2. understanding speech
1. differentiating speech sounds from other
noises
2. recognizing words
3. activating their syntactic and semantic
properties
4. building a grammatical structure
5. interpreting this structure
3. building a grammatical structure?
Do we need to do that?
Well, consider this:
Man bites dog. vs. Dog bites man.
4. … and how about this?
Police kill man with TV tuner.
Life means caring for hospital director.
Retired priest may marry Springsteen.
Kicking baby considered to be healthy.
Brand door kaars geblust.
Slingerend in een jeep heeft de politie vrijdagnacht een
45-jarige Zeisterse staande gehouden.
De burgemeester ging na het telefoongesprek met de
officier van dienst naar bed.
5. upshot
• the intended meaning and the funny meaning do
not result from different word meanings or sth.
• rather, they derive from different arrangements of
words into word groups (phrases)
• so, structure determines meaning
• so, yes, structure building (parsing) is a
necessary component of language
comprehension
9. wait-and-see
• take in words up to a natural boundary (e.g. sentence
ending), and then try to arrange them into a structure,
following the grammatical rules
• comprehension will arise after a sentence has ended
• but: we feel we often know how somebody else’s
sentence will end
• and, if a sentence is interrupted, we nonetheless
understand what was said
10. parallelism
• at any bit of input, create all structures that are
compatible with it
• prediction: the competing structural
representations for an ambiguous piece of input
will all be kept in memory until disambiguating
information comes in
• problem: ambiguity is ubiquitous in natural
language, and memory is limited
11. (conservative) guessing
• at any bit of input, attempt to build as much
structure as possible
• prediction: mistakes will be made, and retracing
(repairing) will occur
19. initial attachment decisions
• Garden Path Theory:
attach incoming words to the evolving structure
in the most economic way, I.e., without involving
building blocks the necessity of which is unclear.
20. for example
He hit the man with the binoculars.
The structure in which “the binoculars” is the
instrument of “hitting” has one node less than the
structure in which it is an attribute of “the man”.
28. another nice one
While she was mending the sock
fell off her lap.
what the parser likes: late closure
29. another nice one
While she was mending the sock
fell off her lap.
what the parser has to do: early closure
note: since ‘while’ introduces a subordinate S, the main
S is expected anyway: minimal attachment is irrelevant
30. Keep in mind that …
…the garden path model assumes that structural
(syntactic) analysis is prior to, and independent of,
semantic and pragmatic interpretation!
32. a note on measurement
• Sometimes a garden path (i.e. parsing difficulty)
is consciously noticeable, like in the horse raced
example.
• However, language is full of ambiguities, and the
majority go by unnoticed.
• So how can we, in such cases, determine
whether the sentence processor has a problem?
33. time
• The answer lies in the assumption that every bit of
work the sentence processor does takes some time.
• If the processor is garden-pathed, it will have to
retrace and correct its previous decisions, in order
to accommodate the incoming words that don’t ‘fit
in’.
• This we can measure by the time it takes to process
the critical words.
40. relative clauses are ambiguous
… in Dutch:
Karel hielp de mijnwerker die de man vond.
Karel helped the mineworker REL the man found
‘mijnwerker’ and ‘man’ can both be finder and ‘findee’
in other words:
“die”, which refers back to “mijnwerker” can be both
subject (subject-relative) and object (object-relative)
41. subject-relative is preferred
1. Karel hielp de mijnwerkers die de man vonden.
Karel helped the mineworkers-PL REL the man-SG found-PL
plural verb needs plural subject; “die” = subject
2. Karel hielp de mijnwerkers die de man vond.
Karel helped the mineworkers-PL REL the man-SG found-SG
sing. verb needs sing. subject; “die” = object
• Less errors, shorter reading times, for 1 than
for 2
42. subject-relative is preferred
Explanation:
readers want to analyse the relative pronoun (“die”)
as the subject of the embedded clause, due to the
Active Filler Strategy
(I.e., this is the most economic option)
if “die” turns out to be the object, the processor has
to re-analyze
Frazier 1987
43. Mak 2001
1. … moeten de inbrekers, die de bewoner
beroofd hebben, nog een tijdje op het …
2. … moet de bewoner, die de inbrekers beroofd
hebben, nog een tijdje op het …
3. … moeten de inbrekers, die de computer
gestolen hebben, nog een tijdje op het …
4. … moet de computer, die de inbrekers
gestolen hebben, nog een tijdje op het …
44. Mak 2001
1. … inbrekers, die de bewoner … hebben …
SR; animate - animate
2. … bewoner, die de inbrekers … hebben …
OR; animate - animate
3. … inbrekers, die de computer … hebben …
SR; animate - inanimate
4. … computer, die de inbrekers … hebben …
OR inanimate - animate
350
386
347
336
ms. on aux + 1
45. summary
• when the two nouns are both animate, SR is faster than
OR
• when there is a difference in animacy, the difference in
reading time disappears
animacy helps deciding which of the two has to be the
subject – immediately
NO REANALYSIS
46. upshot
• Mak has shown that semantics (the animacy factor) has
a very early effect on parsing decisions.
• So it would seem unlikely that semantic interpretation
really follows structural analysis.
• Rather, it looks like the two work in tandem.
… but one could argue that the measurements are not
sufficiently sensitive…
47. does this mean that …
… syntactic and semantic analysis are basically
the same process?
53. upshot
• N400 is specifically sensitive to semantic
information
• P600 is specifically sensitive to syntactic
information
54. upshot
• These two components are different in various
attributes:
– polarity (N vs. P)
– latency (400 vs. 600 ms)
– distribution over the scalp
• So it would seem that different neural networks
generate them
different centers for syntactic and semantic processing
55. wrap-up
• Classical models of sentence processing
assumed that syntactic analysis is prior to and
independent of semantic/pragmatic interpretation
• reading-time evidence casts doubt on the priority
assumption
• electrophysiological evidence supports the
independence (autonomy) assumption