Download Masako`s slides on Goldberg, Chapter 9

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Junction Grammar wikipedia , lookup

Inflection wikipedia , lookup

Causative wikipedia , lookup

Preposition and postposition wikipedia , lookup

Pipil grammar wikipedia , lookup

Construction grammar wikipedia , lookup

Pleonasm wikipedia , lookup

Lexical semantics wikipedia , lookup

Antisymmetry wikipedia , lookup

Agglutination wikipedia , lookup

Transcript
Constructions at Work: Chapter 9
Cross-linguistic generalization
in argument realization
1.0. Background
Widespread claims that linking rules are universal.
Pinker (1989) suggests that “Linking rules seem to be
near-universal in their essential aspects and therefore may
not be learned at all…Linking rules can be universal and
innate in the current theory”.
Naigles and Gleitman (1993) also argue that “there is
sufficient cross-linguistic similarity in these linking rules
to get the learning procedure started…There is an
overwhelming tendency, cross-linguistically, for agents to
appear as subjects and themes as direct objects, with
other arguments appearing in oblique cases”.
Background (continues)
Pinker (1989) offers an explicit proposal for mapping from
semantic role to a surface syntactic position:
1. Link the agent to SUBJECT
2. Link the patient to OBJECT
3. Link the theme argument (first argument of BE or GO)
to SUBJECT unless SUBJECT is already linked;
to OBJECT otherwise
4. Link the goal to an OBLIQUE (prepositional phrase)
argument
5. Link the theme argument in a CAUSE TO HAVE
predicate to the second object in a ditransitive
construction
2.0. Thesis of the chapter
Goldberg, however, argues that what are often considered
as “universals”, as those shown earlier, are only
tendencies, and each tendency is a result of general
cognitive, pragmatic, or processing attributes of
human cognition.
Therefore, argument structure constructions are learnable;
the cross-linguistic generalizations that do exist are not
exceptionless; and they can be motivated by nonlinguistic generalizations.
Generalizations about linking between form and function
provide no evidence for a genetically determined
“universal grammar” related to argument structure
generalizations.
3.0. Actors and Undergoers are
expressed in prominent syntactic slots
Question: What are the cross-linguistic generalizations
about how arguments are linked to syntactic
positions?
Dowty (1991) proposed linking generalizations that are now
widely cited as capturing the cross-linguistic universals:
“If there is a subject and an object, and if there is a protoagent and proto-patient, then the proto-agent role is
expressed by the subject, and the proto-patient argument
is linked to object position”.
Dowty notes that this generalization holds except when the
linking is the opposite in syntactically ergative
languages.
(Continues)
Goldberg, however, points out that this
generalization is oversimplified and can be
problematic in that languages with ergative
properties are typically “split ergative”.
This means that Dowty’s generalization does not
hold across the board for all constructions in a
given language. Moreover, within every
language, there are usually particular
constructions that violate the generalization such
as anti-passives.
SPPS proposed by Goldberg
Therefore, Goldberg proposed a “rephrased”
generalization:
The Salient Participants in Prominent Slots
Generalization (SPPS):
“Actors and Undergoers are expressed in prominent
syntactic slots”.
The SPPS generalization accounts for the fact that an Actor
argument without an Undergoer, and an Undergoer
without an Actor are also expressed in a prominent
syntactic position.
SPPS: (continues)
(1) Actors are salient: a human’s attention is naturally drawn to
actors. For example, a scene of a girl chasing a boy could be
construed as a “chasing” scene or a “fleeing” scene, but speakers
overwhelmingly choose to describe it as “chasing” (Fisher et al
1994). Thus, entities that initiate actions are highly cognitively
“accessible”.
(2) Undergoers are salient: humans have a tendency to attend
closely to one particular type of end point. Regier and Zheng
(2003) note that subjects are better able to discriminate between
events that have distinct end points than distinct onsets; subjects
also use a wider range of more specific verbs to describe endpoint-focused actions (such as putting a key in a lock) than onsetfocused actions (such as taking a key out of a lock).
SPPS: summary
Thus, Actors and Undergoers tend to be
expressed in prominent positions,
because they are highly salient.
And, this follows from general facts
about human perception and attention.
Isomorphic Mapping Hypothesis
Lidz at el (2003) claims that “noun phrase number
lines up as simply as possible with argument
number”. What these Universalists claim is
called an Isomorphic Mapping Hypothesis.
They also suggest that the Isomorphic Mapping
Hypothesis is an aspect of “universal grammar”:
i.e. part of a set of hard-wired principles that are
specific to language and are not the result of
empirical experience.
Goldberg’s challenge
Goldberg challenges by arguing that the Isomorphic
Mapping Hypothesis is in fact far from being universally
valid as a generalization about the surface structure.
For example, even within English, we can find many
particular constructions where the Isomorphic Mapping
Hypothesis is systematically violated.
See Table 9. 1. on page 189, where the number of
linguistically expressed participants (complements)
differs from the number of central semantic participants
(arguments).
Pragmatic Generalization
Instead, Goldberg suggests a more robust, but a weaker
generalization.
Pragmatic generalization: that the referents of
linguistically expressed NPs are assumed to be
directly relevant to the semantic interpretation
conveyed.
This generalization follows from Gricean pragmatic
principles. Grice argues that human interactions
generally, not just those that are specifically linguistic,
are governed by a cooperative principle.
The application of Gricean requirements
This Gricean cooperative principle implies that any
information supplied must be relevant to the
communication at hand.
In the case of language,
(A) Linguistically expressed participants must be
relevant to the message being conveyed.
(B) Sufficient information must be indicated for
the intended message.
Pragmatic Mapping Generalizations
Based on these Gricean requirements, Goldberg further provides the
following generalizations:
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.
Notice that (A) does not specify exactly how the referents of
linguistically expressed NPs should be integrated semantically, nor
does (B) specify exactly how semantic participants may be
indicated. This allows for the possibility that different languages
and different constructions obey principles (A) and (B) in different
ways.
Discourse-conditioned argument omission
Pragmatic Mapping Generalizations:
(B) Any semantic participants in the event being conveyed that
are relevant and non-recoverable from context must be overtly
indicated.
Notice that (B) above makes no predictions about semantic
participants that are recoverable or irrelevant.
This is important, because different languages (and indeed, different
constructions within languages) do different things in these
circumstances.
Many, perhaps the majority of languages in the world, allow
recoverable or irrelevant arguments, both subject and object, to be
routinely omitted.
These languages include, Japanese, Korean, Thai, Hungarian, Russian,
Hindi, and Lao.
“Say no more than you must”
There is a clear motivation from conversational pragmatics
for leaving such arguments unexpressed.
One of the four Conversational Maxims proposed by
Paul Grice, Maxim of Quantity:
“Say no more than you must”.
That is, when topical arguments are recoverable, there is no
need to utter them.
Goldberg provide a typical dialogue in Russian:
Q: Did you introduce Ivan to Marsha?
A: Da, dedstavil.
‘Yes, ( I ) introduced (him) (to her)’
Japanese examples:
The following is a typical statement from a lover in
Japanese:
Looking in the eyes of his lover:
S: Aishiteiru yo.
Aishi-te-iru
yo.
‘( I ) love (you)’
love-L-progressive FP
Although English generally requires all arguments to be
overtly expressed, Japanese allows the omitting of any
recoverable arguments.
Japanese examples continue:
A conversation that took place in a high school classroom in Japan:
The teacher is asking students to turn their homework in;
Student A looks at his friend B and asks a question:
A: Shita. (raising intonation)
‘Did (you) (do) (your homework)?’
B: Shitenai.
‘( I ) didn’t (do) (it)’
What needs to be noted here is that no native speakers of Japanese
would feel that something is missing upon hearing these
utterances.
Instead, they will be construed as “full-sentences”.
Argument omission continues:
Yet each language listed above differs in whether or not recoverable
arguments can be omitted, even though omissability is clearly
conventional in those languages. For example:
In Hebrew, discourse topics, whether subjects or objects, can be
omitted, but other recoverable arguments cannot generally be.
In Brazilian Portuguese, omitted objects must be topics and are
predominantly inanimate or third-person animate, i.e. first- or
second-person objects are not readily omitted.
Since the Pragmatic Mapping Generalizations do not determine
whether arguments that are recoverable are necessarily
expressed, this kind of variation in conventionalized options
can be expected.
These examples indicate that the Isomorphic Mapping Hypothesis is
indeed far from being universally valid as a generalization about
the surface structure.
A case of argument omission in English
Interestingly enough, all languages allow omitted
arguments in certain constructions. Even in English,
which rarely allows argument omission compared to
most languages of the world, object omission can
happen due to a particular confluence of discourse
properties. The following examples illustrate this
phenomenon:
(a) The chef-in-training chopped and diced all afternoon.
(b) Tigers only kill at night.
(c) Pat gave and gave, but Chris just took and took.
Argument omission in English continues
Observe another instance where the argument can
be unexpressed:
In the Deprofiled Object construction (1), or it can
be expressed as in (2):
(1) The tiger killed again.
(2) The tiger killed someone/something again.
Notice that the same scene can be described by
either (1) or (2) and yet the number of overtly
expressed NPs differs in the two examples.
Argument omission: summary
All languages allow omitted arguments in certain
constructions.
The existence of such constructions in these
languages further indicates that the underlying
motivation for the expression of arguments is at
root pragmatic.
Thus, the Isomorphic Mapping Hypothesis is
violated cross-linguistically and even within the
language as well.
Generalizations about reduced forms cross-linguistically
It was discussed in Chapter 3 that highly frequent words and
collocations tend to be reduced in online production and, over time,
the reduced forms often become conventionalized.
There seem to be two motivations for the reduction.
First, it is more economical to produce shorter utterances if the same
informational content can be conveyed.
That is, high frequency is correlated with predictability.
Thus the tendency to reduce highly frequent strings may be
motivated in the same way that omission of recoverable arguments,
that we have observed, is motivated: by the pragmatic dictum of
economy: say no more than is necessary.
Second, highly frequent strings are based in the motor control of the
speaker: highly practiced sequences tend to become routinized
and thus abbreviated.
Reduced forms: example
For example, there is a strong cross-linguistic tendency for “want”
constructions to omit the subject pronoun in same-subject
complements.
The following is the English coreferential “equi deletion”:
(1) I want to play.
(2) I want her to play.
This is clearly a cross-linguistic tendency that should be motivated,
but not predicted by linguistic theory.
The lower clause subject omission in (1) is motivated because it is
statistically predictable. That is, we tend to talk more about a
person’s wishes about their own action.
The statistical likelihood of the lower clause subject being
coreferential with the main clause subject allows the complement
clause to be abbreviated.
Reduced forms: summary
Thus, the above is more evidence of an
independently motivated universal of linguistic
expression: highly frequent predictable
combinations tend to undergo reduction.
The role of analogy: ditransitive
Many languages in the world including English have a
ditransitive form with the recipient immediately
postverbal and without preposition, followed by the theme
argument, where this ditransitive form is associated with a
meaning of “transfer”.
The recurrence of this pattern in unrelated languages leads
us to ask why this is a natural form to use to express
meanings related to “transfer”.
Since constructionists argue that constructions are learned
pairings of form and function, it is necessary to explain
why this generalization exists across unrelated languages
in the world.
The role of analogy continues:
In the semantics of “transfer”, the recipient argument is a
prototypical Undergoer, a causee, since it is caused to
receive the theme argument.
At the same time, the theme argument is a different type of
Undergoer in that it undergoes a change of ownership.
He
gave
her
a book.
agent
recipient theme
Thus, the SPPS generalization (The Salient Participants in
Prominent Slots Generalization): actors and
Undergoers are expressed in prominent syntactic slots
motivates both arguments being expressed in prominent
slots.
The role of analogy continues:
Further, the fact that recipient arguments tend to come
before the theme argument in a ditransitive construction
can be explained as follows:
the recipient argument has certain subject-like properties
vis-à-vis the theme argument, and subjects appear before
objects in all the languages that have ditransitive
constructions. Semantically, the recipient is animate, as
are the great majority of subjects.
It also tends to be already given or topical in the discourse
and is typically expressed pronominally, as are subjects.
The role of analogy: summary
Thus, the syntactic expression of the recipient
argument of ditransitives is based on
simultaneous analogies with cause-objects and
possessor-subjects.
It is natural to express the meaning of “transfer”
with a ditransitive form, because of simultaneous
parallels between recipient and patient-objects on
the one hand, and possessor-subjects on the other.
Word-Order Generalizations
The “head-directionality parameter” is often cited as
an example of a purely syntactic generalization
which requires an appeal to universal grammar.
The Head directionality parameter is a proposed
parameter that provides a choice between:
• Heads are found at the beginning of their phrases
(head-initial languages);
• Heads are found at the end of their phrases (headfinal languages).
Word-Order Generalization continues:
The head directionality parameter can be applied to most
cases, which is shown below:
Adpositions:
English:
at
home (preposition)
Japanese:
ie
de
home at
(postposition)
Relative Clauses:
English:
The man who stole the car
Japanese:
Kuruma-o nusunda otoko
car-ACC
stole
man
Word-Order Generalization continues:
However, exceptions to this Head directionality parameter
can also be found:
Persian is a language in which adpositions come before
their objects while verbs come after the objects. And,
the language has stably remained in this state for
hundreds of years.
This fact casts a doubt on the idea that a “head
directionality” is really hard-wired into our genome.
Children learn each of the particular verbs, prepositions,
and nouns involved. As they do so, the order of
arguments is apparent in the input.
Word-Order Generalization continues:
Goldberg suggests that the close relationship
between verbs and adpositions is likely a
consequence of the fact that adpositions
typically evolve from verbs diachronically.
In addition, the statistical likelihood of
processing difficulty can lead to the
conventionalized preferred ordering of
constituents (Hawkins 1990, 1994).
Word-Order: Sign Languages
Also, Goldberg indicates that the tendency for sign
languages to display SOV order comes from the
fact that the same tendency exists in a nonlinguistic domain, that of gesture.
Hearing adults who know no sign language also use
OV order when gesturing without words
(Meadow et al 2000).
Word Order Generalizations:
summary
Thus, languages tend to have stable head orders
due to diachronic processes and processing
preferences.
Languages tend to develop fixed word order or
case marking in order to avoid ambiguity.
The fact that sign languages appear to prefer OV
order and directional agreement appears to have
its roots in non-linguistic gestures.
Iconicity
Even with iconicity, which has been much attacked
in the literature, one can make some robust
generalizations.
Generally, a tight semantic bond between items
tends to be represented by a correspondingly
tight syntactic bond.
Iconicity continues:
For example, grammatical operators such as
plural markers, determiners, case markers,
classifiers are expressed near their
operands in the linear string.
This can mean that morphology and
semantics are aligned.
Iconicity continues:
Goldberg also discusses the strong
preference in Persian and other languages
for treating various kinds of complex
predicates as single, syntactically integrated
predicates.
It is suggested that this is motivated
iconically by the complex predicates’ status
as a semantically integrated predicate.
Conclusion
We have seen that generalizations typically capture
tendencies, not hard and fast constraints.
Certain semantic/pragmatic functions are so relevant and
useful that they can be found in language after language.
Therefore, it is advantageous to explain universal
tendencies by appeal to independently motivated
pragmatic, semantic, and processing, historical,
iconic, and analogical facts, ultimately buttressed by
experimental data. This will shed light on issues related
to why languages are the way they are, since these would
not be expected to be perfectly exceptionless.