Speech processing and higher level sentence processing Flashcards
Allopenna, Magnuson & Tanenhaus (1998)
Eye tracking: Where do participants look as you are telling them
model predictions
The COHORT model would not predict any fixations on the speaker
The TRACE model does, and it even gets the timing right!
what is parsing
generating a deep structure based on the surface structure (i.e. the words you have identified and the order in which they occurred).
In spoken language, prosodic cues may help.
how does the parser work
The meaning of a sentence can only be fully understood after the syntax of a sentence is parsed
what are the 4 parsing models
Garden-path model (Frazier & Rayner, 1982)
Constraint-satisfaction model (McDonald et al., 1994)
Unrestricted race model (van Gompel et al., 2000)
Good-enough processing (Ferreira et al., 2002)
what is syntactic ambiguity
Garden path sentences
While Mary was mending the sock fell off her lap.
Temporarily ambiguous: Readers tend to think that “Mary was mending the sock” is one clause
At “fell off”, readers tend to:
Boggle and fixate for a long time
Regress back to earlier in the text
Read to end, and then start over again
what is the Garden-path model (Frazier & Rayner, 1982)
Just one possibility is considered based on grammatical rules only, sentence has to be re-parsed if it’s wrong.
what is the Constraint-satisfaction model (McDonald et al., 1994)
Multiple possibilities are considered, but they aren’t activated equally
Semantic properties and frequencies of verbs in syntactic structures are taken into account
what is the Unrestricted race model (van Gompel et al., 2000)
Similar to garden-path model, but structures are considered based on both semantic and syntactic evidence
what is the Good-enough processing model (Ferreira et al., 2002)
Compatible with any of the three above, but says that the result of parsing is not always well-elaborated or even correct
how does the Garden path model deal with ambuguity
The parser chooses one preferred interpretation
Which interpretation is preferred depends on syntactic rules:
Minimal attachment
Late closure
what is late closure
When possible, attach incoming lexical items into the clause or phrase currently being processed (i.e., the lowest possible nonterminal node dominating the last item analyzed).
Translation for non-linguists: While building your deep structure, don’t end a clause (e.g. the relative clause “Since Jay always jogs”) unless you really have to.
what is minimal attachment
Attach incoming material into the phrase-marker being constructed using the fewest nodes consistent with the well-formedness rules of the language.
Translation for non-linguists: Build the tree with the fewest nodes that is grammatical
strengths of garden path model
Fits a lot of evidence, e.g. from eye tracking
No evidence for the competing theories at the time:
Parallel parsing (all alternatives are considered at the same time)
Minimal commitment (the sentence structure isn’t built until you get to the end of the sentence)
A lot of the time, the principles of minimal attachment and late closure are applied even if they don’t make sense semantically:
e.g Readers take longer to read “After the child had sneezed the doctor prescribed a course of injections.”
weakness of garden path model
Listeners shouldn’t take context into account
Spivey, Tanenhaus, Eberhard, & Sedivy (2002)
Task: Look at an array of objects while listening to sentences referring to those objects
Spivey, Tanenhaus, Eberhard, & Sedivy (2002)
garden path model
Put the apple on the towel in the box.”
Listeners should first look at the apple (the referent) and then at the box (the goal).
If they attach “on the towel” incorrectly, they will interpret it as the goal and look at the towel instead of the box at first.
results of Spivey, Tanenhaus, Eberhard, & Sedivy (2002)
Listeners take the context into account:
If there is only one referent (apple), they attach “on the towel” as the goal
Lots of looks at the towel
If there is more than one referent (one apple on a towel and one apple on a napkin) they attach “on the towel” to the referent
Very few looks at the towel
what is the constraint satisfaction model
Connectionist model
The parser considers both semantic and syntactic information straight away
The most likely interpretation gets the most activation
But alternative interpretations cause inhibition and lower that activation
If the most likely interpretation fails, the resulting shift in activation takes some time to resolve
what are the strengths of the constraint satisfaction model
Can explain the effect of semantic processing.
Some verbs are biased towards specific syntactic structures
“Read” is based towards a direct object:
A. The professor read the newspaper had been destroyed.
B. The professor read the newspaper during his break.
This makes B easier to understand than A
“Believe” is normally followed by a subordinate clause
C. The professor believed the newspaper had been destroyed.
D. The professor believed the newspaper.
This makes C easier to understand than A.
Altmann & Kamide (1999)
Visual world paradigm eye tracking study
Where do people look while listening to a sentence?
Participants hear:
“The boy will move the cake” or
“The boy will eat the cake”
Eye tracking:
Will participants anticipate “cake” as a semantically appropriate object for “eat”?
Any of the objects are appropriate for the control verb “move”.
weaknesses of constraint satisfaction model
Maybe too general in its predictions
Maybe a garden-path model with a very fast semantic processing step after the syntactic processing can account for the results as well.
what is the unrestricted race model
van Gompel et al. (2000)
Unrestricted in terms of input:
Syntactic structure is built taking both syntactic and semantic information into account at the same time (like in the Constraint-Satisfaction model)
Initially, all interpretations are built in parallel (there is a race between the possible interpretations)
Once one interpretation wins the race, all other interpretations are ignored (like in the Garden-Path model; unlike in the Constraint-Satisfaction model)
If the preferred interpretation has to be abandoned, this will trigger an extensive reanalysis of the sentence (like in the Garden-Path model).
Unrestriced Race Model: Evidence
van Gompel et al. (2001): Participants read three types of sentences:
A. Globally ambiguous sentences: The burglar stabbed only the guy with the dagger during the night.
Meaning is not constrained by semantics: Both verb-phrase attachment (the burglar uses the dagger) and noun-phrase attachment (the guy has the dagger) are plausible.
B. Verb-phrase attachment: The burglar stabbed only the dog with the dagger during the night.
Syntactically ambiguous but disambiguated by semantics: dogs don’t have daggers.
C. Noun-phrase attachment: The burglar stabbed only the dog with the collar during the night.
Syntactically ambiguous but disambiguated by semantics: you can’t use collars to stab.
model predictions for van Gompel et al (2001)
Garden-path model:
All sentences are globally ambiguous syntactically, so they should take the same time to process
Constraint-satisfaction model:
In the sentences that are semantically constrained (B and C), it should be easier to choose the preferred interpretation, so they should be read faster than the unconstrained sentence (A)
Unrestricted race model:
Sentence A should be fastest since either of the two interpretations can be picked
Sentences B and C should be slower since people might occasionally have to re-analyse these sentences if they picked the wrong analysis initially
why does van Gompel et al support the unrestricted race model
The evidence (from van Gompel et al., 2001) matches the predictions of the Unrestricted Race Model
