L1 - Language I, an introduction to computational models of speech production Flashcards
What is language?
An exchange of information
Where does language processing occur?
Under the level of conscious awareness.
Why do humanities lecturers have more hesitations in their speech compared to science lecturers?
Words in the lexicon compete for selection, and humanities lecturers have more words they can use whereas scientists don’t. There is more spreading activation due to more overlap in the humanities lexicon. Competition slows down speech.
What is tip of the tongue state?
a state in which one cannot quite recall a familiar word but can recall words of similar form and meaning.
What can we tell about speech processing from evidence of tip-of-the-tongue-state?
A concept is selected for production before the phonemes used to say it.
What did Vigliocco et al 1997 find?
Found that Italian speakers in tip-of-the-tongue states know the grammatical gender of the word but cannot access the phonological form of the word. This suggests that sentences are planned and structured in a particular order.
What are the three specific levels of processing which are required to plan and articulate an utterance?
1) Semantic processing
2) Syntactic and morphemic processing
3) Articulation (form processing)
Where do speech errors tend to occur?
Within specific processes.
What do speech errors allow us to do?
To make inferences about speech processing
What happens with errors within semantic processing?
Conceptualisation blend errors - blending semantic related words together.
What happens with errors within syntactic and morphemic processing?
Switching morphemes - e.g slicly thinned for thinly sliced.
What happens with errors within form procssing?
Articulation Word and phoneme exchange - forks of a prong
What have observations of speech errors highlighted?
The processes that are required to plan and execute speech
Describe Fromkin’s model of utterance generation 1971
Box and arrow model, discrete model - each stage must happen in a particular order and they do not overlap.
What is the first stage of Fromkin’s utterance generation?
Meaning is generated
What is the second stage of Fromkin’s utterance generation?
Syntactic structure is generated and associated with semantic features.
What is the 3rd stage of Fromkin’s model of utterance generation?
Position of intonation/stress is planned - stress on relevant sounds in sentence.
What is the fourth stage of Fromkin’s utterance generation?
Lexicon look up - finds words and generates phonological segments - allow us to convey message
What is the fifth stage of Fromkin’s model of utterance generation?
Morphophonemic constraints are added, and phonemes selected for utterance.
How does Fromkin’s model accommodate word exchange errors?
Switching words around - accidentally putting one word in the wrong section. Utterance loses meaning.
What does Dell’s model (a spreading activation theory of retrieval in sentence production) assume?
Assumes processes are fluid, information can go back and forth through the language production processes.
What does Levelt’s model (A theory of lexical access in speech production) assume?
Assumes processes are serially organised and must complete 1 before the other begins - similar to Fromkin.
What is Fromkin’s model known as and why?
Box and arrow model, because it was not implemented as a computational model
What were Dell and Levelt’s models known as and why?
Computational models as they were implemented as functioning computational models.
What do box and arrow models assume about processing?
That processing is occurring in a particular order for behaviour to occur in response to stimulus.
What is a model?
A simplified representation of a thing
What is a statistical model?
A mathematical relationship between variables, that hold under specific assumptions
What are cognitive box and arrow models?
Models that describe the relationships between different mental processes, under the assumption that the mind operates like multi-staged information processing machines.
What are formal cognitive models?
Input –> then usually a mathematical description of the relationship between mental processes usually expressed through computer code –> output (behaviour)
What do nodes represent in neural networks?
Groups of neurons.
Activation of nodes..?
Spreads activation across the network.
Describe Dell’s spreading activation theory?
Interactive and cascading model - information is not restrained in the order in which it has to happen. Based on spreading activation.
What are the three levels in Dell’s spreading activation theory?
Semantic level, Syntax + morphology level, phoneme level.
How does spreading activation work in Dell’s model?
Occurs simultaneously - can go up and down via spreading activation. Some processes receive more activation than others.
Describe spreading activation on a semantic level
Words that are similar
Describe spreading activation on a phonemic level.
Words that sound similar
Describe the assumption about processing levels in Dell’s spreading activation theory
Semantic level information is processed at the same time as phonemic level
Describe Dell’s spreading activation theory
Spreading activation activates concepts related to the item, the concept with the most relevance receives most activation.
Competition between words and concepts.
Processing items that overlap in semantics and/or phonology will result in higher error rates than processing items that do not overlap.
What did participants have to do in Ferreira and Griffin’s experiment?
Participants named target pictures after first reading sentences that primed a semantic competitor, a semantic and phonological competitor or an unrelated word.
What does Ferreira and Griffin’s experiment test?
Semantic and phonological competitors.
We have words that sound the same but have different meaning. Use these to test how we process it. If levels interact then the press flows from meaning and sound up and down.
Primed with sentence in competition with picture they are naming.
Primed to think about ‘nun’ when presented semantic competitor sentence.
Presented with a picture of a priest but then the sentence I thought there would be some cookies left, but there were ______ semantically not related to the image.
Results in more errors as both semantic and phonological errors occur at once.
Results of FERREIRA and Griffin
When attempting to name the target picture “priest”
participants were more likely to mistakenly produce the semantic or phonological + semantic competitor /nʌn/ than they were to produce an unrelated word (e.g., match)
20% of errors when represented with semantic alone. Tells us predictions of Dell’s models. Both of these occur simultaneously and they are active at the same time.
What does the evidence from F&G demonstrate??
Evidence from Ferreira and Griffin (2003) demonstrates that semantic and phonological similarity can result in speech errors suggesting that processing is fluid and interactive.
There is communications between different levels of speech processing.
Dell interim conclusion
The model provides a convincing account of speech errors
Evidence from lab induced speech errors supports the suggestion that processing is interactive.
However some suggest that the parallel processing in the Dell model would predict that we make far more speech errors than we actually do
What is a lemma?
The word form and the syntactic information that links with the word form
What is Levelt et al’s model?
The WEAVER++ model. It contains an extra level called the lemma
What is a lemma?
The word form and the syntactic information that links with the word form.
It contains semantic (meaning) and syntactic (grammatical) information about a word but not its phonological (sound) form.
In models of language production, such as those proposed by Levelt, a lemma is activated during the stage of grammatical encoding. Once the speaker has chosen the concept they want to express, they select the corresponding lemma, which helps in determining the word’s grammatical features (e.g., tense, number, gender) without yet specifying its sounds. After the lemma is selected, the phonological form of the word is retrieved, allowing the speaker to articulate it.
If you want to say “run” in a sentence like “I run every day,” the lemma for “run” would include the concept of the action and its grammatical features (present tense, verb), but not yet the phonological representation (/rʌn/).
What is different about spreading activation in Levelt’s model as opposed to Dell’s?
Spreading activation still occurs, but once the concept has been selected, inhibitory signals are sent out from the node to other relevant concepts and semantics are deactivated, preventing them from being carried forward. This level of inhibition is not present in Dell’s model.
Describe the theory and predictions for both Levelt et al and Dell.
Levelt et al - Processing items which are semantically similar will be inhibitve
Dell - due to the lack of inhibitory connections processing items which are semantically similar will not be inhibitive.
What did the Wheeldon and Monsell experiment make participants do?
To make sure the experiments tested speech production, participants were asked to
1) answer a question out loud
2) name a picture
The answer to the question was either semantically elated to the picture or unrelated. If related to the previous concept, activation of the word would be slower.
What were the results of Wheeldon and Monsell’s experiment?
Graph shows it takes much longer to recall a word semantically related to an image as it has been inhibited, than a semantically unrelated item. Suggest speech production is discrete.
Speech errors in Levelt et al’s model
WEAVER++ assumes speech errors are very rare. Doesn’t account for errors. But errors may occur if the wrong concept is selected and processed to the phonological level.
Evidence to support existence of te lemma -
Tip of the tongue state. In the lemma is the grammatical gender of the word. Have syntactic info in lemma but not the rest.
How did Dell test spreading activation and interaction occurring at all levels? (Producing the word doll)
Tested by having two items on a screen. Target item outline in green, distractor outline in red. Instructed to name green and ignore red.
In Dell’s model, predicts both dog and dol became activated at the same time, down to level of phonology. Both share initial starting phoneme - shared activation. But with bat there is no semantic or phonological overlap,
What did Meyer and Damian find out about Semantic interference?
We see naming is quicker when phonology overlaps - as it is activated twice meaning we have faster access to it. Activating the same phonology means we have access to them quicker.
Under WEAVER++ this shouldn’t happen and phonological relatedness shouldn’t have an impact.
