Language & Cognition - Lecture notes 1-19 PDF

Preview

Summary

Lecture 1: What is special about language? & Word meaningsDefining language: ● A system that consists of the development, acquisition, maintenance, and use of complex systems of communication ● A shift in focus - identify characteristics that language must have, then compare these characteri...

Description

Lecture 1: What is special about language? & Word meanings Defining language: ● A system that consists of the development, acquisition, maintenance, and use of complex systems of communication ● A shift in focus - identify characteristics that language must have, then compare these characteristics to other communication systems (Hockett 1960) Hockett design features of language (1960): ● Vocal-auditory channel - focus on speech ● Broadcast transmission & directional reception - speech is not directional, but comprehender can identify the direction it’s coming from ● Rapid fading - must perceive it in real-time (until recording devices, etc.) ● Most animal communication has these features too ● Interchangeability - anything I can hear, I can say, you can say, we can all say (mating calls) ● Total feedback - we can hear & control what we are saying ● Specialisation - the purpose is communication (compare e.g. to if you are sweating, you know I am hot but this is not communicative) ● A lot of animals have these communication features too ● Semanticity - sounds are associated with meanings, the more specific the better ● Arbitrariness - language forms are arbitrarily linked to their referents ● Discreteness - language can be dissected into smaller units that are categorical rather than continuous ● Displacement - can talk about things that are not present; that do not exist; abstractions; ideas ● Productivity - can produce & understand entirely new utterances ● Traditional transmission - we learn language through interaction with others ● Duality of patterning - limited set of smaller units that are not meaningful & can be recombined in multiple ways ● Prevarication - ability to lie Is language a special cognitive process? ● Few, if any features are specific to human language only, but their combination is unique ● Animal communication only partly resembles language ● Certain characteristics are specific to human language and suggest strong separability of language from other aspects of cognition ● Phonemes:● Basic unit of sound in a particular language, limited inventory of phonemes ● Not meaningful in themselves but combine into meaningful units (morphemes) ● Contrastive, changing a phoneme changes meaning ● Morphemes:● The smallest meaningful elements of words ● Bound morphemes can’t stand alone e.g. ‘ed’ ● Inflectional don’t change the underlying meaning (e.g. dog>dogs), deviational change underlying meaning (e.g. doglike)

Words: ● Smallest unit of grammar that can stand on its own as a complete utterance, separated by spaces in the written language (David Crystal 2010) Phrases & sentences: ● Words have grammatical categories governing the roles they may play in sentences ● Rules of syntax determine how they may be combined in a language ● These differ across languages Processes & problems specific to language: ● Is there any other cognitive task similar to reading words? ● Can language be broken down? ● How can we test the central assumption that language & its components are represented & processed in a modular fashion? Meanings of words: ● Semantics refers to meaning, but in various contexts:● Knowledge - “gist” vs episodic memory ● Truth & language - logic & formal semantics ● Communication & shared meaning - sociolinguistics ● Lexical meaning (vs syntactic/grammatical structure) ● Referential theory of meaning is the idea that words are direct mapping to the real world ● The theory is handy as a starting point but it is not this transparent as concepts & word meanings are not always the same ● Approaches to semantics:● Featural - concepts are made up of smaller components, content is reflected in bundles of features of different types ● Holistic - concepts are unitary, represented in networks where relationships provide the content ● Classical view - word meanings are represented as definitions, features of meaning are necessary & sufficient in order to define a word ● The classical view has difficulty identifying necessary & sufficient features for all words e.g. game is defined as something that has rules but it doesn’t always have rules! ● Classical view also has issues with direct classification e.g. to what extent a sink is a piece of furniture Typicality effect The finding that people are quicker to make category judgments about typical members of a category than they are to make such judgments about atypical members For example, they are more quickly able to judge that a dog is a mammal than they are able to judge that a whale is a mammal

Alternatives to the classical view:

● ● ● ● ●

● ● ● ● ● ● ● ● ● ● ● ● ● ● ●

Neo-classical view:Also based upon features, except it’s two types of features Core features - defining as classical view, define a concept, reflect deep thinking of words Characteristic features - extra things that are relevant, response for typicality effects e.g. not all chairs have 4 legs but having 4 legs is characteristic of chairs But this approach does not solve the issue of core features still being quite hard to identify (e.g. a game), participants disagree on if certain features are necessary (Hampton 1995) Core features are not identifiable for most concepts (Smith & Medin 1981) Prototypes & features:Best example or summary representation, feature types, etc. Attempts to solve issues of classical theory Hard to define what kind of features abstract concepts have Holistic accounts, network models:Concepts are holistic rather than defined by primitives Emphasis is upon the relationship between concepts Concepts are represented as nodes Holistic accounts can explain verification speed ‘A robin is a bird’ is verified faster than ‘a robin is an animal’ (Collins & Loftus 1969; 1975) If verifying meanings = checking whether concepts are linked, following more links = slower responses Network models can scale up to real vocabulary Cucumber, celery & pumpkin are represented at a similar level which cannot be explained or accounted for by the model Full semantic characterisation of word meanings is still very much lacking, research now addresses real world problems or specific research questions

Summary: ● Human language differs from communication in other species, especially in combination of semanticity, displacement, productivity & the duality of patterning ● Language differs from other aspects of cognition & human communication in many of the same ways, particularly the use of specialised representation not apparently present in other modes ● Word meaning may be conceived as a combination of meaningful features, or as meaningful links between word representations ● Large scale theory-building & testing are still underway ● Many current approaches to meaning do not adopt a strong theoretical perspective or attempt to fully capture words’ meanings, but instead, use indicative measures that suffice for the questions being addressed

Lecture 2:

Visual world recognition A simplistic model of language comprehension: ● Input visual/auditory signal (this is what we want to process) ● Recognise each letter/phoneme ● Recognise each word ● Retrieve word’s meaning (semantics) ● Determine relationship to other words/concepts (syntax) ● But language is ambiguous at all levels of representation Visual word recognition: ● More straightforward mapping from the input than for speech Words & the mental lexicon: ● Models of word recognition assume that a representation for each word is housed in the mental lexicon ● I.e. we have a store of knowledge of all the words we have learned in our lifetime ● Aim of word recognition - to access correct entry in the lexicon (e.g. which word is it?) ● Visual word recognition is fast, within 500-600ms we can start to read a word out loud & decide if a string of letters is real or not (e.g. shark vs. sharm) ● What does it mean to say that we have recognised a word?:● Identified its constituent letters ● Processed information about their order ● Uniquely identified which word in your lexicon is present How do we move our eyes during reading? ● Detailed visual information is provided by the fovea ● Readers move their eyes in sequence of saccades to fixate on different words ● It is not a smooth motion, we do a lot of staccato jumps across the page (this is not exclusive to reading, it is all the time) ● The visual system gives our brain detailed sequences of snapshots ● Fixations usually last 200-250 ms & saccades are 8 letter spaces ● We do not have a preprogrammed movement of our eyes, it varies based on what we are reading ● 10-15% of saccades are backwards movements (regressions) ● Even though 4-5 letters lie within the fovea, preview information about shape of upcoming word & its initial letters are available from outside the fovea Does our knowledge about words help letter recognition? ● We don’t rely on visual information alone ● Word superiority effect - letters are easier to identify in a real written word (Reicher, 1969) ● E.G. what is the fourth letter when sli is presented, and the 4th letter is ambiguous, could be P or F, people are more likely to say it is p, because slip is a word and slif is not What factors affect word recognition?

● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●

Not all words are equal, some are easier than others It is hard to figure out what the factors are that makes some easier to recognise Many variables are correlated e.g. why is dog easier than chihuahua? Is it because dog is shorter? More common? Easier to spell? More recently encountered? Lexical properties:Frequency (e.g. house vs shark), Length (e.g. rat vs hippopotamus) Age of acquisition (controversial), semantic ambiguity Neighbourhood effect (e.g. sink has many neighbours, where you change one letter and it’s a different word but yacht has no neighbours) Concreteness (e.g. book vs hope) The ‘MegaStudy’ Approach (Balota et al., 2004):2428 one-syllable words Measures of frequency, length, sound-to-sound spelling consistency, imageability Lexical decision experiment Multiple regression used to separate out the effects of different variables I.e. is there an effect of word frequency when you factor out all the other effects? All the above variables affected speed & sometimes accuracy Recent experience:Recognition speed is influenced by what precedes the word Priming experiments look at this, faster reading for target word preceded by helpful context e.g. the dog barked vs tree bark Interference, slowing reading of words when preceded by unhelpful context

Is word recognition serial or parallel? ● How do we match the word form that we see against the correct entry in our mental lexicon? ● I.e. mapping one word that you do not know against one target word that you do know ● A serial process - check against every word that we know, one at a time? (see Forster 1976) ● Would be like looking up a word in a dictionary, checking stores in your brain one at a time like a digital computer ● A parallel process - all possible words checked simultaneously? ● E.g. interactive activation model (McClelland & Rumelhart 1981) moves away from the ‘brain as a digital computer’ idea, instead says the brain is an interconnected network of neurons with massively parallel processing ● There is no strong evidence either way but the serial search is implausible from a neural perspective ● The brain performs individual computations relatively slowly compared with a computer ● However, its interconnected structure allows for massively parallel processing ● The general consensus is that word recognition is parallel in nature

How do we code information about the order of letters within a word?

● ● ● ● ●

Words can be mucked up a lot and still understood The interactive activation model does ‘slot based coding’, so a unit will only turn on if it is the right letter in the wrong place I.e. each letter is coded according to its absolute position within the word Our ability to understand mucked up words suggests absolute position idea is not completely true The issue of how letter positions are encoded during recognition is still uncertain

Does the sound of a word play a role in its recognition? ● Many English words have regular spellings, pronunciation can be worked out on the basis of letter sound knowledge ● Children are taught to make use of this regularity (phonics, sounding words out to read them) ● Do fluent adult readers do this too? ● Semantic relatedness judgments (Lesch 1998), slower responses for pairs like ‘sand’ - ‘beech’ ● Homophone experiment^ suggests sound information is important but there is still controversy ● Most agree that the primary route does not involve phonological (sound-based) representation ● DRC model of reading (Coltheart et al 2001) includes letter to sound representations and the 2 routes we can take to get from print to speech Summary: ● Snapshots of detailed information are provided by rapid saccades ● There is careful control of eye movements according to the properties of words ● We are better at identifying letters within a familiar word ● There are many factors that affect how easily we recognise a word ● Word recognition is more likely to be a parallel process ● We code information about the order of letters within a word via coding absolute and relative position ● The sound of a word does play a role in its recognition, especially for lowfrequency words and for poor readers

Lecture 3:

Spoken word recognition, word meaning access & processing sentence structure Speaking: ● Sound is produced by flow of hair from lungs, vibration of vocal cords (in larynx) and constrictions in vocal tract ● Sound is filtered by the cavities of the vocal tract, articulators for different shapes for each sounds ● For vowels, the vocal tract is in open configuration, different vowels are produced by varying the shape of the vocal cavities ● Consonants are produced by varying manner of articulation, air stream from lungs is constricted fully or partially ● Phonemes - spoken word equivalent of a letter ● Mapping between letters & phonemes is not one-to-one i.e. k sound is in both Cat & Kit ● Phonemes are the sound units of language & are defined relative to a particular language (which is why learning a new language as an adult is hard) The nature of speech: ● Spoken word recognition is different to visual word recognition:● Speech is more variable & has more temporal properties ● Variability:● Place of articulation (e.g. where your tongue is) & voicing vary continuously (timing of which vocal cords vibrate) ● There are no sharp boundaries between the acoustic properties of similar phonemes (e.g. /d/ vs /g/) ● Large differences between speakers e.g. accent, gender, age ● Also large differences within a speaker (every sound I produce is not the same, e.g. speech rate, emotion, volume) ● Phonological context effects - speech sounds are influenced by adjacent sounds ● Coarticulation - smearing between sounds e.g. hambag sounds like handbag ● Solutions to variability - Categorical perception:● Our speech perception system imposes categories ● Chunking world into categories, and discriminating between things that straddle the boundaries ● Different languages have different boundaries, some categories are innate & some are learned ● This sensitivity is lost after 10 months and is not uniquely human ● Solutions to variability - Lexical knowledge:● We use things spoken in our language e.g. Ganong effect (Ganong, 1980) ● Solutions to variability - Knowledge about specific speakers:● (Nygaard & Pisoni 1998) ● Solutions to variability - Use of visual information:● Listeners use information about the speakers lip position (The McGurk effect)

The timing of speech is different to text:

● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●

With speech, input develops over time, but with reading whole word is simultaneously available There are no gaps between words, sometimes we pause but it is not the same Segmentation problem - divide the speech stream into word chunks e.g. i scream vs ice cream I.e. you hear the start of the word before the end When do you recognise a word? Do you wait until you have the whole word? Evidence suggests we do early recognition & we do not wait for the end of the world E.g. capt is ambiguous between captain & captive, are we retrieving both of these in parallel? Are we jumping the gun? Hear ambiguous fragment -> lexical decision to visual target Capt primes ship & guard (related to captain vs captive) Captain only primes ship Suggests we initially activate both in parallel, then rapidly select one on the basis of subsequent input (Zwitserlood 1989) Cohort model:Marslen-Wilson et al, multiple versions published, model always changing As a word is heard, all the words consistent with the initial sequence enter the cohort (or become activated) Words stay in the cohort if they match, but drop out of the cohort if there is mismatch (or their activation is reduced) A word is recognised when it is the only word left in the cohort (or its activation is greater than any other word) Segmentation cues:Acoustic cues e.g. vowel lengthening, ice cream vs i scream Segmentation follows recognition, in cohort model words are recognised before their offset, telling the listener where the next onset will be This works well for long words but not short words, as many words are not unique until they have finished Up to 60% words do not become unique before offset e.g. cat... (Bard et al., 1988) Metrical segmentation strategy (MSS) by Cutler 1990 - most words in english start with strong syllable, so if we assume every stressed syllable is a word onset it acts as a good cue (not 100% accurate though)

Word meaning access: ● How do we know what words mean? ● Cognitive mechanisms:● Automatic retrieval of multiple meanings in parallel ● Rapid selection of single meaning ● Occasional need for subsequent reinterpretation ● Cues that help with word-meaning selection:● Rapid, fluent access requires integration of many different statistical (probabilistic) cues ● Sentence context e.g. the bark of the tree/dog ● Long-term experience with the word (dominance) ● Recent experience with the word ● Knowledge about the speaker/writer Sentence context & dominance:

● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●

Ambiguous words with balanced vs biased meanings (Duffy et al 1988) For biased words, context was consistent with lower frequency meaning Variation where the context is put in the sentence (early vs late disambiguation) Method - eye tracking There was a complex interaction between dominance & context Early disambiguation - people are slower with identifying big frequency difference (e.g. port) Late disambiguation - getting port early doesn’t cause much disruption in eye tracking but is then slowed when context is given (due to need to reinterpret) Shows that context matters as well as relative frequency, and the two interact When the cues are consistent, reading is easy, multiple cues cause an issue Meaning preferences differ across individuals & change over time Rodd et al 2016 studied this in rowers, accessibility of rowing meanings depended on recent (same day) experience & long-term (years) experience Recent experience:Prime phase - semantic relatedness task (e.g. seal & bank) Filler task - digit span Test phase - word association task Does prime influence response in the test phase (compared with a control who did not see the sentence)? Significant increase in responses consistent with context meaning Likely meanings are more accessible!

Understanding sentence structure: ● Understanding word meanings alone is not enough, we need to know how the word meanings relate to each other ● Syntax - the rules that govern how words can be combined in a particular language, separate to semantics (meaning of word) ● Word lists have semantic content but no syntactic structure (lecture difficult students baffled) ● Syntax tells us how the individual words relate to each other ● Syntactic rules vary over languages, they have the same kinds of principles but different weightings ● Syntactic trees are a hierarchical way of demonstrating the structure of a sentence, they represent roles of words ● Syntactic pars...