Sentence Processing Flashcards
difficulty in sentence comprehension
- Sentences are infinitely variable
• Can’t rely on recognition
• Sentence comprehension is a constructive process that relies on domain general processes eg working memory, inferential reasoning - Grammar and syntax
• Rules about how words are combined to form sentences eg Phrase structure grammar*
• Sentence processing difficulties provide insight into normal sentence processing
theories of sentence parsing
central question: which comes first - syntax or semantics?
- syntax first: garden-path mode
- semantics first: good-enough processing
- parallel syntactic and semantic analysis: multiple constraints model
local syntactic ambiguity
“Time flies like an arrow” vs “Fruit flies like a banana”
Multiple possible syntactic categories activated, e.g.
– time = V or N?
– flies = V or N?
– like = V or adverb (describing the timing, describing the flying)?
➔Multiple possible phrase structure trees can be constructed
• Real world knowledge and probabilities can sometimes disambiguate meanings:
My son has grown another foot
and sometimes not:
Visiting relatives can be boring
global syntactic ambiguity
“garden path” sentences
The horse raced past the barn fell
Garden path: initial syntactic analysis (“parse”) is incorrect so we need to revise syntactic structure, i.e. the beginning of the sentence “leads you down the garden
path” to the wrong meaning
grammatically correct but violates phrase structure tree
- The beginning of sentence is consistent with more than one syntactic analysis (i.e., more than one parsing is possible)
- BUT the main verb is actually ‘fell’; ‘raced’ is part of a ‘reduced relative clause’ describing the horse
- can often be disambiguated by using prosody, punctuation or wording to mark the reduced relative clause
syntax first: garden-path model
Comprehenders pursue a single interpretation at a time (try to build a phrase structure tree), based on syntactic parsing only, not meaning
–> encapsulated, autonomous syntactic language module
Processing limited by attentional resources: construct the simplest tree structure
➔the one that minimises working memory load
Achieved by applying:
• Late closure heuristic
Continue to build one phrase as long as possible
e.g. Since Jay always jogs a mile seems like a short distance to him.
• Minimal attachment heuristic
Build the simplest phrase structure tree: attach a word to current clause if possible
e.g. The spy saw the cop with the binoculars.
IF initial parse yields an incoherent semantic interpretation,
➔re-analyse to construct alternative syntactic structure
evidence for GPM
Eye movement studies of sentence reading • The length of time people fixate on a word indexes amount of processing required
Eye movements in Garden Path sentences (Frazier & Rayner,1982)
(1) Since Jay always jogs a mile seems like a short distance to him.
(2) Since Jay always jogs a mile this seems like a short distance to him.
• Longer fixations on seems in Sentence 1 than Sentence 2 BUT fixations on this in Sentence 2 not delayed
➔consistent with late closure heuristic: re-analysis required for Sentence 1 but not Sentence 2
• No disruption for ‘this’: ambiguity alone does not affect processing
➔People do not compute all possible meanings
predictions of GPM
- Parsing is serial and incremental: Syntactic role is assigned word by word
Cross-modal priming:
(1) If you walk too near the runway landing planes IS/ARE….
(2) If you’ve been trained as a pilot landing planes IS/ARE….
➔Faster to ARE for (1): landing = adjective eg landing planes are dangerous
➔Faster to IS for (2): landing = verb eg landing planes is easy - Parsing is autonomous: when ambiguity encountered, a single interpretation is selected based on syntax, not meaning
➔Context/lexical meaning should not affect GP effects BUT Eye movement data
(1) The defendant examined by the lawyer turned out to be unreliable
(2) The evidence examined by the lawyer turned out to be unreliable
Identical syntactic structure but only (1) causes ‘garden path’
➔Semantic context influences parsing
➔Challenges claim that syntactic parsing is autonomous and occurs before semantic processing
good enough comprehension
Garden path model assumes that
comprehenders always generate a complete, accurate representation of meaning
BUT Processing sometimes faster for ambiguous than unambiguous sentence
➔Comprehenders do not always resolve
ambiguities
➔ Goal of comprehension is a representation that is ‘good enough’ for current task goals
swets et al (2008): who did the scratching?
Who did the scratching?
Ambiguous:
The maid of the princess who scratched herself in public was terribly humiliated.
N1 attachment:
The son of the princess who scratched himself in public was terribly humiliated.
N2 attachment:
The son of the princess who scratched herself in public was terribly humiliated
Manipulated comprehension demands:
• Questions about relative clause after every sentence eg ‘Who scratched…?’
• Superficial questions after every sentence eg ‘Was anyone humiliated?’
• Superficial questions after ~10% of sentences
–>people take longer for comprehension question in N1 attachment, also different attachment patterns (N1 longer in compre question, no difference in superficial question)
–> ambiguous condition faster in superficial questions
➔People adopt a ‘good enough’ strategy UNLESS full comprehension required for task
Multiple constraint model
Parsing is parallel and interactive
‘Connectionist’ framework (cf TRACE model)
extended to semantic and world knowledge
• Lexical knowledge includes syntactic information eg word class; verb biases
• Early activation of contextual/ semantic knowledge
• Interpretations compete for selection through competitive mechanisms similar to TRACE model
➔Multiple possible ‘parses’ of sentence are
evaluated simultaneously
➔Comprehension is a constructive process that requires integration of multiple knowledge sources
evidence: semantic vs ‘world knowledge’
Hagoort et al., 2004
semantic vs world knowledge N400 effect
correct: the dutch trains are yellow and very crowded
world knowledge violation: the dutch trains are white and very crowded
semantic violation: the dutch trains are sour and very crowded
finding:
- world knowledge violation as disruptive in N400 as semantic violation
- activating world knowledge early in processing in same area as semantic violation
- same mechanism in treating both kinds of knowledge
semantic vs discourse context
peanut salted/in love study
stronger N400 to salted (correct predicate, contextually disfavoured) than to in love
summary
Multiple word meanings activated; more than one syntactic analysis may be constructed
Sentence comprehension is derived from word meanings, context, syntactic analysis and depends on working memory resources
Alternative parses held in working memory to allow revision of syntactic structure in case of ambiguity
Comprehension goals influence criteria for ‘good enough’ comprehension
Similar processes involved in spoken and written language comprehension