Language Processing Flashcards
(26 cards)
SPITSYNA et al 2006
Listening and reading, although starting analysis using brain regions specific to these modalities (hearing and vision), have the info streams converging on the same area of the brain’s left temporal lobe (fMRI). this is more economical way of processing language.
SEGAERT et al 2013
Producing language (speaking) and comprehending language (listening) use the same brain structure - confirmed by fMRI. The language system is unitary.
PRE-LEXICAL MODEL OF SPEECH SEGMENTATION
Relies on characteristic of speech stream of where the boudary is likely to be. Silent gaps are not sufficient, as the ones within words are often longer than the ones between words. They can be useful though if they are long. Another cue comes from rhythm of speech. In English it’s a strong syllable followed by one or more weak syllables. Strong syllables are good landmarks in the speech stream & most words that carry meaning start with a strong syllable. Cutler & Norris conducted an experiment that showed English speakers use rhythm to identify word boudaries. But there are words that defy these rules (such as confess), which means pre-lexical cues do not always lead to correct boundary identification.
LEXICAL MODEL OF SPEECH SEGMENTATION
Rely on our knowledge of phonological representation of particular words, to guide segmentation. This means recognizing word sequentially and thus predicting new word at the boudary of the existing. However this will only work for longer words, where there is time to guess the ending of the word within the time span of their acoustic waveforms.
SAFFRAN et al (1996)
Segmentation mechanisms develop as a consequence of exposure to language. Saffran et al. devised an artificial language and presented 8-month old babies with stream of made up words, with no acoustic or rhythmic cues in it - from one loudspeaker, and jumbled up syllables from that language in another. After just two minutes of speech infants started turning their heads towards the speaker with jumbled up syllables, indicating that they’ve learnt statistical info about how the syllables should be arranged and preferred to listen to novel instances. Implicit learning ability.
PARALLEL ACTIVATION
Constant evaluation & re-evaluation of words against potential candidates for the identity - in order not to waste time! Advantage - it can lead to word identification before the whole word is heard.
The cohort model (Marsley-Wilson). As soon as the beginning of a word is heard, the word-initial cohort (all the words that match that beginning) is activated. As more of the word is heard, the list is narrowed down to finally reach uniqueness point when only one word will match.
MARSLEN-WILSON (1987)
The meaning of all likely candidates is accessed briefly during hearing the word. M-W used cross-modal priming and presented part. with spoken prime word, followed by a visual target word. Part. had to decide if the target was a real word asap. If the words were semantically connected, RT were faster. However even primes such as ‘confe’ which could have fitted with ‘confetti’ primed ‘wedding’ responses. This suggest that all the possible candidates are activated before the uniqueness point, in accordance with the cohort model.
TRACE MODEL
Connectionist model that assumes 3 levels of representation of speech: phonetic feature level (manner of production), the phoneme level and the word level.Speech changes patterns of activation on phonetic feature level, where phonemic speech representation is constructed. A word node has connections from all the phonemes within that word.
Secondly, nodes at word level are connected with inhibitory links, which helps select between active words. THe more the word node is activated (the more consistent all phonemes in the input are) the more it inhibits other potential candidates - the winner takes it all approach.
WORD SUPERIORITY EFFECT
Letters are easier to detect when they form part of a word. This could be explained by IAC (interactive activation and competition) model of visual word recognition. For example ‘i’ is easier to detect in ‘slim’ than in ‘spim’ because the word node for slim provides a secondary source of activation for recognition of ‘i’, whereas there is no secondary source for non-word spim (top-down feedback connections).
COLTHEART ET AL.
Dual-route model of reading, DRC. There is a rule-based route to pronunciation via a grapheme-phoneme system (works for regularly pronunced words) - assembled phonology, and a ‘lexical’ route, which relies on stored pronunciation. Regular words are named fasted than irregular, because they can be named via either of the routes, whereas irregular words need to use the lexical route. This advantage however is only there for low-frequency words. For high frequency words it is assumed that lexical route is just as quick as the rule-based route.
GLUSHKO
Properties of neighbouring (in spelling) words can influence speed of naming a word. Although have is a regular word, its neighbours have inconsistent pronunciation - wave, gave, save. The naming of this word is slower as a result of this inconsistency. Simple dual-route model cannot explain this.
VAN ORDEN
Phonological representations are involved in silent reading, even when it’s detrimental to performance. In a semantic categorization task, participants found it difficult to reject homophones to a category member. This suggests they were activating the pronunciation of the homophone. Same was found for pseudohomophones.
STARR & RAYNER 2001 offprint
Eye movement during reading is claimed to reflect high level cognitive processes. Studies of parafoveal information uptake (right of fixation) - part. shown a word on a screen with a window showing only a certain number of letters to the left and right of the fixation point, which moved with the gaze. The experimenter gradually increased the size of the window until normal reading speed was achieved. It meant that the text beyond the window range is not used for normal reading.This supports the parallel processing in reading. Frequency &predictability of a word has infuence on fixation times suggesting top-down processing - again supporting parallel processing.
MENTAL LEXICON
Where word information, such as phonology, morphology, semantics, orthography and syntactic role is stored. It also contains knowledge of associations between words.
MORPHEME
The smallest meaningful unit in a word. Discussion if morphemes have their own lexicon representation. Marslen-Wilson used priming to test that. If words are broken down into morphenes, there should be strong priming between words containing the same morpheme. They did find it for words that had some shared meaning, such as ‘cruel’ and ‘cruelty’, but not ‘casual’ and ‘casualty’. Morphological decomposition only occurs if there is some benefit to be had
SPREADING ACTIVATION MODEL
Model of how word meanings are represesented mentally. This model, similarly to IAC and TRACE assumes words are respresented by units/nodes. The links between the nodes here represent semantic relationship, such connecting ‘canary’ and ‘bird’ with a semantic link ‘is’.
FEATURAL THEORY
Theory of semantic representation - word meanings are represented as a set of semantic features or properties (a bit like concepts) associated with that word ie. features relevant to ‘canary’ would be ‘has wings’ ‘ can fly’ etc
SEMANTIC PRIMING
Way of investigating what type of semantic information is stored in mental lexicon. Relies on assumption that when there is a semantic relationship in a pair of words, response to target word will be quicker if the semantic relationship is represented in the mental lexicon. For example - pairs of associated words- 90% part. presented with word ‘cheddar’ replied ‘cheese’ as the first word that comes to their head. This suggests that associative links between words are represented in the lexicon is some way.
SEMANTIC AMBIGUITY
Autonomous view - all meanings of ambiguous words are accessed, then context compatibility decides which alternative is correct.
Interactive view - sentential context rules out inappropriate meanings before they are fully accessed.
Swinney, using cross-modal semantic priming found evidence for the first view. Part. heard homonyms such as ‘bugs’ embedded in a sentence. in the unbiased sentence, the word was ambiguous, it could mean instects, but also spying equipment. In the biased sentence, the word could only mean instects. Part. were then asked to make a lexical decision based on one of the meanings of the prime, or an unrelated word. In both types of sentences, both related targets were primed, meaning in the biased sentence, even though interpretation of the homonym was unambiguous, the other meanings were also accessed - support of autonomous activation model.However when it was repeated with targets appearing about a second later, only the contextually appropriate meaning was activated.
PARSING
The constructive process in reading or hearing, of taking the individual components (words) and combining them to produce something that bears relationship to what the speaker or writer intended; building up a mental model of the information
SYNTAX
Mutually agreed conventions for word order.
GARDEN PATH MODEL OF PARSING
Assumes that parsing is incremental, i.e. each new word is given a semantic role as soon as it’s encountered, refining the set of plausible new syntactic structures every time a new word is recognized. The model predicts that sometimes the parse chosen at a point of ambiguity will be incorrect (the listener will be led ‘down the garden path’) and re-analysis will be necessary.
CONSTRAINT BASED MODEL OF PARSING
MacDonald, assumes parsing is parallel and interactive. More than one syntactic evaluation of a sentence is possible at the same time (unlike in garden path model) - a bit like a cohort model of word recognition. It’s interactive because frequency and semantic plausability can influence parsing immediately. Involvement of the lexicon - some info about how a word goes with other words is stored in the lexicon.
SEMANTIC CONTEXT (constraint on parsing)
Trueswell et al ran an eye-tracking experiment. Using sentence fragments that could lead to a garden path interpretation, such as words that could be interpreted as verb, or as reduced relative clause (‘the car towed…’), they looked for evidence of processing difficulty (such as longer eye-fixations just after the point that introduces the ambiguity). In more constraining sematic contexts they found these contexts to reduce or even eliminate the garden path effect. Therefore the initial evaluation of the role of a word in a sentence is not only based on syntactic factor (not autonomous as garden path model assumes).
