1 dependency structure and cognition richard hudson depling2013, prague
TRANSCRIPT
1
Dependency structure and cognition
Richard Hudson
Depling2013, Prague
2
The question• What is syntactic structure like?
– Does it include dependencies between words (dependency structure)?
– Or does it only contain part-whole links (phrase structure)?
She looked after him
after him
She looked after him
3
Relevant evidence: familiarity
• University courses teach only one approach.
• School grammar sometimes offers one.– Usually dependency structure– even in the USA
• Reed-Kellogg sentence-diagramming
– especially in Europe– and especially in the Czech Republic!
4
What Czech children do at school
blossomed out kingcups
by stream
near
yellow
Jirka Hana & Barbora Hladká 2012
5
or even …
6
Relevant evidence: convenience
• Dependency structure is popular in computational linguistics.
• Maybe because of its simplicity:– few nodes– little but orthographic words
• Good for lexical cooccurrence relations
7
Relevant evidence: cognition
• Language competence is memory
• Language processing is thinking
• Memory and thinking are part of cognition
• So what do we know about cognition?
• A. Very generally, cognition is not simple– so maybe syntactic structures aren't in fact
simple?
8
B. Knowledge is a network
me Gaynor
Lucy
Peter
JohnGretta
Colin
9
C. Links are classified relationsperson
woman man
relative
parent child
mother father
is-a
10
D. Nodes are richly related
me Gaynor
Lucy
Peter
JohnGretta
Colin
m
s
mf f
d
gf
bb
wh
s ss
s
11
E. Is-a allows default inheritance
• Is-a forms taxonomies.– e.g. 'linguist is-a person', 'Dick is-a linguist'
• Properties 'inherit' down a taxonomy.
• But only 'by default' – exceptions are ok.– e.g. birds (normally) fly– but penguins don't.
12
Penguins
bird
robin
'flies'
penguin 'doesn't fly'
robin* 'flies' penguin* 'doesn't fly'
13
Cognitivism
• 'Cognitivism'– 'Language is an example of ordinary cognition'
• So all our general cognitive abilities are available for language– and we have no special language abilities.
• Cognitivism matters for linguistic theory.
14
Some consequences of cognitivism
1. Word-word dependencies are real.
2. 'Deep' and 'surface' properties combine.
3. Mutual dependency is ok.
4. Dependents create new word tokens.
5. Extra word tokens allow raising.
6. But lowering may be ok too.
15
1. Word-word dependencies are real
• Do word-word dependencies exist (in our minds)?– Why not? – Compare social relations between individuals.
• What about phrases?– Why not? – But maybe only their boundaries are relevant?– They're not classified, so no unary branching.
16
Punctuation marks boundaries
• At the end of the road, turn right.
• Not:– At the end of the, road turn right.– At the end, of the road turn right.– At the end of the road turn right,
• How do we learn to punctuate if we can't recognise boundaries?
17
No unary branching
• If S NP + VP, then:
S
NP VP
N V
Cows moo. Cows moo.
N V
But if a verb's subject is a noun:
18
2. 'Deep' and 'surface' properties combine.
• Dependencies are relational concepts.• Concepts record bundles of properties that
tend to coincide– e.g. 'bird': beak, flying, feathers, two legs, eggs– 'mother': bearer, carer
• So one dependency has many properties:– semantic, syntactic, morphosyntactic– e.g. 'subject' ….
19
'subject'
The typical subject is defined by• meaning
– typically 'actor' or …
• word order and/or case– typically before verb and/or nominative
• agreement– typically the verb agrees with it
• status– obligatory or optional, according to finiteness
20
So …
• Cognition suggests that 'deep' and 'surface' properties should be combined– not separated
• They are in harmony by default– but exceptionally they may be out of harmony– this is allowed by default inheritance
21
3. Mutual dependency is ok.
• Mutual dependency is formally impossible in standard notation
• And is formally impossible in phrase structure theory
• So if it exists, we need to – resist PS theory– change the standard notation
22
Mutual dependency exists
• I wonder who came?
• Who is subject of came, – so who depends on came.
• But who depends on wonder
• and came can be omitted:– e.g. Someone came – I wonder who.
• So came depends on who.
23
Standard notation
A
B
B
A
A 'dominates' B
so A is above B
so B cannot 'dominate' A
24
4. Dependents create new word tokens.
• General cognition: – every exemplar needs a mental node.– no node carries contradictory properties.– so some exemplars need two nodes.
• E.g. when we re-classify things.– NB we can remember both classifications
25
What kind of bird?
bird
blackbird
B
B*
?mate
26
And in language …
word
LIKE-verb
like
?subject
I
like*NB like* is a
token of a token
27
The effect of a dependent
• When we recognise a dependent for W, we change W into a new token W*.
• The classification of W* may change.• W* also has a new meaning
– normally a hyponym of W– but may be idiomatic
• If we add dependents singly, this gives a kind of phrase structure!
28
typical French house
HOUSE housemeaning
house
house*
house**
French
housemeaning
French housemeaning
typicalFrench house
meaningtypical
29
Notation
houseFrenchtypical
house*
house**
houseFrenchtypical
30
5. Extra word tokens allow raising.
rainssubjectit
keepsit* raining
subjectit
subject predicative
31
Raising in the grammar
A
A*
C
B
higher parent
lower parent
shared
A* is-a A, so A* wins.
32
6. But lowering may be ok too.
• Raising is helpful for processing– the higher parent is nearer to the sentence root.
• But sometimes lowering is helpful too– e.g. if it allows a new meaning-unit.
• Eine Concorde gelandet ist hier nie.
a Concorde landed has here never.
A-Concorde-landing has never happened here.
33
German Partial VP fronting
gelandet ist hier nie
Eine Concorde
Eine Concorde*
higher parent
lower parent
lowered
34
Conclusions
• Language is just part of cognition.• So syntactic dependencies are:
– psychologically real– rich (combining 'deep' and 'surface' properties)– complex (e.g. mutual, multiple).
• And dependency combines with – default inheritance– multiple tokens