Models of Sentence Processing Flashcards
Frazier&Rayner’s Garden Path Model (1982)
2-stage parser in which initial attachments are made purely on the basis of syntactic information
- Stage I: Initial attachments made on the basis of the principles of Minimal Attachment and Late Closure (with MA taking precedence)
- Stage II: Initial attachments confirmed or revised on the basis of semantic and pragmatic information generated by an independent thematic processor
Minimal Attachment
parser always tries to make the simplest attachment possible
Simplest’ means the attachment that requires the creation of the fewest additional nodes
Attachment arises as the natural outcome of a race between alternative sets of rules
-Compared eye movements for Ss where Minimal Attachment should/should not result in garden path
-S1 - The criminal confessed that his sins harmed many people
(‘that’ forces parser to treat NP ‘his sins’ as Subject of Complement ‘harmed many people’)
- S2- The criminal confessed his sins harmed many people
- Minimal Attachment leads parser to treat Noun Phrase ‘his sins’ as a Direct Object of the Verb ‘confessed’ and not the Subject of Complement ‘harmed many people’
GPM predicts reading of S1 straightforward
S2 should result in temporary garden path and blip when reader reaches ‘harmed’ (disambiguating region)
The criminal [confessed his sins] harmed (?) many people – This was true
Late Closure
-Compared reading times for sentences consistent/inconsistent with Late Closure
-Since Jay always jogs a mile and a half this seems a short distance to him
Requires reader to keep VP ‘jogs …’ open and attach NP ‘a mile and a half’
-Since Jay always jogs a mile and a half seems a short distance to him
Requires reader to close VP after ‘jogs’
-GPM predicts reading of S1 straightforward
-S2 should result in temporary garden path and blip when reader reaches ‘seems’ (disambiguating region)
Since Jay always [jogs a mile and a half] seems (?) a short distance to him
-F & R found reading times longer for S2 than S1
-F & R found first fixation in disambiguating region longer for S2 than S1
Garden Path Model Evidence
Ferreira &Clifton (1986)
- Compared GP sentences where MA reading was or was not semantically anomalous
- The defendant examined … by the lawyer turned out to be unreliable
- The evidence examined … by the lawyer turned out to be unreliable
- GPM predicts ‘examined’ will be treated as Main Verb versus Verb of reduced relative in both cases
- Eye-movement data showed that in both cases subjects had to backtrack when reached ‘by’ (i.e. had initially treated ‘examined’ as a main verb) – We follow the garden path rule even when it makes no sense – Evidence cannot examine!!
Mitchell (1987)
- Compared self-paced reading times on GPs where Late Closure reading was/was not anomalous
- After the child had visited the doctor … prescribed a course of injections
- After the child had sneezed the doctor … prescribed a course of injections
- GPM predicts ‘the doctor’ will be treated as Direct Object in both cases
- Results showed that in both cases subjects had problems when reached ‘prescribed’ (i.e. had initially treated ‘the doctor’ as the Direct Object) – Following late closure even when the sentence they are constructing makes no sense
Breedin and Saffran (1999)
- Evidence from neuroscience suggests that semantic and syntactic processing are independent
- DM, who had a significant and pervasive loss of semantic knowledge as a result of dementia.
- Yet his semantic deficit had no apparent effect on his syntactic abilities. He performed extremely well at detecting grammatical violations (e.g., he knew that “what did the exhausted young woman sit?” was ungrammatical).
Garden Path Model
Against
GPM research tends to make the unrealistic assumption that readers don’t have prior expectations that influence how they behave in parsing experiments, whereas CBM research takes these expectations seriously and shows that difficulties that appear to reflect parsing principles like MA can actually be understood as reflecting the prior expectations that participants bring with them into the experiment.
Crain&Steedman (1985)
The horse raced past the barn quickly
The horse raced past the barn fell
-Without any preceding context, natural assumption is there is only one horse
-This assumption consistent with S1, but not S2
Proponents of the constraint-based models argue that researchers favouring the garden path model use techniques that are not sensitive enough to detect the interactions involved, or that the non-syntactic constraints used are too weak.
Constraint-based models of parsing
Sentence interpretation the product of interaction between multiple constraints/sources of information
- (i.e. syntactic, semantic, pragmatic and frequency - based information)
- Current interpretation one most strongly supported by these multiple constraints and hence most active
- Garden paths occur when the correct analysis of a local ambiguity receives little activation
Constraint-based models of parsing
Evidence
Altmann& Steedman (1988) - contextual
- Measured reading times on sentences like:
- The burglar blew open the safe with the dynamite and made off with the loot
- The burglar blew open the safe with the new lock and made off with the loot
- GPM predicts you should be good at interpreting the 1st sentence
- Manipulated prior context to mention 1 or 2 safes
- If 1 safe, S1 easier because information in PP of S2 is redundant and so unexpected
- If 2 safes, S2 easier because information in PP of S2 is necessary to specify which safe
- The context biases them towards expecting information about the safe, people follow minimal attachment when there is one safe but do the opposite when there is two.
- Contextual information influences how you process a sentence
Trueswell, Tanenhaus & Garnsey (1994) - Word meaning
-Compared reading times for GPs where animacy/inanimacy of initial NP implied main verb/reduced relative reading
-The book read … by the student was very difficult to understand
-GPM predicts GP effect, however people were not surprised by ‘by’
Strong semantic information was protecting people from being garden pathed
-Direct evidence against the garden path model.
Trueswell, Tanenhaus & Kello (1993)
- Compared self-paced reading times and eye-movement data for GPs with verbs for which direct object reading more or less likely
- The general revealed the weapon … was ready to be used
- The general pretended the weapon … was ready to be used
- Garden Path M predicts ‘the weapon’ will be treated as Direct Object in both cases
- Results showed subjects only garden pathed by first kind of sentence
- Pretend tends to come in sentences which don’t have a minimal attachment structure to pretect them from being garden pathed by minimal attachment – they are not obeying minimal attachment because of what they know about the word pretend – pretend tends to take place in certain structures and not others
- The frequency in which words pop up in certain structures means the word is used to predict the sentence structure – so you are not just following the garden path
Mason, Just, Keller, & Carpenter, 2003
-workload was spread among the superior temporal gyrus (including Wernicke’s area) and the inferior frontal gyrus (including Broca’s area), hinting that multiple processes are involved in ambiguity resolution. In particular, Broca’s area might be involved in generating abstract syntactic frames, and Wernicke’s in interpreting and elaborating them with semantic information
Constraint-based models of parsing
Issues
GPM makes strong predictions which are readily falsifiable, whereas the CBM can be adapted to explain almost any result and is hence very difficult to falsify
Proponents of the garden path model argue that the effects that are claimed arise because the second stage of parsing begins very quickly, and that many experiments that are supposed to be looking at the first stage are in fact looking at the second stage of parsing
Why do subjects sometimes appear to prefer a particular interpretation in spite of the preceding semantic and pragmatic context? – Follow garden path even when the sentences are crazy
Van Gompel and Pickering (2001) “After the child had sneezed the doctor … prescribed a course of injections” eye-movement methodology readers experience difficulty after “sneezed.” suggest that the first stage of parsing is short-sighted and does not use semantic or thematic information.
Part of the problem is that the information that you would use to prevent being garden pathed is quite slow to come online, not because you can’t use it but that it takes time to compute. Reading takes time to work out the meaning and words used are low frequency so the semantic might not be that strong. What about listening? Can give the context in front of them – immediately available.
Tanenhaus, Spivey-Knowlton Eberhard & Sedivy (1995)
- In reading experiments, context not immediately available and so has to be retrieved from memory
- Put the apple on the towel in the box
- GPM predicts that ‘on the towel’ should be interpreted as the destination of the apple until the subjects hear ‘in the box’
- When subjects hear ‘on the towel’ they should look at the empty towel
- When subjects hear ‘in the box’ they should look back at the apple on the towel
- Eye movement data suggest ‘on the towel’ interpreted in line with GPM in one referent condition
- In the two referent condition subjects hardly ever looked at the empty towel
- Provides strong evidence that context can be used immediately if available (i.e. against GPM)
Autonomous and Interactive Models
How modular is the human sentence processing mechanism?
Are initial attachments made only on the basis of syntactic information? Or are they influenced by semantic and contextual factors?
Do you process meaning from words and the structure and grammar at the same time? Or just from grammar?
How do we process temporary structural ambiguities?
The criminal confessed his sins harmed many people
Can semantic and pragmatic information prevent us from being garden pathed?
Or are they just used to guide subsequent reanalysis?
