Introduction to Psycholinguistics

Lecture 3: Sentence Processing

Matthew W Crocker

Computerlinguistik
Universität des Saarlandes

What makes up Psycholinguistics?

 Computational models of the representations, architectures and
mechanisms that underlie human language processing.
 Linguistics:
 Wow people represent linguistic knowledge
 Psycholinguistics:
 How people use this knowledge to produce and understand language
 Computational Models:
 Implementations of psycholinguistic theory, using linguistic representations.
 More complete theory: models and predicts human behaviour
 Experiments & Data:
 Descriptive studies of human language processing
 Off-line: grammaticality judgement, completions, global reading time (?)
 On-line: “word-by-word”, self-paced reading, eye-tracking, ERP
 Test prediction of theories and computer models
 Corpus data: word-frequency, category and sense bias, subcategorization

© Matthew W. Crocker Introduction to Psycholinguistics 2

1
Sentence Processing
 Sentence processing is the means by which the words of an utterance
are combined to yield and interpretation
 All people do it well
 It is a difficult task: complexity and ambiguity
 Not simple ‘retrieval’, like lexical access
 Compositional: interpretation must be built, rapidly, even for novel
word/structure input

 What are the architectures, mechanisms and representations

underlying this process?
 Architectures: modularity vs. interaction
 Mechanisms: how is input mapped to interpretations using knowledge
 Representations: How is knowledge encoded

© Matthew W. Crocker Introduction to Psycholinguistics 3

A Simple Theory of Grammar

The Grammar The Lexicon
S NP VP  Det = {the, a, every}
 NP PN  N = {man, woman, book,
 NP Det N hill, telescope}
 NP NP PP  PN = {John, Mary}
 PP P NP  P = {on, with}
 VP V  V = {saw, put, open,
 VP V NP read, reads}
 VP V NP PP

© Matthew W. Crocker Introduction to Psycholinguistics 4

2
A Generated Sentence
 the man read every book
S
ei S NP VP
NP VP
ty ru NP Det N VP V NP
Det N V NP
g g g tu NP Det N
the man read Det N
g g
every book

© Matthew W. Crocker Introduction to Psycholinguistics 5

Semantic Composition
 Theories of meaning and knowledge representation
 Semantic composition:
 lexical competence + semantic operations
 “the man read every book”
S ∀x book ( x) ♦ read ( j , x)
ei
NP
j VP λ x ∀y book ( y) ♦ read ( x, y)
ty ru
Det N V NP λ P ∀y book ( y) ♦ P( y )

g g g tu
the man read Det N
man
λ Pιx P(x ) g g
every book
λ Pλ Q ∀y P( y ) ♦ Q( y ) book

Introduction to Psycholinguistics 6

3
Semantic Ambiguity
 Word sense ambiguity: a word may have more than one SENSE
 Stock (soup vs investments), bank (money repository vs side of river),

 Anaphoric underspecification: anaphoric expressions derive their

meaning from context:
 She, it, the book, every man

 Scope ambiguity: the interpretation of sentence constituents may be

controlled by the interpretation of other constituents
 “In New York, a man is mugged every 10 minutes”

© Matthew W. Crocker Introduction to Psycholinguistics 7

Syntactic Ambiguity
 Lexical ambiguity occurs when a word may be rewritten by more than
one category:

 N {saw, hammer, book ...}

 V {read, saw, witnessed, ...}

 Structural ambiguity occurs when a sentence may be generated in

more than one way by the PS rules:

The spy saw the cop with the gun/binoculars

 The gun is usually interpreted as a modifier of the cop

 The binoculars is usually interpreted as an instrument of saw

© Matthew W. Crocker Introduction to Psycholinguistics 8

4
PP Attachment Ambiguity
 PPs may attached to NPs, and VPs:
 John saw the man with the telescope
S
ei
NP VP
g ry
PN V NP PP
John saw ty tu
Det N P NP
the man with the telescope

© Matthew W. Crocker Introduction to Psycholinguistics 9

Local Ambiguity: NP/S complements

 Local ambiguity occurs during incremental parsing, when there is
insufficient local information to determine the correct structure:

S
ei When the NP is
encountered we don’t
NP VP know whether to attach it
6 rp as direct object, or
embedded subject
The athlete V NP or S
g g ro
realised NP VP
6 6
his goals ... were unattainable

© Matthew W. Crocker Introduction to Psycholinguistics 10

5
Processing and Ambiguity
 What if an utterance may be interpreted in more than one way:
“I saw the man on the bench in the park with a telescope”
 choose a singe parse/interpretation
 determine all possible interpretations

 Incrementality
 Local ambiguity:
“I knew the solution to the problem was incorrect”
 How much is interpreted incrementally?
 What are the implications for parsing and understanding?

 Most evidence suggest people …

 … are consciously aware of only one interpretation at any time, and
 … construct interpretations on a word-by-word basis (at least!)

© Matthew W. Crocker Introduction to Psycholinguistics 11

Mechanisms for syntactic processing

 Human syntactic processing requires a solution to the problem of local
and global ambiguity

 Serial/backtracking:
 Initial disambiguation: rule (or structure?) selection strategy
 Reanalysis: reparsing? parse repair? …
 Parallel:
 Preferences: ranking strategy
 Limitations: what structures to forget
 Reanalysis: reranking/adjusting
 Parsing/Ranking strategies:
 Structural, syntactic
 Interactive: semantics, discourse, ...
 Probabilistic

© Matthew W. Crocker Introduction to Psycholinguistics 12

6
From Theory to Data
 We want to understand, and ideally model, sentence comprehension
 Organisations, mechanisms, representations, acquisition, interaction …

 What methods can we use to get at these issues:

 Introspection is notoriously unreliable
 Direct: Neuroscientific methods are not (yet!) that revealing …
 Behaviour: Rather we focus on observed behaviour
 Judgements on meaning or grammaticality
 Unconscious measure, e.g. reading times, priming …

 Why do we focus on ambiguous or “pathological” sentences?

 Ambiguity is more common than you think, yet usually not problematic
 Understanding how people cope with ambiguity can reveal a lot about the
underlying architectures and mechanisms
 Easier to investigate minimal pairs: similar sentences, with one difference

© Matthew W. Crocker Introduction to Psycholinguistics 13

Linking Hypotheses
 Linking Hypothesis:
 Need to relate the theory to some observed measure
 Typically impossible to predict measures completely

 Common view: Theories of parsing typically determine …

 which information sources are used when
 which representation is preferred/constructed when ambiguity arises
 If the sentence is consistent with that representation, processing should
be easier than if it is not
 Preferred sentences should have faster reading times in the
disambiguating region than dispreferred

 Don’t overfit the data: remember,our theories are about sentence

comprehension, not reading times …
 Must explain competence, and possible other behavioural measures

© Matthew W. Crocker Introduction to Psycholinguistics 14

7
Evidence from reading times
 Language comprehension entails the incremental recovery of an
interpretation for an utterance/sentence:
 Grammar, lexicon, parser, semantics, world knowledge, situation
S S
ei ei
NP VP NP VP
6 ri 6 rp
The athlete V NP The athlete V S
g 6 g ro
realised his goals … realised NP VP
6 6
his goals were out of reach
his shoes were out of reach
 Ambiguity: two possible structures
 How do we know which one people build first?
 Reading times increase when disambiguated towards the dispreferred
interpretation.
Pickering, Traxler & Crocker, 2000
© Matthew W. Crocker Introduction to Psycholinguistics 15

The Modularity Issue

 What is the architecture of the mind and brain?

 How is computation achieved/organised?
Understanding
Signal

Syntax
Lexicon

© Matthew W. Crocker Introduction to Psycholinguistics 16

8
Architectures and Mechanisms
 Are there distinct modules within the human language processor?
 What does “distinct” mean?
 Representational autonomy: e.g. parse trees vs conceptual representations
 Possibly shared procedures
 Procedural autonomy: e.g. parser versus interpreter
 Possibly shared representations

 If so…
 How are any such “distinct subsystems” for language processing organised?
 How do they interact?

 How does the architecture affect possible mechanisms?

 Serial (backtracking), parallel (bounded?), underspecified …
 Kinds of ambiguity resolution strategies?
 What are the implications for semantic processing?

 What are the arguments for and against ‘modularity’?

 theoretical, computational and empirical

© Matthew W. Crocker Introduction to Psycholinguistics 17

Towards a theory of parsing

 Syntax mediates the mapping sound to meaning
 If syntax exists, the construction of syntactic representations must precede
semantics or be part of the same system.

 The construction of syntactic dependencies and semantic

interpretations occurs incrementally, word-by-word.

 Little evidence of “conscious parallelism”

 Ambiguity and incrementality entail making decisions and building

interpretations in the face of “uncertainty”:
 What kinds of mechanisms are used to deal with ambiguity?
 What kinds of linguistic knowledge inform the decision making process?
 What does this tell us about the architecture of the sentence processor?

© Matthew W. Crocker Introduction to Psycholinguistics 18

9
A Modular Model

saw(man, …)

S
Semantics tu
NP VP
Lexical Access

ty g
Det N V
Syntactic Parsing the man saw

Det N V ...
Category Disambig
the man saw ...

© Matthew W. Crocker Introduction to Psycholinguistics 19

The Garden Path Theory

 Parser operates incrementally:
 Each word it attached into the Current Partial Phrase Marker
 The parser operates serially:
 A “race” to find an analysis: first wins
 Reanalyse if the analysis is thematically impossible

 Ambiguity resolution strategies:

 Minimal Attachment: Adopt the analysis which requires postulating the
fewest nodes
 Late Closure: Attach material into the most recently constructed phrase
marker
 Active Filler Strategy: Associate fillers with possible gaps (traces) as early
as possible.

© Matthew W. Crocker Introduction to Psycholinguistics 20

10
Minimal Attachment: VP Attachment
 John saw the man with the telescope
S
ep
NP VP
g qgp
PN V NP PP
John saw 2 tu
Det N P NP
the man with the telescope
S
ei
NP VP
g 3
PN V NP
John saw 3
NP PP
2 tu
Det N P NP
the man with the telescope

© Matthew W. Crocker Introduction to Psycholinguistics 21

Late Closure
 Prefer ‘low attachment’
S
ei
NP VP
6 ru
The reporter V S
g to
said NP VP
5 5 AdvP
the plane crashed 5
last night

© Matthew W. Crocker Introduction to Psycholinguistics 22

11
NP/S Complement Ambiguity
 The student knew the solution to the problem.
 The student knew the solution was incorrect.

S
S ei
ei NP VP
NP VP 6 ru
6 ru The student V S
The student V NP g ro
g 6 knew NP VP
knew the solution to ... 6 6
the solution was incorrect

© Matthew W. Crocker Introduction to Psycholinguistics 23

Ambiguities revisited: [preferred/dis-preferred]

 NP/VP Attachment Ambiguity:
 “The cop [saw [the burglar] [with the binoculars]]”
 “The cop saw [the burglar [with the gun]]”
 NP/S Complement Attachment Ambiguity:
 “The athlete [realised [his goal]] last week”
 “The athlete realised [[his shoes] were across the room]”
 Clause-boundary Ambiguity:
 “Since Jay always [jogs [a mile]] the race doesn’t seem very long”
 “Since Jay always jogs [[a mile] doesn’t seem very long]”
 Red. Relative-Main Clause Ambiguity:
 “[The woman [delivered the junkmail on Thursdays]]”
 “[[The woman [delivered the junkmail]] threw it away]”
 Relative/Complement Clause Ambiguity:
 “The doctor [told [the woman [that he was in love with]] [to leave]]”
 “The doctor [told [the woman] [that he was in love with her]]”
© Matthew W. Crocker Introduction to Psycholinguistics 24

12