Title: Constructions at Work by Adele Goldberg Oxford U Press 2006
1Constructions at Workby Adele GoldbergOxford U
Press 2006
- Presented by Laura A. Janda
2Overview of the book
- Part I (Chap 1, 2, 3) Theory, comparison with
Generativism - Part II (Chap 4, 5, 6) Learning, how and why
constructions are learned - Part III (Chap 7, 8, 9) Constraints, how they
work language-internally and cross-linguistically - Chap 10 Comparison with Generativism and other
construction grammars (Fillmore, Croft,
Langacker) - Chap 11 Conclusion
3What is a construction?
- a conventionalized pairing of form and meaning
- This works at many levels simultaneously -- its
constructions all the way down - Any given utterance contains many constructions
- Constructions are not componential, meaning can
be wholistic/emergent, though some items can be
put in slots
4Part I (Chap 1, 2, 3) Theory
5Ch1 Constuctionism (CCxG) vs. Generativism
- Where they agree
- Language is a cognitive system
- It is possible to combine structures to create
novel utterances - Theory must account for acquisition, for
potentially infinite output given finite input
6Ch1 Constuctionism (CCxG) vs. Generativism
- Where they disagree
- Input is not so impoverished
- Relationship of structure and meaning
- Nature of meaning
- Distinction between core language (grammar
rules) and periphery (lexicon storage) - Status of idioms
- How grammar (core) is transmitted
- Transformation/derivation
- Levels of language
- Underlying representations, empty elements
- Relationship of language to cognition
7Ch1 Basic Assumptions of CCxG
- Language is the aggregate network of
constructions - Language is learned on the basis of input and
general cognitive mechanisms - Storage and grammatical rules (generalizations)
overlap, coexist and cooperate - Idioms are not special, no clear boundary between
idioms and grammar
8Ch1 Basic Assumptions of CCxG
- Semantics are always important, there are no
purely formal generalizations - Some semantic prototypes may exist
cross-linguistically because humans are more
alike than different - Semantics is recursive language works because I
can imagine what you think about what someone
else believes
9Ch2 Surface generalizations
- Each surface pattern should be considered on its
own terms - Various constructions may be related to each
other, but they are not derived from each other - Constructions can be combined freely
(juxtapositions blends as well as addition),
when conflicts arise they are resolved by
construal or coercion - Note
- Goldberg is dealing with constructions
specifically at the level of the verb phrase, but
her model can be used at other levels
10Ch2 Notation
- Sem
- gives arguments labels, but does not assume that
they come from a universal set, can be determined
by the construction - Syn
- surface structure
- Solid lines
- argument role of construction must fuse with
independently existing participant role of verb - Dashed lines
- argument role can be contributed by construction,
doesnt have to be inherent to verb
11Ch2 The Surface Generalization Hypothesis
- There are typically broader syntactic and
semantic generalizations associated with a
surface argument structure form than exist
between the same surface form and a distinct form
that it is hypothesized to be syntactically or
semantically derived from - Each argument structure pattern is best analyzed
in its own terms, without relying on reference to
a possible alternative paraphrase
12Ch2 Interaction of roles
- Interaction of participant roles (ltverb) and
argument roles (ltconstruction) accounts for
overlap in meaning of paraphrases
ambiguity/multiple interpretations - Each sense of a verb is associated with certain
participant roles
13Ch2 Two constraints on fusion of participant
roles (ltverb) and argument roles (ltconstruction)
- Semantic Coherence Principle participant role
of verb and argument role of construction must be
semantically compatible the more specific
participant role of the verb must be construable
as an instance of the more general argument role
a categorization task - Correspondence Principle profiled participant
roles of the verb must be encoded by profiled
argument roles of the construction lexical
semantics and discourse pragmatics are aligned - Correspondence principle can be overridden by a
construction such as passive, which deemphasizes
a normally profiled role like agent.
14Ch2 Arguments and Adjuncts
15Ch3 Item-Specific knowledge and generalizations
- Information about specific exemplars is stored
- Abstractions are created locally, on the basis of
small numbers of exemplars - Language learning must involve memories of
individual examples because the end state of
grammar is only partially general - Oftentimes an idiom is constructionally the same
as something that is very common
16Ch3 Item-Specific knowledge and generalizations
- It must be that both generalizations and
instances are stored - Children seem to both learn instances and make
generalizations
17Ch3 Usage-Based models of language
- Grammars are usage-based if they record facts
about actual use of linguistic expressions such
as frequencies and individual patterns that are
fully compositional alongside more traditional
linguistic generalizations - Language learners are conservative they do not
generalize significantly beyond the evidence in
the input, and they seek generalizations that are
consistent with the evidence presented - Language learners seek both local consistency and
global consistency. Local consistency makes
learners aim to be conservative and stick closely
with the local instances that they have
witnessed. Global consistency makes learners seek
out generalizations among instances so that the
overall system coheres
18Part II (Chap 4, 5, 6) Learning
19Ch4 How constructions are learned
- Growing evidence that at least certain patterns
in language are learnable on the basis of general
categorization strategies - Poverty of stimulus point of view is biologically
implausible - What is crucial is the uncontroversial notion
that there do in fact exist correlations between
formal linguistic patterns and meaning.
20Ch4 Skewed input
- In input from mothers, go, put, give, make are
more frequent than all other verbs, and are used
in only one construction each - go Intrans Motion
- put Caused Motion
- give Ditransitive
- make Resultative
- Certain verbs are templates for the learning of
other verbs and constructions - children gradually abstract a more general and
purely syntactic pattern
21Ch4 The advantage of skewed input an anchoring
effect
- A high-frequency token is likely to be considered
a prototype by the learner - Categories with a salient stable prototype are
easier to learn - Frequency and order of acquisition play key roles
in category formation - Input is structured in such a way as to make the
generalization of argument structure
constructions straightforward - Knowledge of constructions is a straightforward
extension, by generalization, of knowledge of
groups of words that behave similarly
22Ch5 How generalizations are constrained
- Factors
- a) entrenchment (token frequency)
- b) statistical pre-emption (repeated witnessing
of a word in a competing pattern) - c) type frequency
- d) variety of items that occur (degree of
openness) - How to minimize overgeneralizations more
specific knowledge always pre-empts general
knowledge in production when all else is equal
23Ch5 Collaboration between storage and
conservative generalization
- Specific knowledge vs. generalizations
- referee preempts reffer
- children preempts childs
- went preempts goed
- moister co-exists with the abstractly created
more moist - Pre-emption gives learners indirect negative
input since an instance of X suggests that the
context is probably not good for any other Y - Constructions that have appeared with many
different types are more likely to appear with
new types than constructions that have only
appeared with few types
24Ch6 Why generalizations are learned
- Verbs are good cues to sentence meaning cf.
childs early learning of verb-centered argument
structure patterns (verb islands) - But children also generalize beyond the verb
this is useful in predicting overall sentence
meaning - Constructions are often better predictors of
overall meaning than verbs
25Ch6 Cue validity of verbs vs. constructions
- Most verbs appear in more than one construction
- Pat got the ball over the fence (VOL pattern -gt
caused motion) vs. Pat got Bob a cake (VOO
pattern -gt transfer) - Here got has lower cue validity than construction
- Cue validity for VOL construction in predicting
caused-motion meaning in mothers speech 63-85
vs. Cue validity for verbs 68 (near 100 for a
few, extremely low for most) - Cue validity for VOO construction 61-94 vs. Cue
validity for verbs 61 (a few with 100 and very
low for all the rest)
26Ch6 Cue and category validity for caused-motion
- Corpus of mothers speech
- 100s Eng verbs can appear in the C-M
construction, for nearly all of them, cue
validity is close to 0 (highest is put at 62) - Only 3 constructions express caused motion, and
the VOL construction has 83 category validity - constructions are better cues to sentence meaning
than verbs
27Ch6 Summary
- Verb is best single word predictor of overall
sentence meaning, but constructions have equal
cue validity and much higher category validity,
thus the construction is at least as reliable and
much more available - Because many verbs have low cue validity in
isolation, attention to the contribution of the
construction is essential - Hearing or producing a particular construction
makes it easier to produce the same construction - Instead of learning a myriad of unrelated
constructions, speakers do well to learn a
smaller inventory of patterns in order to
facilitate online production
28Part III (Chap 7, 8, 9) Constraints
29Ch7 Island constraints
- Generative grammar Islands are certain
syntactic constructions that restrict
extraction (complex noun phrases, complex
subjects, complements of manner-of-speaking
verbs, adjunct clauses) - CCxG so-called movement involves combinations
of constructions and clashes in their information
structure - Keenan Comries (1977) accessibility hierarchy
subject gt direct object gt oblique object gt object
of comparison -- this corresponds to foregrounded
gt gt gt backgrounded
30Ch7 Island constraints, contd
- It all boils down to this -- Backgrounded
constructions are islands - The restriction on backgrounded constructions is
clearly motivated by the function of the
constructions involved. - Elements involved in unbounded dependencies are
positioned in discourse-prominent slots - It is pragmatically anomalous to treat an element
as at once backgrounded and discourse-prominent - Example ditransitive recipient argument, which
is almost always (old) given information -- it is
thus backgrounded and resists unbounded
dependencies
31Ch7 More examples
- Subordinate clauses most, but not all, are
backgrounded - Reason clauses may or may not be backgrounded,
same goes for adjuncts - Both restrictive and non-restrictive relative
clauses are backgrounded - Displacement from canonical position creates
additional processing load and this combines with
the pragmatic clash to result in unacceptability
32Ch7 Other examples
- Quantifier scope ( use of a, one, some with all,
every, each) is also strongly correlated with
topicality - The information-structure properties of
constructions predict their predominant
assignment of scope
33Ch8 Subject-Auxiliary Inversion (SAI)
- SAI found in yes/no ques, non-subject wh-ques,
counterfactual conditionals, sentences with
negative adverbs, exclamatives, comparatives,
negative conjuncts, positive rejoinders. - It has been claimed that these contexts are
unrelated, so this is just syntax - CCxG shows that syntactic form of SAI is
motivated by semantic/pragmatic function - Goldberg shows that SAI forms a radial category
34(No Transcript)
35(No Transcript)
36Ch8 The SAI prototype
- The prototype is a non-prototypical sentence
non-positive, non predicate focus, non-assertive,
dependent, non-declarative - The various types of SAI are linked to this
prototype - But WHY should non-positive contexts be indicated
by an inversion of subject and auxiliary?
37Ch8 Why SAI?
- Auxiliaries carry information about polarity as
well as tense and aspect - By positioning the auxiliary in a non-canonical
position, the construction conveys that the
polarity involved is not the canonical, positive
polarity - Note that SAI is rare cross-linguistically, but
UG assumes SAI is a universal -- while the formal
approach may be descriptively adequate, it does
not, in the case of SAI, have any explanatory
force
38Ch9 Mapping of roles is learnable and learned
- Universalist vs. Emergentist debate
- Linking rules (link agent to SUBJECT, patient
to OBJECT, etc.) are a problem for UG, assumed to
be universal, not learned - CCxG solution not universals, but tendencies,
that result from general cognitive, pragmatic, or
processing attributes of human cognition - Isomorphic Mapping Hypothesis ( of NPs and
linking) - Not exceptionless
- Motivated by non-linguistic generalizations
39Ch9 Isomorphic Mapping Hypothesis
- Exceptions
- Pat was killed (missing agent)
- The tiger killed again (missing patient)
- Pat buttered the toast (incorporated patient)
- Pat laughed a hearty laugh (object of an
intransitive) - These exx show that we cannot claim universal
validity it makes more sense to accept a weaker
pragmatic generalization...
40Ch9 Pragmatic Mapping Generalizations
- (A) The referents of linguistically expressed NPs
are interpreted to be relevant to the message
being conveyed. - (B) Any semantic participants in the event being
conveyed that are relevant and non-recoverable
from context must be overtly indicated. - This formula states only what has to be there,
not what might be required by the grammar of a
given language. - Participants that are irrelevant or recoverable
dont have to be expressed, and languages differ
on how they deal with this some require them,
some dont.
41Ch9 Other examples
- Goldberg shows that may supposed universals can
be accounted for by non-linguistic motivations - Discourse-conditioned argument omission
- Recoverable arguments need not be expressed
- Ditransitive construction
- Recipient is subject-like
- Word-order generalizations
- Processing simplicity
- Iconicity
- Proximity both functionally and
temporally/spatially
42Ch9 Summary
- Generalizations typically capture tendencies, not
hard and fast constraints - It is advantageous to explain universal
tendencies by appeal to independently motivated
pragmatic, semantic, and processing facts, since
these would not be expected to be perfectly
exceptionless
43Ch10 Differences between CCxG and UG
- UG
- Assumes derivational approach to syntax
- Ignores speaker construal
- Assumes pairing of underlying form and coarse
meaning rather than surface form and detailed
function - Assumes only certain patterns are constructions
words and morphemes are stored separately - Assumes syntax makes no reference to semantics or
function - Assumes constructions are universal and
determined by UG - Does not address language-internal
generalizations across constructions
44Ch10 Differences between CCxG and UG
- CCxG assumes
- Each verb sense lexically specifies the number
and semantic type of arguments it has, and which
of those arguments are obligatory (profiled),
along with its rich frame semantic meaning - Each argument structure construction specifies
its formal properties, its semantic and
information-structure properties, and how it is
to combine with verbs and arguments - Constructions specify which if any arguments they
contribute - There is a cline of productivity and regularity
- The role of the lexicon includes phrasal patterns
with their own idiosyncratic syntactic or
semantic properties - It is the interaction of the argument structure
of the verb and construction that gives rise to
interpretation
45Ch10 Various Construction Grammars
Ronald Langacker
William Croft
46Ch11 Conclusion
- Speakers knowledge of language consists of
systematic collections of form-function pairings
that are learned on the basis of the language
they hear around them - The usage-based model of grammar is supported not
only by linguistic facts, but also by what we
know about how non-linguistic categories are
represented - Far from being an arbitrary collection of
stipulated descriptions, our knowledge of
linguistic constructions, like our knowledge
generally, forms an integrated and motivated
network - Child learners can make statistical
generalizations, use semantics and pragmatics for
making interpretations and generalizations - Constructions can be learned, and learned
quickly, on the basis of the input