible verbs that could appear after the noun
was roughly the same for both conditions. Similarly, the nouns in both conditions were matched
for frequency. After controlling for these factors (particularly, cloze for this experiment), we had
predicted that the eect of the co-composition process would show up in the form of longer reaction
20 / JSAL volume xx, issue xx xx xx
FIGURE 4: Mean reaction times for Experiment 4 (cross-modal lexical decision task with matched
cloze probabilities).
Productivity and Argument Sharing / 21
times in the light condition, while the results show the opposite pattern.
We can interpret the results in two ways: First, the collocational frequency of the predicating
noun and the light verb together resulted in faster reaction times at the verb, superseding any
predictability eects for the individual constructions. Another way to interpret these results is the
language-wide productivity of the construction in Hindi: Hindi native speakers develop a greater
eciency in co-composition, where the light verb construction is being composed faster due to
greater practice with this predicational strategy in the language overall.
3 General Discussion
This paper explored how the systemic, language-wide frequency of a predicational strategy aects
cognitive processing, using light verb constructions in Hindi as a test case. Experiment 1 used a
self-paced reading paradigm, and while people slowed down reading anomalous constructions, there
was no dierence in reading speed between light and non-light constructions. Experiments 2 and 3
used a cross-modal lexical decision paradigm, which has been shown to be more sensitive to semantic
composition processes. However, regardless of the timing of the lexical decision task, we failed to
detect any dierence in reaction times to probes presented while people listened to light vs non-light
constructions. Experiment 4 used the same cross-modal task, where light and non-light constructions
were controlled for predictability. Crucially, in this nal study, light verb constructions led to faster
reaction times than non-light constructions. Most of these data stand in contrast to German and
English data, where longer reaction times and higher processing costs were found for light verb
constructions (Piñango et al. 2006, Wittenberg and Piñango 2011, Wittenberg et al. 2014, but see
Wittenberg, 2013).
These results show that for speakers of Hindi, where the language-wide systemic frequency of
light verb constructions is greater than in English or German, the event co-composition process does
not incur a measurable computational cost when measured at the verb, or in the region immediately
following the verb. We interpret these ndings to imply that the process of event co-composition is
facilitated by a greater exposure to the composite argument structure of light verb constructions in
the language.
Thus, we interpret these datasets as evidence for language-specic eects of the systematic preva-
lence of a predicational strategy on cognitive processing. In Hindi, a greater proportion of predicates
are expressed using complex verb phrases; light verb constructions make up more than a quarter of
predicates, more than double that of English or German. Thus, Hindi native speakers have signi-
cantly more experience in processing these constructions, and this practice eect overrides any cost
of co-composition.
Crucially, this interpretation hinges on two assumptions: The rst of these assumptions is that
the English and German data are reliable. While we have not attempted a replication of those data,
we hope to do so in future work, once in-person data collection is possible again. This plan is not
rooted in mistrust, but scientic prudence: In the ten years that have gone by since these data were
collected, statistical methods and conventions on sample size, for instance, have changed, and so
may have usage patterns, with some constructions being now perhaps more or less frequent than
they used to be. Related to this point is a valid discussion on the reliability of the relatively rarely
used interference paradigm, the parametrization of probe timings, and their explanatory power (see
Wittenberg, 2013, for discussion). This evaluation of the paradigm is beyond the current discussion,
but converging evidence from both dual tasks and Event-Related Potentials alleviates this concern
(Piñango et al. 2006, Wittenberg et al. 2014, Wittenberg and Piñango 2011).
Importantly, the present results are decidedly not in line with earlier data, which brings us to
the second assumption our interpretation hinges on: Namely, that in light verb construction, co-
composition is indeed the explanandum. However, within the co-composition assumption, several
variations on the theme are conceivable, and there are also there are several broader theoretical
alternatives to consider. We discuss these in turn.
22 / JSAL volume xx, issue xx xx xx
3.1 Early vs. late co-composition
The model of event co-composition that has been discussed in this paper assumes that the eventive
meaning of the noun is incomplete until the integration of the light verb. This means that co-
composition is not complete until after the processing of the light verb. We can refer to this as ‘late’
co-composition. Another possibility is a noun-driven composition account, which would predict that
the process will be initiated before the verb is comprehended.
This alternative model of event composition has the noun as the sole predicator in the light verb
construction (Grimshaw and Mester 1988, Kearns 1988). According to such an account, the noun
is the primary predicator, while the light verb is merely a theta-marker, supplying an agent role to
the subject of the clause (e.g. Mohan in Figure 1).
From a processing point of view, this set of accounts would predict that the co-composition process
will take place early, i.e. at the noun itself, rather than after the verb is encountered. Although the
noun is not usually a predicating element, in the context of the light verb construction, its eventive
meaning is strongly predictive of the entire predicate. We note that this account does not rule out
co-composition as a phenomenon. Rather it implies that during the real-time processing of light
verb constructions the noun is so strongly predictive of the light verb that there is is a negligible
cost when reaction times are measured at the verb.
Both models predict a process of co-composition, but in processing terms, the noun-driven com-
position account would predict the process to be initiated earlier than the event co-composition
account, and the quality of co-composition would dier. The faster reaction times at the verb in Ex-
periment 4 could indicate that event composition has already taken place at the noun. On the other
hand, if the noun was the only predicating element, then we should have observed faster reaction
times when cloze probabilities for the light condition were higher (as in Experiment 1-3). Hence our
experiments do not strongly support either the late or early account, and could be compatible with
both. We hope to address the time-course of co-composition in future studies.
3.2 Alternatives to event co-composition
In this paper, we have adopted event co-composition as the model to explain the process of light
verb construction interpretation, but there may be alternate mechanisms contributing to the data.
In this section we review some of these explanations.
Aspectual implicatures. Wittenberg and Levy (2017) have shown that the conceptualization of
event duration diers between a simple transitive verb (A kissed B) and a light verb construction
(A gave B a kiss). In English, a computational cost for processing light verb constructions may be
attributed to incorporating these aspectual implicatures such as telicity or volitionality during the
co-composition process. In Hindi, one possibility is that these aspectual implicatures are missing in
light verb constructions, which could result in a reduced computational cost when measured at the
verb.
One reason for this would be the absence of denominal and light verb alternations in Hindi. Only
a small group of nouns in Hindi have both denominal and light verb forms, whereas in English both
forms will co-exist in the language (i.e., a kiss vs. to kiss). In Hindi, the light verb construction is
often the only way to express a certain meaning; no simplex verb alternative exists. Hindi native
speakers must be adept at the process of using the construction to ‘verbalize’ new predicates into
the language. Consequently, the verb in Hindi provides much less semantic content to the light verb
construction as compared to German and English.
We do not have psycholinguistic evidence for the availability of aspectual implicatures in Hindi
light verb constructions, but we do know, from Hindi corpus studies, that predicating nouns in
light verb constructions tend to form semantically coherent groups (Sulger and Vaidya 2014). For
instance, it is possible to form a light verb construction as give a sigh/grunt/cry but not *take a
sigh/grunt/cry: A predicating noun, such as sound emission nouns, will combine with only certain
types of verbs, and not others. This suggests that there may be lexical semantic properties of the
light verb that control combinatorial possibilities. If light verbs in Hindi were simply verbalizers,
Productivity and Argument Sharing / 23
such restrictions should not exist- indeed they would be compatible with any noun. We note that this
is still indirect evidence, and more studies need to examine the availability of these implicatures.
Collocational frequencies. It would also be possible to interpret our results as the eects of
frequency and productivity alone, rather than increased practice with the co-composition process.
This would amount to treating light verb constructions as any other type of multiword expression
with high collocational strength, like verbal idioms or compounds, where prediction is faster due
to storage as a whole construction (also see discussion in Wittenberg 2016). However, this is not
supported by the results of the low cloze experiment (Experiment 4) - there is potentially more than
one light verb that can appear after the noun, and yet, reaction times are faster. Additionally, on the
basis of collocational strength alone, we should have also found faster reaction times in Experiment
1, which we do not. This leaves us to conclude that the process of co-composition does not appear
as a processing cost as it does for English or German. Rather, this cost disappears due to language
specic dierences in Hindi.
Alternative syntactic and semantic congurations. As discussed above, the composite ar-
gument structure of noun and light verb is distinct from that of other verbal predicates, in Hindi
and other languages. Unlike canonical one-to-one mapping between syntactic and semantic argu-
ment structure, the light verb construction needs to combine thematic roles originating from both
noun and light verb. It is possible, however, that this composite argument structure is an erro-
neous assumption (see for discussion e.g. He and Wittenberg 2020, Wittenberg 2016); rather, light
verb constructions may simply have a number of semantic roles that correspond to its syntactic
arguments, where the predicating noun lls a ‘metaphorical’ thematic role slot. Alternatively, the
predicating noun need not ll a semantic role slot at all – it simply forms a single predicate together
with the light verb, and the structure has two semantic roles corresponding with two syntactic ar-
guments. In both these scenarios, there would be no reason to believe that a composite argument
structure is derived as in Figure 1, and therefore we nd no computational cost associated with
processing the light verb construction.
However, there is experimental evidence for co-composition in light verb constructions (again, on
English). In Wittenberg and Snedeker (2014), participants were trained to categorize pictures of
events according to number of thematic roles, for instance, sleeping children into a one-role category;
monkeys eating bananas would be a two-role event; and a child giving an apple to a teaching
would be a three-role event. In the test phase, participants also had to sort sentences containing
light verb constructions containing the verb give (e.g., give a kiss/kick/hug…), base verbs (e.g.,
kiss/kick/hug…), or non-light constructions (e.g., give a ower/plate/ticket…). The predictions were
that if the thematic argument structure of light verb constructions is constructed following surface
syntactic arguments, participants will categorize light verb constructions as three-role events. If
they are understood as stored constructions to describe the same as base verbs, then they would
be categorized as two-role events. Results showed that sentences with light verb constructions fell
between the two categories – they were categorized dierently from two-role events and dierently
from three-role events, suggesting that light verb constructions may be associated with two types of
argument structure simultaneously.
In a followup to this study, Wittenberg et al. (2017) conducted an eye-tracking experiment. Here,
participants implicitly learned to classify two- and three-role sentences, without being instructed
about their valency properties. Participants were able to do this successfully for non-light and
base verb sentences, but when light verb constructions were encountered, they again displayed
an intermediate pattern, again between the two and three-role alternatives, indicating that native
speakers need to resolve two sets of thematic roles coming from the noun and light verb respectively.
None of these studies alone can answer the question of how light verb constructions are learned,
stored, comprehended, and produced. However, all of them together indicate that at least in English
and German, light verb constructions behave dierently from canonical constructions on several
dierent levels; and the evidence suggests that the assembly of the argument structure plays a
crucial role.
24 / JSAL volume xx, issue xx xx xx
3.3 Pairwise comparisons.
Experiments 2 and 3 in our study yielded no dierences between the light and non-light condition.
We expected to nd no dierences in Experiment 2 (based on the English and German results), but
we found a null eect in Experiment 3 as well. In order to explain it, we examined the experiment
design for the previous experiments on light verb constructions.
Interestingly, the cross-modal lexical decision task for English did not nd dierences between
light and non-light when measured 300ms after the verb (Piñango et al. 2006). Rather, they found
a pairwise dierence between the light and heavy condition, i.e. between Mr. Olson gave an order
last night to the produce guy (light) and Mr. Olson typed an order last night for the produce guy
(heavy). There was no dierence between the light and non-light condition at 300ms after the verb.
For German, on the other hand, the light condition did result in slower reaction times than both
heavy (same noun) and non-light constructions 300ms after the verb.
This seems to suggest that there could be other types of pairwise comparisons that are possible,
particularly grammatically plausible ones (such as the heavy condition) rather than the semanti-
cally implausible anomalous condition used in our experiments for Hindi. Perhaps future work can
examine such comparisons in more detail.
4 Conclusion
We presented four studies on comprehending Hindi light verb constructions, compared to their
non-light counterparts, and anomalous sentences. In summary, there appear to be considerable
dierences in the speed of co-composition carried out by Hindi speakers as compared to their English
and German counterparts. Our results imply that Hindi native speakers are adept at the process
of understanding light verb constructions as ‘verbalizing’ predicates, much more so than speakers
of Germanic languages. One potential explanation of these data is that the process of argument
sharing is not universal, but limited to Germanic languages. However, the gist of the theoretical
proposal seems to hold across languages, and was originally developed for languages like Hindi and
Urdu (Butt 2010). Thus, we argue that these data provide evidence for a case of specic linguistic
experiences shaping cognition: Cost disappears with practice.
Acknowledgments
We thank the audiences at SALA 2018 in Konstanz, Germany, and X-PPL 2019 in Zurich, Switzer-
land, for helpful comments and discussion, and gratefully acknowledge Sidharth Ranjan (IIT Delhi)
for creating the custom JsPsych plugin for the presentation of the stimuli. Ashwini Vaidya was sup-
ported by the DST-CSRI (Dept of Science and Technology-Cognitive Science Research Initiative)
post-doctoral fellowship at IIT Delhi (PDF-67/2015).
References
Ahmed, Tafseer, Miriam Butt, Annette Hautli, and Sebastian Sulger. 2012. A reference dependency bank for
analyzing complex predicates. In Proceedings of the Eight International Conference on Language Resources
and Evaluation (LREC’12).http://www.lrec-conf.org/proceedings/lrec2012/pdf/474_Paper.pdf.
Alsina, Alex, Joan Bresnan, and Peter Sells. 1997. Complex Predicates: Structure and Theory. In Complex
Predicates. CSLI Publications, Stanford. https://trove.nla.gov.au/version/17082409.
Boland, Julie E. 1997. The relationship between syntactic and semantic processes in sentence comprehension.
Language and Cognitive Processes 12(4):423–484.
Briem, Daniela, Britta Balliel, Brigitte Rockstroh, Miriam Butt, Sabine Schulte im Walde, and Ramin
Assadollahi. 2009. Distinct processing of function verb categories in the human brain. Brain Research
1249:173–180. https://doi.org/10.1016/j.brainres.2008.10.027.
Butt, Miriam. 2010. The Light Verb Jungle: Still Hacking Away. In M. Amberber, M. Harvey, and B. Baker,
eds., Complex Predicates in Cross-Linguistic Perspective, pages 48–78. Cambridge University Press. https:
//doi.org/10.1017/CBO9780511712234.004.
Productivity and Argument Sharing / 25
Butt, Miriam, Tracy Holloway King, and Gillian Ramchand. 2008. Complex Predication:How Did the Child
Pinch the Elephant? In L. Uyechi and L. Wee, eds., Reality Exploration and Discovery: Pattern Interaction
in Language and Life. CSLI Publications, Stanford. http://ling.uni-konstanz.de/pages/home/butt/
main/papers/mo-final.pdf.
Cuetos, Fernando, Don C. Mitchell, and Martin M.B. Corley. 1996. Parsing in dierent languages. In
M. Carreiras, J. Garcia-Albea, and N. Sabastian-Galles, eds., Language Processing in Spanish. Erlbaum.
https://www.taylorfrancis.com/books/e/9780203773970/chapters/10.4324/9780203773970-11.
Culicover, Peter W and Ray Jackendo. 2005. Simpler syntax. Oxford University Press. https://global.
oup.com/academic/product/simpler-syntax-9780199271092?cc=us&lang=en&#.
Davison, Alice. 2005. Phrasal predicates: How N combines with V in Hindi/Urdu. In T. Bhattacharya, ed.,
Yearbook of South Asian Languages and Linguistics, pages 83–116. Mouton de Gruyter. http://www.
uiowa.edu/~linguist/faculty/davison/phrasal.pdf.
de Leeuw, J. R. 2015. jsPsych: A javascript library for creating behavioral experiments in a web browser.
Behavior Research Methods 47(1):1–12. https://doi.org/10.3758/s13428-014-0458-y.
Drummond, Alex. 2007. Ibex farm. https://spellout.net/ibexfarm/. accessed August 8, 2020.
Durie, Mark. 1988. Verb serialization and “verbal-prepositions” in oceanic languages. Oceanic linguistics
27(1):1. https://www.jstor.org/stable/3623147.
Grillo, Nino and Joäo Costa. 2014. A novel argument for the universality of parsing principles. Cognition
133(1):156 – 187. https://doi.org/10.1016/j.cognition.2014.05.019.
Grimshaw, Jane and Armin Mester. 1988. Light verbs and theta-marking. Linguistic Inquiry 9(2):205–232.
https://www.jstor.org/stable/4178587.
He, Angela Xiaoxue and Eva Wittenberg. 2020. The acquisition of event nominals and light verb construc-
tions. Language and Linguistics Compass 14(2):e12363. https://onlinelibrary.wiley.com/doi/abs/
10.1111/lnc3.12363.
Hook, Peter. 1974. The Compound Verb in Hindi . University of Michigan, Ann Arbor.
Jespersen, Otto. 1965. A Modern English Grammar on Historical Principles, Part VI, Morphology. George
Allen and Unwin Ltd.
Kachru, Yamuna. 2006. Hindi. John Benjamins.
Kamienkowski, Juan E, Harold Pashler, Stanislas Dehaene, and Mariano Sigman. 2011. Eects of practice on
task architecture: Combined evidence from interference experiments and random-walk models of decision
making. Cognition 119(1):81–95. https://doi.org/10.1016/j.cognition.2010.12.010.
Kearns, Kate. 1988. Light verbs in English. Manuscript, MIT (revised 2002), http://citeseerx.ist.psu.
edu/viewdoc/summary?doi=10.1.1.132.29.
Kennedy, A., J. Pynte, W.S. Murray, and S.A. Paul. 2013. Frequency and predictability eects in the
dundee corpus: an eye movement analysis. Quarterly Journal of Experimental Psychology 66(3):601–618.
https://doi.org/10.1080/17470218.2012.676054.
Kilgarri, Adam, Siva Reddy, Jan Pomikálek, and Avinesh PVS. 2010. A Corpus Factory for Many Lan-
guages. In Proceedings of the Seventh Conference on International Language Resources and Evaluation
(LREC’10).http://www.lrec-conf.org/proceedings/lrec2010/pdf/79_Paper.pdf.
Kuperberg, Gina R. 2013. The proactive comprehender: What event-related potentials tell us about the
dynamics of reading comprehension. Unraveling the behavioral, neurobiological, and genetic components
of reading comprehension pages 176–192.
Kuperberg, Gina R, Trevor Brothers, and Edward W Wlotko. 2020. A tale of two positivities and the
n400: Distinct neural signatures are evoked by conrmed and violated predictions at dierent levels of
representation. Journal of Cognitive Neuroscience 32(1):12–35.
Masica, Colin. 1993. The Indo Aryan Languages. Cambridge University Press. https:
//www.cambridge.org/in/academic/subjects/languages-linguistics/other-languages-and-
linguistics/indo-aryan-languages?format=PB&isbn=9780521299442.
McElree, Brian and Teresa Grith. 1995. Syntactic and thematic processing in sentence comprehension: Ev-
idence for a temporal dissociation. Journal of Experimental Psychology: Learning, Memory, and Cognition
21(1):134.
Mitchell, Don, Fernando Cuetos, Martin Corley, and Marc Brysbaert. 1995. Exposure-based models of human
parsing: Evidence for the use of coarse-grained (nonlexical) statistical records. Journal of Psycholinguistic
Research 24:469–488. https://link.springer.com/article/10.1007/BF02143162.
26 / JSAL volume xx, issue xx xx xx
Mitchell, Don C. 2004. On-line methods in language processing: Introduction and historical review. In
M. Carreiras and C. C. Jr., eds., The on-line study of sentence comprehension: Eyetracking, ERPs
and beyond, pages 15–32. New York, NY: Psychology Press. https://www.taylorfrancis.com/books/
9780203509050.
Mohanan, Tara. 1997. Multidimensionality of representation- NV complex predicates in Hindi. In A. Alsina,
J. Bresnan, and P. Sells, eds., Complex Predicates. CSLI Publications, Stanford.
Norclie, Elisabeth, Alice C. Harris, and T. Florian Jaeger. 2015. Cross-linguistic psycholinguistics and
its critical role in theory development: early beginnings and recent advances. Language, Cognition and
Neuroscience 30(9):1009–1032. https://doi.org/10.1080/23273798.2015.1080373.
Piñango, Maria Mercedes, J. Mack, and Ray Jackendo. 2006. Semantic combinatorial processes in argument
structure: Evidence from light-verbs. In Proceedings of Berkeley Linguistics Society 32nd Annual Meeting.
http://dx.doi.org/10.3765/bls.v32i1.3468.
Reddy, Siva and Serge Sharo. 2011. Cross Language POS Taggers (and other Tools) for Indian Languages:
An Experiment with Kannada using Telugu Resources. In Proceedings of the Fifth International Workshop
On Cross Lingual Information Access, pages 11–19. Chiang Mai, Thailand: Asian Federation of Natural
Language Processing. https://www.aclweb.org/anthology/W11-3603.
Sadeghi, Ali Ashraf. 1993. On denominative verbs in Persian. In Farsi Language and the Language of
Science, pages 236–246. Tehran: University Press.
Sag, Ivan A., Timothy Baldwin, Francis Bond, Ann Copestake, and Dan Flickinger. 2002. Multiword Ex-
pressions: A Pain in the neck for NLP. In Proceedings of the 3rd International Conference on Intelligent
Text Processing and Computational Linguistics (CICLing’02), pages 1–15. https://link.springer.com/
chapter/10.1007/3-540-45715-1_1.
Seiss, Melanie. 2009. On the dierence between auxiliaries, serial verbs and light verbs. In Proceedings of
the LFG09 Conference.https://web.stanford.edu/group/cslipublications/cslipublications/LFG/
14/papers/lfg09seiss.pdf.
Staub, Adrian. 2011. The eect of lexical predictability on distributions of eye xation durations. Psycho-
nomic Bulletin and Review 18(2):371–6. https://doi.org/10.3758/s13423-010-0046-9.
Sulger, Sebastian and Ashwini Vaidya. 2014. Towards Identifying Hindi/Urdu Noun Templates in Support
of a Large-Scale LFG Grammar. In Proceedings of the Fifth Workshop on South and Southeast Asian
Natural Language Processing at COLING 2014.https://www.aclweb.org/anthology/W14-5501.pdf.
Tu, Yuancheng and Dan Roth. 2011. Learning English Light Verb Constructions: Contextual or Statistical.
In Proceedings of the Workshop on Multiword Expressions (MWE 2011), 49th Annual Meeting of the
Association for Computational Linguistics: Human Language Technologies (ACL-HLT 2011).https:
//www.aclweb.org/anthology/W11-0807.
Vaidya, Ashwini, Martha Palmer, and Bhuvana Narasimhan. 2013. Semantic roles for nominal predicates:
Building a lexical resource. In Proceedings of the 9th Workshop on Multi-word Expressions, NAACL-13.
https://www.aclweb.org/anthology/W13-1018.
Vincze, Veronika, István Nagy T, and Gabór Berend. 2011. Detecting noun compounds and light verb
constructions: a contrastive study. In Proceedings of the Workshop on Multiword Expressions: from Parsing
and Generation to the Real World (MWE 2011).https://www.aclweb.org/anthology/W11-0817/.
Wittenberg, Eva. 2013. Paradigmenspezische eekte subtiler semantischer manipulationen. Linguis-
tische Berichte 2013(235):293–308. https://buske.de/paradigmenspezifische-effekte-subtiler-
semantischer-manipulationen.html.
Wittenberg, Eva. 2016. With light verb constructions from syntax to concepts, vol. 7. Universitätsverlag Pots-
dam. https://publishup.uni-potsdam.de/opus4-ubp/frontdoor/deliver/index/docId/8236/file/
pcss07.pdf.
Wittenberg, Eva, Manizeh Khan, and Jesse Snedeker. 2017. Investigating thematic roles through im-
plicit learning: Evidence from light verb constructions. Frontiers in Psychology 8:1089. https://www.
frontiersin.org/article/10.3389/fpsyg.2017.01089.
Wittenberg, Eva and Roger Levy. 2017. If you want a quick kiss, make it count: How choice of syntactic
construction aects event construal. Journal of Memory and Language 94:254 – 271. http://www.
sciencedirect.com/science/article/pii/S0749596X16302480.
Productivity and Argument Sharing / 27
Wittenberg, Eva, Martin Paczynski, Heike Wiese, Ray Jackendo, and Gina Kuperberg. 2014. The dierence
between “giving a rose” and “giving a kiss”: Sustained neural activity to the light verb construction.
Journal of Memory and Language 73:31–42. https://doi.org/10.1016/j.jml.2014.02.002.
Wittenberg, Eva and Maria Mercedes Piñango. 2011. Processing light verb constructions. The Mental
Lexicon 6(3):393–413. http://dx.doi.org/10.1075/ml.6.3.03wit.
Wittenberg, Eva and Jesse Snedeker. 2014. It takes two to kiss, but does it take three to give a kiss?
categorization based on thematic roles. Language, Cognition and Neuroscience 29(5):635–641. https:
//doi.org/10.1080/01690965.2013.831918.
28 / JSAL volume xx, issue xx xx xx
Appendix
Region Context Subject Object Nominal
modier Noun Verb Post-verb-1 Post-verb-2 Sent end
Light vs. Nonlight t-value 1.53 -0.10 -1.43 0.65 -0.76 0.45 -0.56 -0.92 -0.14
p-value 0.12 0.91 0.15 0.51 0.44 0.65 0.57 0.35 0.89
Anomalous vs. Nonlight t-value 0.24 -1.28 -1.24 0.32 1.0 4.35 2.73 0.48 0.75
p-value 0.8 0.19 0.21 0.74 0.32 0.00001* 0.006 0.63 0.45
TABLE 4: Region-wise results for the entire sentence showing t-values (top row) and p-values (bottom
row) in Experiment 1. The critical region is at the verb.
Region Context Subject Object Nominal
modier Noun Verb Post-verb-1 Post-verb-2 Sent end
Light
Mean RT 714.12 837.92 752.61 641.66 753.38 724.82 591.28 580.27 593.32
SD 904.04 874.07 539.24 574.82 802.61 842.96 509.89 360.2 718.02
Non Light
Mean RT 706.26 838.85 801.85 607.39 709.97 676.71 566.36 573.79 569.59
SD 1820.22 945.37 809.12 500.8 480.26 577.13 406.14 344.25 615.56
Anomalous
Mean RT 745.14 785.58 790.86 616.92 783.61 813.32 637.5 587.79 584.82
SD 2047.39 678.09 767.12 562.57 677.58 784.26 533.13 325.04 564.32
TABLE 5: Region-wise Mean Reaction times and standard deviations for each condition in Experiment
1. The critical region is at the verb.
1 Items in Experiment 1,2 and 3
1. अपने समय का बंधन करना मुिकल है, इसीलए अयापक ने वाथ को भाषण/कै लंडर/*सलसला दया और कु छ आसान
उपाय भी बताये It is dicult to manage one’s time, hence the teacher gave the student a
speech/calendar/*happening and mentioned a few tips. PROBE: गाय ; cow
2. अपनी भूल को कबूल करते ए, लंडन के अधकार ने टश अबेसेडर को अपना इतीफा/पासपोट/*अफ़सोस दया लेकन वे
फर भी नाराज़ थे While accepting his mistake, the ocer in London gave the British ambassador
his resignation/passport/*regret but he was still upset. PROBE: चाँद; moon
3. समलेन के बाद ोफे सर बासु ने अपने वाथय को भारत मण करने का वचन/नशा/*आज़ाद दया और साथ म उनको कसे
सुनाये After the conference, Professor Basu gave his students a promise/map/*freedom and
recounted a few anecdotes. PROBE: चोर; thief
4. दवाली के दन मीरा ने एक गरब को बाजार म इशारा/पोशाक/*फायदा दया और उसको अपने बगल म बैठाया On the day
of Diwali Meera gave a poor man a sign/clothing/*advantage and made him sit beside her.
PROBE:साँप; snake
5. मुंबई के टूडयो म, मशर गायक रायेयाम ने जय को एक मौका/एबम/*अपराध दया और जय को बड़ ख़ुशी ई In the
Mumbai studio the famous singer Radheshyam gave Jay a chance/album/*oence and Jay was
very happy. PROBE: धूल; dust
6. इस फम म भगवान ने करना के सपने म उसे दशन/उपयास/*तापमान दया लेकन उठने के बाद वह सपना समझ नह पायी In
this lm, God gave an audience/novel/*weather to Kareena in her dream, but on waking up
she could not understand the dream. PROBE: जाल; web
7. ऊपर से ऑडर आने के बाद सीनयर इंजीनयर ने अपने वभाग को ज बनाने का नदश/डजाइन/*काशन दया और मीडया के
लए ेस कां स क योजना बनाई After getting an order from his superiors, the senior engineer gave
his department an order/design/*publication and made arrangements for a press conference.
PROBE: नाक; nose
8. वायरस के फै लने के बाद इस संथा ने गरब लोग को अपताल के बाहर जानकार/जगह/घोषणा दी और उनको बड़ राहत
मली After the virus began to spread, this organization gave the poor people informa-
tion/space/*announcement outside the hospital and they were very relieved. PROBE: तंबाकू ;
tobacco
Productivity and Argument Sharing / 29
9. उसे मोशन मलने पर दादी ने रोहन को बधाई/बाइक/*पसद दी और पूरे बिडंग को दावत भी दी After he got his
promotion, grandma gave Rohan compliments/motorcycle/*liking and threw a party for the
whole building. PROBE: कु स; chair
10. आज अख़बार म लखा था क उस ांतकार गुट ने ही लोग को उर/हथयार/*वहार दया और उसके बाद वे जंगल
म वापस चले गए Today the newspapers reported that the revolutionary group gave an an-
swer/weapon/*behaviour to the people and then they went back into the jungle. PROBE: रेत;
sand
11. एक नत समय पर रपोटर ने बाहर खड़े ए सहायक को संके त/प/*चचा दया और उसके बाद उसे कु छ समझाने लगा At a
particular time, the reporter gave a signal/letter/*argument to his aide standing outside and
then began to explain something to him: PROBE: जूता; shoe
12. के ट क ेनग के लए बाबूजी ने रोहन को टेडयम म खेलने के लए ोसाहन/कराया/*तारफ़ दया और अगले महीने उसका खेल
देखने का वादा भी कया In order to train for cricket, Babuji gave Rohan encouragement/fare/*praise
and promised to come see him play next month. PROBE: कमल; lotus
13. आज के कायम म धानमंी ने नए अय को कृ ष सेवा का उदहारण/वभाग/यास दया और कृ ष पर इस सरकार का बत
ज़ोर है In today’s event, the prime minister gave the new ocer a(n) example/division/*eort
of the agriculture ministry as this area is important to him. PROBE: कपड़े; clothes
14. उस गाँव के छोटेसे मंदर म पंडतजी ने अभय को आासन/साद/आकलन दया यक अभय अपनी परा के लए काफ परेशान
था In the small temple in the village the priest gave Abhay reassurance/oerings/*assessment
because Abhay was worreid about his exams. PROBE: सड़क; street
15. उस ज़ले म लोक कलाकार ने इस आंदोलन को अपना समथन/संगीत/*संबंध दया और आंदोलन ने दो महीने से काफ तेज़ी पकड़
ली है In that district, the folk artists have given this campaign their support/music/*connection
and the campaign has intensied in the last two months. PROBE: दाल; lentil
2 Items in Experiment 4
1. कोचग लासेज बत पैसे लेते ह, लेकन कोचग वाल का यह फायदा है क वे इस साल क परा म आनेवाले सवाल का
अंदाजा/पेपर देते है, जससे अछा अयास हो सकता है. Coaching classes charge a lot of money, but one
advantage of these classes is that they give a hint/test about the exam questions for this year
and this results in better preparation. PROBE: चाँद ; moon
2. सुमन के परवार वाल को कसी भी हालत म बेटा चाहए था, और डलीवर के बाद अपताल कमचारय ने बड़ हमत करके
सुमन को सुरक् /बेटी षा दी और कसी को खबर नह क. Suman’s family wanted a boy at any cost and after
her delivery the hospital workers courageously gave Suman protection/a girl and made sure
nobody found out. PROBE: साँप; snake.
3. सारंग गैलर' म सुमन के कौशय को देखकर यूरोपयन आस कसल ने उसे बलन आने का आमंण/ऑशन दया जहाँ वह
अपनी कला को आगे बड़ा पायेगी. . After being impressed with Suman’s work in the Sarang Gallery,
the European Arts Council gave her an invitation/option to come to Berlin, where she could
progress with her talent. PROBE: चोर ; thief
4. २१वी शतादी म भारत क सफलता को नजर म रखते ए यू.न को भारतवासय को एक नया दजा/वन देना है जसका हम अब
एहसास हो रहा है. In the twenty-rst century, keeping India’s achievements in mind, the U.N.
should give Indians a new status/dream, which we are only now beginning to realize. PROBE:
धूल; dust.
5. पछले कई साल से भारत म इतहास शोध क हालत देखते ए, भारत सरकार क नयी योजना के ारा पैस से इतहासकार
को मान/आकषण दया है और यह बात काफ अनपेत है. Despite the sorry state of historical research in
India over the last few years, the Indian government’s new scheme of monetary rewards gave
historians recognition/inducement, which was quite unexpected. PROBE: तंबाकू ;tobacco.
6. उराखंड के देहरान शहर के पास 'वूडलड' कू ल म सोनू के ायापक ने उसके माँ-बाप को छावृ के फॉम के लए थोड़
तकलीफ/फस दी और उह दो दन के लए वही काकर रखा था. Near the city of Dehradun in Uttarakhand in
Woodland school, Sonu’s teacher gave his parents some trouble/fees and made them stay for a
couple more days there. PROBE: जाल ; web.
7. इस इलाके के लोग ने सभी अख़बार म पाक म पेड़ कटने के बारे म एक टपणी लखी और पाक म हरयाली बचाने के लए
अधकारय को जोर/व दया यूंक यह अवैध प से हो रहा था. The people in this area wrote an article in
30 / JSAL volume xx, issue xx xx xx
the papers about trees getting cut in the park and they gave the authorities a bother/deadline
about the illegal nature of this work. PROBE: कु स ; chair
8. परा के दन शहर म तेज़ बारश ई, और देर से आने वाले छा के लए कॉलेज ने वशेष रयायत देकर, उनको राहत/का दी
जससे वे परा पूर कर पाये. On the day of exams, it rained heavily in the city and for the students
who arrived late, the college made arrangements and gave them assistance/(a) classroom where
they could complete their exams. PROBE: गाय; cow.
9. मेडकल शोधकता ने अपने शोध के जरये एस जैसी गंभीर बीमार के खलाफ लड़ने के लए एक नया सुझाव/हथयार दया
है जससे मरज को बत फायदा हो सकता है. The discoveries made in medical science have given those
suering from serious diseases like AIDS a new possibility/weapon which will surely benet
them a great deal. PROBE: रेत; sand.
10. उस े के सभी जल म दंगा-फसाद के कारण लोग अपनी खेती बार छोड़कर भाग गए लेकन सरकार ने उन लोग को न कोई
सुरा/भूम दी न उह मुआवज़ा मला. In all the districts in this area people have left their homes and
elds due to the riots but the government has neither given any protection/land to these people
nor have they received any other compensation. PROBE: कमल;lotus.
11. मनु चाचा को कू ल म पढ़ाना पसंद था, फर भी उहने नया वसाय शु करने के लए योतष बाबा को अपनी सहमत/कुं डली
दी और अपना इतीफ़ा लख दया. Uncle Manu liked to teach in school but in order to start a new
business, he gave the astrologer his consent/horoscope and wrote his resignation. PROBE: जूता;
shoe.
12. हम घर के बाहर गाड़ का इंतजार कर रहे थे तभी रामदासजी नजर आये और उहने हम नमल क शादी का यौता/तोहफा दया जो
एक बड़े लाल और पले डबे म रखा था. We were waiting outside the house for the car, when Ramdasji
came into view and he gave us an invitation/gift for Nirmal’s wedding which was kept in a
large red and yellow box. PROBE: दाल; lentil.
13. पोस मनी क ऑडर मलने के बाद टीम के कोच जोशी जी ने नए हॉक टीम को आकार/कप दया जससे बत लोग को
ेरणा मली है. After receiving the order from the Sports Ministry, the team coach Joshi gave the
new hockey team a shape/cup which has given a lot of people hope. PROBE: सड़क; road.
14. दस साल के बाद इस महीने पहली बार पंचायत क भेट ई, जसम ाम पंचायत के अलग-अलग काय के लए पंचायत मुख ने एक
करोड़ क पूंजी नवेश क सफारश/फाइल दी और आगे के काम क योजना बनाई. After 10 years, the village council
met this month, where the council chief gave a recommendation/le for one crore rupees to
take care of various activities and make a plan for the days ahead. PROBE: कपड़े; clothes.
15. कृ ष ववालय से डी ात करने के बाद ,महेशजी रामपुर गाँव गए और वहाँ उहने जैवक खेती करने के लए गाँववाल
को बढावा/चैलज दया और अपने चाचा क ज़मीन पर अपने आप योग करने लगे. After obtaining a degree from
the agricultural university, Maheshji went to Rampur and gave the villagers motivation/(a)
challenge to farm organically and began to experiment on his uncle’s land himself. PROBE:
नाक; nose.
