Testing the Strength of the Spurious Licensing Effect of NPIs

1
Testing the Strength of the Spurious Licensing
Effect of NPIs

• Ming Xiang
• Brian Dillon
• Colin Phillips
• University of Maryland

CUNY 2006
2

• A puzzle
• The man who had no beard was ever happy.

3

• Negative polarity items (NPIs) are lexical items
  that need to be licensed in certain environments,
  prototypically negation
• John wouldnt ever do that.
• John would ever do that.
• John didnt know any French.
• John knew any French.
• John didnt lift a finger to help Bill.
• John lifted a finger to help Bill.
• There hasnt been an accident in years.
• There has been an accident in years.

4

• NPIs in general need to be c-commanded by their
  licensors
• a. Nobody would ever do that.
• b. The man who nobody liked would ever do that.

5
Previous work on NPIs

• Vasishth et al. 2005, Drenhaus et al. 2005
• (1) a. Kein Mann, der einen Bart hatte, war
  jemals glücklich.
• no man who a beard had was ever
  happy
• 'No man who had a beard was ever happy.
• b. Ein Mann, der einen Bart hatte, war jemals
  glücklich.
• a man who a beard had was ever
  happy
• 'A man who had a beard was ever happy.
• c. Ein Mann, der keinen Bart hatte, war jemals
  glücklich.
• a man who no beard had was ever
  happy
• 'A man who had no beard was ever happy.'

6

• Mean accuracy and reaction time (n24)
• accuracy reaction time
• Accessible 85 540ms
• licensor
• b. Inaccessible 70 712ms
• licensor
• c. No licensor 83 554ms

7
(No Transcript)
8

• From these data
• There is a spurious licensing effect, despite the
  structural constraints.
• ERP data suggests that the spurious licensing is
  early.

9

• An activation based model (Lewis and Vasishth
  2005)
• Sentence processing is derived from general
  cognitive principles, specifically, it can be
  viewed as a series of skilled associative memory
  retrievals modulated by similarity based
  interference and fluctuating activation.
• The spurious licensing is caused by an imperfect
  feature match
• a man who had no beard is ever happy
• syntactic features mismatch
• semantic features match
• The model is consistent with the finding that
  there is not any timing dynamics to the spurious
  licensing effect.

10
Surprising.

• The sentence just sound odd!
• There is a lot of empirical evidence showing that
  people are very sensitive to syntactic violations
• Sensitivity to syntactic islands (Stowe 1986,
  Traxler Pickering 1996, Yoshida et al. 2004)
• Traxler Pickering 1996
• We like the book/the city that the author wrote
  unceasingly and with great dedication about ___
  while waiting for a contract.
• We like the book/the city RCthat the author who
  wrote unceasingly and with great dedication
  saw__ while waiting for a contract.

11

• Binding principles in parsing binding principles
  as initial filter (Nicol and Swinney 1989, Sturt
  2003, Lee and Williams 2005, Kazanina, Lau et al.
  2006)
• Jonathan remembered that
• the surgeon had pricked himself/herself with a
  used needle.
• The surgeon who treated Jonathan had pricked
• himself/herself with a used needle.
• (Sturt 2003)

12
Even more surprising.

• NPIs are potentially different from other long
  distance dependencies
• The unique property of NPI licensing might
  suggest the parser should respect structural
  conditions in a even more rigid way

13
Common Long-distance Relations

• Filler-Gap
• Who did John talk to _ ?
• Backwards Anaphora
• While she was reading a book, Mary remembered
• Antecedent-Pronoun
• John told Mary that he wanted to be a linguist.

14

• In these long distance dependencies, the parser
  builds a relation between two lexical items, or a
  lexical item and a gap position.
• AB
• In the process of retrieving A from the memory,
  an element with similar property to A might cause
  interference
• AA B

15

• But NPI licensing potentially doesnt involve
  searching for individual licensors.

16
Diversity of licensors

• Some examples of licensors (data from Linebarger
  1987)
• Negation
• John didnt know any French.
• Few
• Few people have any interest in this.
• Only
• Only John has a hope in hell of passing.
• Relative clauses headed by every
• Everyone who knows a damn thing about English
  knows this word.
• Questions
• Have you ever met George?
• Too
• John is too tired to give a damn.
• Antecedent of conditionals
• If you steal any food, theyll arrest you.
• Comparatives
• He was taller than we ever thought he would be
• Adversative predicates (refuse, doubt, etc.)
• He refused to budge an inch.

17
NPIs are diverse too

• NPIs range from single lexical items to complex
  idioms any, ever, much, anymore, in years, yet,
  a red cent, the slightest difference, give a
  damn, budge an inch, etc.
• Not all the NPIs are compatible with all the
  licensors (e.g. Zwarts 1998)

18

• Searching for individual licensors implies that
• The parser needs to keep track of all the
  possible licensors for each NPI and go through
  them every time an NPI is encountered.
• The parser also needs to keep track of selection
  (strength) constraints of each licensor and NPI.

19

• Sometimes it doesnt seem like an individual
  licensor is available
• a. Has John ever lifted a finger to help you?
• b. The reason one ever bothers to decant a wine
  is to leave the sediment behind in the bottle
  SouthWest Airlines Spirit 199447)
• c. There is a shred of evidence to suggest that
  he is the murderer.
• d. There isnt a shred of evidence to suggest
  that he is the murderer.
• e. Well take a shred of evidence and try to
  turn that into a story.
• f. Go ahead. Get her on the witness stand and
  try her with your shreds of evidence. Mr.
  Liedecker in the movie Laura
• (Horn 1996, Israel 1998)

20
Alternatively

• It is the semantic properties of the whole
  proposition matters for the parser, not the
  individual licensors.
• The relevant semantic/pragmatic property is
  incrementally parsed along the way, and
  automatically licenses the NPI if it encounters
  one.

21
So, what is the relevant semantic property?

• Logical properties of downward entailment
  (Ladusaw 1979)
• John bought a car. ? John bought a red car.
• John didnt buy a car. ? John didnt buy a red
  car.
• The more general property of negation
  direct/indirect negation.
• The deeper motivation is likely tied to the
  pragmatic function of NPIs of making stronger
  statements.

22
A likely consequence of the semantics driven
approach

• To derive the necessary semantic properties to
  license a NPI, the structural conditions
  (c-command) have to be respected, because
  semantic composition is parasitic on structure
  building.
• a. Nobody bought a car.
• ? Nobody bought a red car.
• b. The man that nobody likes bought a car.
• ? The man that nobody likes bought a red car.

23
Summary

• The German results are counterintuitive.
• There is evidence showing that the parser
  respects syntactic constraints.
• The unique properties of NPIs imply a processing
  mechanism that respects the syntactic
  constraints.

24
Testing the spurious licensing effect

• First Try
• The spurious licensing could be due to the fact
  that no is a prototypical NPI licensor. It is an
  artifact of the high co-occurrence frequency of
  no and ever.

25
Materials

• Three licensors no, only, and few
• They differ in their co-occurrence frequency with
  ever
• No 9 (19 with other transparent negations
  not, nobody, nothing)
• Only 5
• Few 2
• (Gigaword corpus)

26

• The same manipulation as in Vasishth et al. 2005
• Licensor in an accessible position.
• Licensor in an inaccessible position.
• No licensor.

27
Sample item

• Accessible licensor
• a. No bills that the Democratic senators have
  supported will ever become law.
• Very few bills that the Democratic senators have
  supported will ever become law.
• c. Only three bills that the Democratic senators
  have supported will ever become law.
• Inaccessible licensor
• d. The bills that no Democratic senators have
  supported will ever become law.
• The bills that very few Democratic senators have
  supported will ever become law.
• The bills that only three Democratic senators
  have supported will ever become law.
• No licensor
• g. The bills that the democratic senators had
  supported will ever become law.

28
Expt 1 Speeded acceptability judgment

• Speeded presentation, 400 ms per word, 3 s
  judgment time. N 21.
• Replication and extension of Vasishth et al 2005.

29

0.84
0.42
0.19
Frequency didnt play a role. Spurious licensing
replicated for all three licensors.
30

• What happens if participants have more time?
• Expt 2 offline judgments

31
Expt 2a 5-point scale ratings

4.27
2.37
1.88
Items 28, N 14
32
Expt 2b yes/no judgments

0.81
0.22
0.05
Item 28, N21
33

• Could the effect due to the incomplete structure
  building? For instance, the parser could
  mistakenly think the NPI is located within the
  relative clause.
• Expt 3 auditory presentation

34
Expt 3 - Auditory presentation

• Auditory presentation with a clear (but natural)
  prosodic break at the relative clause boundary. N
  28.

35

0.87
0.26
0.07
36

• All the tasks so far are offline tasks.
• Expt 4 self-paced reading, N49

37
Expt 4 Self-paced Reading Times
No1 bills2 that3 the4 Democratic5 senators6 have7
supported8 have9 ever10 become11 law12
38
Summary

• In the offline judgment tasks, there is a robust
  spurious licensing effect in English. It is
  modulated by the specific tasks, but the general
  pattern isnt affected by the co-occurrence
  frequency between the licensor and the NPI.
• In the online task, there is no immediate
  spurious licensing effect at the critical region
  and the spill over region. This is in contrast to
  the ERP result in Vasishth et al. 2005 Drenhaus
  et al. 2005.
• Licensor frequency doesnt have effects in the
  online reading task.

39
Discussion

• The contrast between RT and ERP data
• The contrast is crucial because of the different
  conclusions we can draw for the time course of
  the spurious licensing effect.
• Further investigation is needed. Self-paced
  reading time might not be a sensitive enough
  measure.
• On the other hand, although ERP measures could
  reflect the process of semantic integration, they
  could also tap into surface associative
  mechanisms.

40

• Fischler et al 1983 found the N400 varied as a
  function of the relatedness of the two NPs in the
  sentences below, but not of the truth conditions
• a. A robin is a bird.
• b. A robin is not a bird.
• c. A robin is a vehicle.
• d. A robin is not a vehicle.
• Suggests that N400 might not uniquely reflect
  sentential semantic integration.

41

• Deacon et al 2000 obtained N400 priming effects
  in a masked priming paradigm, even when primes
  were masked so that subjects were unable to
  consciously identify them.
• They conclude that the N400 does not reflect
  post-lexical integration

42

• If the current results are on the right track
• It is consistent with the model suggested in
  Sturt 2003 for anaphora processing (Binding
  principle A), where he found that inaccessible
  antecedents have a late effect on the reflexives.
• Grammatical principles constrain the initial
  parsing of NPIs
• While encountering an unlicensed NPI, the parser
  could initiate task strategies to pick up some
  available but structurally inaccessible
  semantic/pragmatic inferences. This kind of
  errors might arise in the later stage of the
  reading task, but even more so when subjects are
  in a judging mode.

43

• The consistent lack of the frequency effect in
  both offline and online tasks supports the
  possibility that parsing NPIs is not simply
  building relations between two items. A parser
  that makes use of complex semantic composition is
  more plausible.

44
Conclusions

• Reading time data receives no boost from a
  structurally inaccessible but linearly preceding
  licensor, suggesting that structural constraints
  are respected.
• Offline judgment data shows some spurious
  licensing effect. Further investigations are
  needed to confirm if the interference effect
  occurs online.
• Preliminary evidence suggests that complicated
  semantic composition is established online very
  rapidly.

45
Acknowledgments

• Many thanks to
• Rebecca Baier
• Michael Israel
• Ed Kenschaft
• Ellen Lau
• Matt Wagers
• NSF BCS-0196004