Transcript Document

Constructional Profiles as the
Basis of Semantic Analysis
Suzanne Kemmer
Rice University
[email protected]
Introduction

Construction Grammar defines constructions as
linguistic units that necessarily have some noncompositional semantics

Constructions have some aspect of meaning that is
not reducible to (or “predictable from”) its component
parts (or other constructions)

And, constructions are argued to be necessary as a
construct in any theory of grammar.
Why do we need constructions?
 One argument for the indispensability
of constructions in grammatical
analysis comes from Coercion Effects
(Michaelis 2002)
Constructions and Coercion
 Constructions explain how certain expected
semantic anomalies do not materialize.
 Give me some pillow!
 Elements in some sentences are expected
to clash by virtue of their incompatible
semantics, based on their distribution
outside the construction.
Constructions override lexical meaning
 Constructions fill in semantic substance and
overcome semantic incompatibility of
component parts through override
 I slept my way across the Atlantic.
 Sleep -- lack of motion specification
 Sentence as a whole -- describes motion with
concomitant sleeping
 pit the cherries, dust the furniture, bone
the filet
 conventionalized semantic elements added:
motion, directionality
Investigating constructional semantics


We can investigate the semantics of constructions in various
ways. Most relevant here:
1. Observation of distributional properties at sentence level;
contrasts




Syntactic properties are diagnostics or clues to semantics (as
per assumption of close nature of the syntax-semantics
relation of CL) via acceptability patterns in arrays of minimally
contrasting examples.
Semantics can be investigated by observing lexical items in
mainly clause level contexts and observing anomalies and
compatibilities that make such utterances less or more
acceptable.
Lakoff (1987) and Langacker (1987, 1991) inter alia analyze
constructions (also lexical items) using this methodology.
2. Observation of distributions of recurring elements in a
construction in large samples of language use
Investigating constructional semantics
 We should employ any methodologies that
prove useful
 Ideal: convergence of multiple sources of
evidence gathered via different methodologies
 Second method gives insight into some
semantic properties otherwise inaccessible
(on assumption that frequency is a
reflection of degree of entrenchment, which
is itself a part of the system).
What else can we get?
 Observing constructions also
contributes to an understanding of
the important mechanism of coercion
 How does it work?
 How and to what extent can it forestall
anomaly?
Who uses corpora?

Who uses large corpora for the cognitive semantic analysis of
constructions? Consider:

British, Scandinavian, and American schools:


Sinclair, Stubbs, Stig Johanson, Hunston and Francis, Biber.
Corpora yes, cognitive semantic analysis of constructions, no or
minimal
Construction Grammarians:
 Fillmore, Kay, Michaelis, Croft, Goldberg, and their students
Corpora, no or not primarily; semantic analysis is largely done by the
first method above
Corpus-construction grammarians:


Boas, Fried, Gries, Lambrecht, Michaelis, Östman, Stefanowitsch,
and students.
Corpora yes, cognitive semantic analyses, yes


Barlow, Geeraerts, Kemmer, Verhagen and students
Corpora yes, cognitive semantic analysis yes


Corpus-cognitive linguists:
Results and ongoing research
 The Corpus-CG and Corpus-CL groups have
in common:
 Bottom-up analyses of particular constructions.
Visual inspection for patterns via keywords in
context and sorting; typically statistical data
 Analysis of constructional semantics, identification
of theoretical issues in CL/CG
 Inclusion of units at varying levels of specificity
See: Rohde 2001; Boas 2002, Gries and Stefanowitsch 2002;
Stefanowitsch and Gries 2003, Fried 2004; studies
appearing in Achard and Kemmer 2004 and Oestman and
Fried 2005; Hilpert (to appear a, to appear b); Taylor (to
appear)
Diachrony
 Several diachronic studies in Cognitive
Linguistics and, more recently,Construction
Grammar, using corpora
 Carey 1994, Israel 1995, Ziegler 2002, Diewald,
Gilquain 2004, Kemmer and Hilpert (forthcoming)
Two perspectives
 Focus on individual lexemes



many analyses of 1980s and 1990s; still current
polysemy networks containing linked senses of
prepositions, cases, etc.
Kemmer (1993) added links to conceptually neighboring
concepts in a multidimensional semantic/conceptual
space (constructions not foregrounded)
 Focus on the construction as the unit of observation

Integration of lexical semantic information with
constructional specifications
Item-based vs. construction-based
 Lexical perspective continues a venerable
tradition of focus on words
 Words (more precisely: lexical roots) are
certainly salient cognitive units
 Speakers access word meaning much more
easily than meaning of linguistic units of
other sizes or greater schematicity
 sublexical morphemes, constructions, sentences
 BUT: there are some very good reasons for
the switch to viewing linguistic knowledge
from a constructional perspective.
Motivating the constructional view

Traditional polysemy networks, although focusing on lexical
items, actually require observation of lexemes in larger
chunks of language


The standard methodology for distinguishing word senses is
to posit different linguistic contexts in which the senses
become clearly disambiguated



Linked senses are structured by relative closeness of
senses as determined by the closeness of the relations of
the contextualized examples (representing types of uses)
Example: the ring on my finger vs. he was knocked right
out of the ring
the desired sense is the only one compatible with the
surrounding material.
Thus: Distribution of lexical items in larger structures is
how we can tell the range of senses/uses of a lexical item
Senses often correlate with
constructions

For relational units, the various senses of a given lexical unit are
actually associated with constructions.

Static vs. dynamic senses of the English prepositions are parts of
constructions that express static and dynamic spatial senses



over (the hill) : hovered over the hill, flew over the hill
Prep NP : V-static over NP, V-motion over NP
Senses of verbs are in most cases correlated with different
constructions (e.g. roll (tr., caused motion) vs. roll (intr., autonomous
motion)
Following Goldberg (1995), Construction Grammarians reject
positing ‘extra’ verb senses
 an extra sense for sneeze in she sneezed the napkin off the
table seems absurd
Bidirectional links

But Langacker (2003) points out: there are many verbs that have
a strong associative link to a particular construction.





give is extremely frequent in the ditransitive compared to other verbs
The ditransitive construction is extremely frequent with give.
The usage-based model predicts, based on frequency, that there is a
highly conventionalized link to the ditransitive that is part of our
knowledge of give.
If so, give is an access point to the ditransitive construction and its
associated frame
Hypothesis: The links between lexical item and constructions reach
in both directions

We posit both nodes as units if both are conventionalized. Give may
activate the ditransitive just as the construction primes the word.
The usage-based model
 In the usage-based model, links in a
linguistic knowledge network are viewed as
activation pathways with potentially
bidirectional activation flows (cf. Lamb 2000)
 Predicts that strongly entrenched links could
potentially go in either direction.
 Converges with findings from neurology suggesting
that links between neurons and between cortical
columns have physically distinct pathways that can
have differential activation strength.
Predictions
 Viewing the links as potentially bidirectional
makes some predictions:
 It should be possible for there to be dissociation
between the relative strengths of an activation
path directed from a lexical item to a construction,
and another in the opposite direction.
 Evidence from distributional studies show that
some individual verbs occur very frequently in a
particular construction, like wriggle in the way
construction (the baby wriggled its way out of the
playpen) and not at all often, relatively speaking,
outside the construction. Other examples in Gries
and Stefanowitsch 2005.
The usage-based model

Conversely, a verb like make occurs frequently in the construction,
but only as a function of its high overall frequency.


Thus, affinity to the way construction not as tight
Prediction: wriggle should have a stronger priming effect on the
construction than make.


Psycholinguistic evidence: Goldberg 2004 for give and ditransitive

Still, the way construction does have a strong link to make.

The construction allows a wide range of verbs, but given no special
semantic properties the speaker desires to convey, the speaker is likely
to choose make over other possibilities because s/he has a great deal
of experience of that choice, which effectively increases, by repeated
memory, the likelihood of activation of make --unless there is some
overriding reason to make another choice (desire to express manner of
motion, for example).
I think of this as ‘long-term priming’.
Prediction:

More on preference for constructions:
Constructions and events




Argument structure constructions are used to form more
event-sized conceptualizations than single words.
Such constructions are likely to be rather crucial units relied
on by speakers to make an intial chunking of reality into a
manageable and manipulable portion of conceptual
structure, as suggested by Talmy 2000.
If this is true, then starting from the construction and
investigating it with regard to what smaller units occur in it
habitually/conventionally is more likely to result in an
analysis that is something like what the speaker knows
about how words are used.
The speaker’s main experience with word usage, in fact,
even from the earliest childhood, is in the context of larger
utterances that the adult uses before child can even talk:
generally speaking, constructions.
Constructions and events

The constructional perspective puts us into the realm of meaning
which is much more like the kinds of meanings that speakers deal
more normally with in language use.





It is difficult for speakers to define either words or constructions
When they do try to define words, they have to activate memories of
contexts similar to those in which they heard the word used, i.e. they
have to try to come up with an imagined context similar to one they
experienced, or an actual remembered context for the usage.
Novice linguists sometimes try to imagine a situational context without
a linguistic context, which gives them much more trouble coming up
with accurate usage generalizations.
They have to be trained to think of specific linguistic utterances that
will allow them to more precisely explore the word in something more
like the typical contexts in which they have heard it, and in which it
contributes to an overall semantically integrated conceptualization.
It is an empirical fact that speakers most typically process
constructions in usage, and not isolated words, given that most
conversation occurs in construction-like chunks and not in isolated
words.
Constructions and processing

Constructions have more utility to speakers for choosing
compatible lexical items to use with the construction than vice
versa.


In the process of arguing for the necessity of constructions in
grammatical theory, some important empirical findings about
constructions have emerged (Goldberg 2004):




But: by hypothesis, lexical elements which hardly occur outside the
construction can trigger it, so that the effects might go the other way.
Speakers use constructions to interpret lexical items
Argument structure constructions aid children to learn new words
These findings make perfect sense if we accept the view that
constructions are extremely important in language processing,
even if they are below the level of consciousness (like most
grammatical units).
For purposes of guiding the precise activation links that will allow
instantaneous accessing of appropriate lexical items,
constructions, on this view, would be crucial processing units.
Rohde 2001
 Study of the relations between lexical items and
constructions that used substantial corpora as a basis
for drawing cognitive semantic conclusions about
those relations
 Rohde investigated a large range of English motion
verbs and their co-occurrence patterns with a range of
prepositions in a [motion verb + preposition + NP]
schematic construction frame.
 Found strong distributional correlations between
motion verbs and particular prepositions; and in fact
with particular senses of the prepositions, essentially
corresponding to particular constructions.
 Results showed that particular verbs have affinities for
particular prepositions and vice versa (and--links have
different strengths as measured by frequency.)
Rohde 2001




Patterns of affinity were also found at a more general level,
involving verb classes defined by particular semantic properties.
Affinity patterns reflect entrenched semantic properties of
particular units
Heightened compatibility makes the lexical items more likely to be
used in the constructions; reduced compatibility reduces the
frequency of the less compatible item.
Compatibility has a strong effect on frequency.


For example, the verb escape occurs extremely frequently with source
prepositions like from. Most other motion verbs in English prefer goal
prepositions, with varying degrees of preference. Rohde concluded
that a source (rather than goal) image schema is strongly
conventionally linked with the lexical item escape.
We can also take the constructional point of view of the same
phenomenon and say that the semi-specified construction [X
escape PREP N] is strongly conventionally linked with source
prepositions, whereas a similar construction with go is linked with
goal prepositions.
Coercion in action


Evidence for the key role of constructions in constructing
meaning comes from acceptability judgments.
It has been demonstrated that constructions coerce lexical
items’ interpretations (Michaelis 2002 and refs.), whereas the
opposite has not been demonstrated.
Example of semantic incompatibility:
I walked into the room.
*I squinted into the room.
[supposed to be bad—and is if one is thinking of the caused
motion construction]
Coercion

Now consider:
(1)
(2)
(3)
(4)
(5)
She
She
She
She
She
looked into the room.
squinted into the room.
squinted through the window.
squinted through her glasses.
peered through her glasses.
(1) expresses an event of directed vision. (2) has a context
involving a somewhat similar visual event for squint. It’s
better than it was on previous slide.

Subsequent sentences get better and better. Why?
Coercion explained




The frame semantics of squinting are associated with a
typical purpose of squinting; to make ones vision better
temporarily by deforming the cornea with the surrounding
eye muscles.
Squinting, then, often occurs to improve vision that is
somehow impaired, possibly in part due to a barrier to
vision such as intervening dirty window glass. The more
overtly we add a barrier to visibility to the context, the
more we improve semantic compatibility and thus the
better the sentence sounds.
Claim: these sorts of acceptability judgments tap into some
speaker knowledge relating knowledge of lexical items to
constructions. (That knowledge itself was originally derived from
experience using those lexical items in larger constructions).
If true, we would expect that speaker knowledge to be also
reflected in frequency patterns. In fact we predict the following:
Predictions (requiring testing)
 An increasing degree of semantic ‘distance’ of a
lexical item (i.e. it contains some incompatibility)
from a conventional construction type will


increase unacceptability of the particular lexical item in
the construction
increase need for additional linguistic context with
additional elements that highlight some compatible
semantic feature of the target item, if that item is to be
produced or interpreted as acceptable.
This is the implicit basis of many linguistic arguments for
semantics of particular elements.
More predictions
 If an utterance is ambiguous, increasing
distance/incompatibility should decrease the likelihood
that the reading with greater semantic incompatibility
will be chosen.
Decrease likelihood of production (compared with more
compatible units)
 Decrease frequency of production (compared with more
compatible units)
Extra contextual information (linguistic or just situational)
needed for production.




I gave it a little kick.
I gave the lever (on the handlebar) a little kick with my
thumb.
Problems
 How to quantify, or at least operationalize how to
judge, degree of semantic similarity/difference,
closeness/distance, compatibility/incompatibility.
 How to corroborate: Linguists interested in semantics
can usually come up with an analysis that is intuitively
satisfying, but little psycholinguistic experimentation
has been done
Productivity: When is coercion “OK”, when not?
Individual variation
Relation of innovation of an extension and its spread
Testability
 Unfortunately, the last two predictions, about the
decreasing likelihood of production (and hence
decreasing frequency) are rarely examined.
 The subtle contrasts used by linguists to test the
boundaries of a construction are rarely present in
corpora, because as phenomena near the boundary
there are not likely to be many instances, if any.
 However, we can still make the prediction and with a
large enough corpus, it should be borne out if the
assumptions about the relation of usage and
linguistic knowledge are correct.
Potential problems
 Why don’t Construction Grammarians in
general try to determine the semantics of
constructions they work on by examining
the lexical items that occur in them most
often? Why is this left to a few Corpusconstruction people?
 Possible reasons:
 Not all constructions occur with specific
recurrent lexical items (or classes of items).
 (but even some surprisingly , general
argument structure constructions do, like
the passive)
Potential problems

Or, it might be more serious:




Construction Grammarians (unlike Cognitive Linguistics) often
do not take the usage-based model to heart and concern
themselves with mechanisms of production and interpretation.
Construction grammar analyses and theoretical descriptions
simply do not bring processing considerations into the picture.
Possibly this is because Construction Grammarians are
agnostic about what exactly happens when language is used.
Backgrounding of processing aspects to CG may be the real
reason that it is so rare to encounter investigations which
use frequency data by most Construction Grammarians.
Corpus-Construction
Grammarians and Corpus
Cognitive Linguists

The small community of Corpus-Construction Grammarians
also are by and large strong cognitivists familiar with the
implications of the usage-based model for the theoretical
side of the model, and not just for selection of data.

So they, along with Corpus Cognitive Linguists, are
prepared to use corpora seriously to discover the cognitive
semantics associated with constructions.
The make-causative
 The make-causative shows strongly marked
distributional asymmetries suggesting it is
not a generalized causative construction,
but instead has specific semantic
properties.
 3 senses of the make causative
construction in English
 mechanical action we need to make it work
 emotional reaction it made me feel good
 compulsion make you can’t make me marry him
 Frequency patterns of animacy of causer
and causee shows clusterings of 3 senses.
(Kemmer 2001; 2005)
The make-causative


The make causative has its own unique ‘constructional profile’ of elements
that typically occur with it and which relate to its function as a construction.
These characteristic distribution patterns are found with many constructions




(e.g. English passive; Dutch laten and doen causatives; German lassen causative;
English let and have causatives; into causative, split infinitives, Swedish future)
Also, distributional observation of the predicate complements of make
causative show a relatively narrow range of complements.
The set of verbs with the widest range of predicates, correlating
semantically with the compulsion use, is the least frequent in type
frequency and overall token frequency.
But most predicates are far more restricted, falling into a number of
semantic subclasses, but grouping into just 2 main classes (mechanical
action; emotional reaction). These also correlate with the clusterings of
animacy causer/causee types. Thus we have two pieces of formal,
distributional evidence that ‘fingerprint’ the construction; their
interpretation results in a unique semantic profile: the constructional
semantics of the make causative.
The make-causative
 Moreover, the make causative shows some
interactional effects between the predicates selected
in it, and the senses of those predicates.
 The make causative effectively constrains the
interpretation of the predicates, where they are
polysemous.
 For example,
 In
 I worked
 And
 It worked
 The default readings are senses of work that
respectively correlate with the animate and inanimate
subjects in the examples.
The make-causative
 However, in the make causative, the sense of work
that occurs cannot be predicted from the animacy
of the causee:
It really made it work.
It really made him work.
 In the causative construction, the second example
is at best ambiguous, but more likely to be
interpreted as mechanical action make rather than
social compulsion make.
The make-causative

Such coercion effects reveal the nature of make + OBJ + INF as a
conventionalized construction




specifically support the analysis of the make causative as being
primarily about mechanical action causation, and secondarily its
extension emotive reaction causation.
the least motivated (entrenched) sense of the make causative is one
which a human is acting on another human to socially compel him/her
to do what the causer wishes to be done.
The normal sense of work (voluntarily perform labor) is suspended
here because the entrenched constructional semantics induce a
reading of involuntary, mechanical action. The sentence could
perhaps be used that way, but it would take some supporting
semantic factors that were coherent with a social frame of strong
social force.
This would be required to overcome the slight semantic
incompatibility of the construction with the normal sense of work
that is found with human subjects.
The historical dimension


We can also observe the history of the make causative. In this history we
find support for the 3 senses of the current construction, and find their
progressive emergence as distinct senses.
When we examine the history of the construction, we find that the
construction did not always occur with its current constructional profile.






In fact, in the earliest days in our data, make was primarily a main verb taking
nominal and adjectival predicates.
It began to take infinitives in early middle English (the do-causative was fading by
then), but as an early causative construction it was principally (most frequently in
type and token frequency) used with the Mechanical Action type of predicates.
Although all 3 senses were apparently attested, the primary use as attested by
frequency was the Mechanical action.
As time went on, emotion predicates began to dominate, as they still do today,
but the number of compulsion predicates began to move past the one or two
found in the earlier periods.
Furthermore, we find the beginnings of coercion effects as the construtcion begins
to exert an interpretational effect on the senses of verbs like work and look.
Effectively, it begins to coerce its component elements to become more
compatible with its own developing semantics.
The historical dimension
The historical dimension



Observing the history of make in this way allows us to see
for the first time how a construction emerges by gradual
extension by speakers until we can see the construction
exert a coercion effect which itself motivates the analysis of
the make causative as a full-fledged construction.
The development followed a trajectory of a changing
constructional profile in terms of preferred predicates and
predicate types.
The changing constructional profile is itself both a symptom
and a mechanism of change, because speakers are
sensitive to frequency and the changing frequency will have
the effect of inducing them to reorganize the construction
into one increasingly resembling the modern make
causative in its constructional profile and its semantic
characteristics, including coercive force.
Conclusions--Utility of
Constructional Profiles
 Observing constructional profiles from
large sets of usage data allows us to
draw conclusions about the semantics
of a construction
 Examining diachronic changes in
constructional profile shed light on
the emergence of the construction
Conclusions: Coercion;
diachrony
 We can find early evidence for the
conventionalization of a construction by
pinpointing the first visible instances of
coercion.
 These show that the construction has
acquired some semantics of its own that
can ‘override’ semantic anomalies.
Constructions are a fruitful perspective for not
only synchronic, but diachronic
investigation.
Acknowledgments
 Thanks to Martin Hilpert; Michael
Barlow
References
Achard, Michel, and Suzanne Kemmer, eds. 2004. Language, culture and mind.
Stanford: CSLI Publications.
Barlow, Michael, and Suzanne Kemmer, eds. 2000. Usage Based Models of
Language. Stanford: CSLI Publications.
Hilpert, Martin. To appear (a). On Swedish future constructions. Proceedings of the
High Desert Linguistic Society Meeting. Albuquerque: HDLS.
Hilpert, Martin. To appear (b). Collograms in the English split infinitive and other
grammatical constructions. Constructions. Special issue on Collostructional
Analysis.
Kemmer, Suzanne. 2001. Causative constructions and cognitive models: The Make
Causative.The First Seoul International Conference on Discourse and Cognitive
Linguistics: Perspectives for the 21st Century, 803-846. Seoul: Discourse and
Cognitive Linguistics Society of Korea.
Kemmer, Suzanne and Martin Hilpert. 2005. Constructional grammaticalization in
the English make-causative. Presented at ICHL in Madison, Wisc. August
2005.
Kemmer, Suzanne and Arie Verhagen. 1994. The grammar of causatives and the
conceptual structure of events. Cognitive Linguistics 5(2), 115-156.
Lakoff, George. 1987. Women, fire and dangerous things. Case studies. Chicago:
University of Chicago Press.
Langacker, Ronald. 1987,1991. Foundations of Cognitive Grammar Vols I and II.
References, cont.
Michaelis, Laura A. 2005. Entity and event coercion in a symbolic
theory of syntax. In Jan-Ola Oestman and Miriam Fried, eds.,
Construction Grammar(s): Cognitive Grounding and Theoretical
Extensions. (Constructional Approaches to Language 3.)
Amsterdam: Benjamins, 45-88.
Stefanowitsch, Anatol, and Stefan Gries. 2003. Collostructions:
Investigating the interaction of words and constructions.
International Journal of Corpus Linguistics 8, 209-243.
Taylor, Christopher. To appear. The X to where Y construction.
Proceedings of the 31st Meeting of the Berkeley Linguistic Society
(BLS 31, 2005). Berkeley: BLS.