Lexical semantics - Göteborgs universitet Semantics-eng.pdf · Meaning potential and semantic operations Words have meaning potentials and have their specific meaning determined

Lexical semantics

What is meaning and how is it linked toword forms and to larger linguisticunits?

• Word semantics• – Prototype theory• – Semantic distinctive features• – Semantic fields• – Semantic features and operations in aphasia• Sources of data and theories• – Deep dyslexia• – Category-specific anomia• – Right-hemisphere lesions and lexical-semantic

problems• – Semantic dementia

– Methodological considerations– – Group studies versus case studies

• – Central or access disorder• – Mapping of form and meaning

– Types of process models– – Word comprehension

• – Word production• Simulation• Links to therapy

Nominalismthere are objects and phenomena in theworld and words to denote them; nothingelse (i.e., no intermediate representation) isneeded.Conceptualism (most often found inneurolinguistic studies):there are concepts in our minds that linkwords to objects and phenomena in theworld.Conceptual realismconcepts exist in themselves,independently of human minds (e.g., so-called Platonic ideas).

Basic units of meaning

• What is or can be the basic units ofmeaning?

• words, parts of words, or evensmaller features

• sentences, utterances, contributions,longer texts, monologues, ordialogues

• Or could they all be?• And what is the relationship between

the meanings of linguistic units andtheir context of use?

Anomia

Anomia (difficulties in finding“content words”) is a symptomshared by almost all persons withaphasia.(In the type of aphasia called“anomic aphasia,” this is the onlysymptom; while in the other types ofaphasia, it is one of severalsymptoms.)Anomia creates problems in taskssuch as naming objects (visuallydisplayed, from verbal descriptions,or from memory).

Hypotheses about the”mental lexicon”

Aphasic ways of searching for words andaphasic semantic word substitutions(semantic paraphasias) have been used totest different hypotheses about how a“mental lexicon” can be organized.Central or access disorder?A basic question is whether the “lexicon”can be disturbed in itself – that is, can itsorganization be affected? – or whetheronly the ways of accessing the lexicon canbe disturbed.Since we are dealing with patterns ofactivation in the brain, this is really apseudo-question, or rather a questionregarding the quality, degree, or variationin activation patterns

Example

Target word: LADDERA: yes there we have it there… he had

one of those yesterday the one whopicked apples I know so well whatit’s called we have one over there bythe shed no cannot

Target word: SHOEMAKERA: carpen … no it isn’t that shoebake

… shoebaker, no carpen carpenT: shoemake …A: shoebakerT: no shoemake …P: shoemaker

Prototype theory

Rosch (1975)• Categories are organized around the

most prototypical exemplars, whichare central and more “stable.”

• “prototypical” items, are namedmore easily and consistently thanmore “atypical” exemplars,

• exemplars that are borderlinebetween categories are associatedwith the “fuzzy” fringes of the fields.

Prototype theory andaphasia

• Prototype theory• Whitehouse, Caramazza, and Zurif (1978)

(in an experiment designed by Labov,1973).

• The task was to determine whether drawnobjects were “cups,” “plates,” or “glasses.”Stimuli gradually became more similar andit was noted where each person drew thelimits between the different types ofobjects.

• Broca’s aphasics used the same strategiesas controls for classifying objects

• Anomic aphasics either showed a totalinability to name objects systematically orused rules that were too simple. They werealso unable to use the added context in apicture as a cue.

Superordinate-Basic-Subordinate level

Prototype theory has been combinedwith considerations of differentlevels of abstraction in a hierarchicalorganization of semantic categoriesinto- superordinate (animal)- basic (dog)- subordinate (poodle)levels.It has been claimed that base-levelwords are more likely to beprototypical, and remain more“robust” for many persons withaphasia

Therapy effects

Studies by Kiran and colleagueshave shown that there is asemantic complexity effect, inthat training on atypicalexamples of animate andinanimate categories and theirsemantic features leads toimprovement in the naming ofmore typical exemplars as well,whereas the reverse does nothold

Semantic DistinctiveFeatures

• The idea is that a category can beidentified and described withreference to the necessary andsufficient conditions for belongingto that category.

• This process therefore involvesdecomposition into more primitiveor basic features.

• Examples of semantic distinctivefeatures are

• [+living], [+male], [– adult] for boy.

Semantic features andaphasia

• The words the participants had to classifywere mother, wife, cook, partner, knight,husband, shark, trout, dog, tiger, turtle,and crocodile.

• The controls tended to classify the wordsinto the human and animal categories,

• Broca’s aphasics divided them into human+ dog versus other animals.

• Controls then sorted the animals accordingto species.

• Broca’s aphasics sorted them into more orless dangerous groups.

Controls used semantic features in asystematic, “technical” way.

Broca’s aphasics used temporary andconspicuous features and were moredependent on emotional andsituational influences.

Wernicke’s aphasics usedindeterminate and deviant featuresand were generally “out of focus.”

Semantic fields ornetworks

• Semantic fields or networks aregroupings of words according tosemantic similarity, or contiguity(co-occurrence), relations.

• Hierarchical and similarity-basedword association network used byCollins and Quillian (1969), basedon “is a” relations between words;for example, a poodle is a dog, a dogis an animal.

• Especially for natural kinds, this is agood way of describingparadigmatic, similarity-basedrelations.

• Collins and Loftus (1975)• model of word meanings based

on semantic similarity.

• Words are activated byspreading activation in thenetwork when related words arearoused.

• This is the basis for semanticpriming of word recognition

• (i.e., that a word is recognizedfaster if it has been preceded bya semantically related word).

There are, however, alsosyntagmatic semantic fieldrelations between words, basedon contiguity or co-occurrence.Such relations exist betweenword pairs such as cat and dog,eat and food, or red and light.

Similarity - Contiguity

• Similarity and contiguity relationsinteract in determining associationstrength between words.

• For example, cat and dog bothbelong to the same category in ahierarchical, similarity-based field,

• but they also tend to co-occur inexpressions such as cats and dogs.

• Semantic fields can be based onrelations between words, which arein turn based on relations betweenobjects, events, and properties in theworld.

Semantic fields andaphasia

Wernicke’s aphasics in particularhave difficulties in recognizing andusing the relations.Luria (1976) claimed that wordshave graded semantic associativefields.Goodglass and Baker (1976)- parts of these fields could bedamaged selectively- object naming was affected by howmuch of the semantic field wasaccessible.

Task: to name objects and then judge whetherother words were associated with theobject namesThe speed of recognizing associated wordswas measured.All subjects had shorter response times forwords associated with objects that they hadbeen able to name.Certain associated words were consistentlyeasier, while others were harder torecognize.

Meaning potential andsemantic operations

Words have meaning potentials andhave their specific meaningdetermined by the current context(Allwood 1999, 2003).Context determination of meaning iscrucial and lexical semantics worksby semantic operations on meaningpotentials.Semantics and pragmatics are, thus,unified.This perspective is also importantwhen looking at communicativecontributions in context

Semantic features andoperations in AphasiaAn intended target word is very often replaced by aword that is semantically and/or phonologicallyrelated to it.

The semantic relation can be based on similarity orcontiguity (i.e., co-occurrence ).

Some semantic relations that are often found inthese situations aresame semantic categorysuperordinatesubordinatepart for wholeattributespatial relationfunctional-causal relation.

Many semantic substitutions andparaphrases (circumlocutions) describesituational features, indicating a need tocontextualize.

More substitutions seem to move in thedirection from abstract to concrete than inthe opposite direction, indicatingdifficulties with abstraction

(Allwood & Ahlsén, 1986, 1991).

• Same category: cat for dog• Superordinate: dog for poodle• Subordinate: poodle for dog• Part for whole: trunk for elephant• Attribute: yellow for banana• Spatial relation: head for cap• Functional causal relation: kick for

ball• Circumlocution: he had one of those

yesterday, the one who picked appleswe have one over by the shed forladder

Category-specificAnomia

anomia that affects a certaincategory of words, but leavesother categories unaffected.

Nouns vs verbs

Nouns and verbs can be selectivelydisturbed in relation to each other.Verbs - nonfluent, agrammatic aphasics withfrontal lobe lesions.Nouns - persons with anomia, who have temporal-parietal lesions.

The verb disorder can be considered in relation to apossible frontal encoding of verbs of movement andactionIt may also be related to the role of verbs ingrammatical structure.The noun disorder would be more related to theassociation of sensory features.

Concrete vs abstract

Concreteness effect:concrete nouns are more easily activated.Why?- concrete nouns have richer representationalstructures in memory, including, for example, anonverbal image coding?- more contextual information?- a larger set of semantic features?

But some (albeit few) persons with aphasia activateabstract nouns more easily.

Another argument is that concrete nouns arespecifically associated with sensorimotor (includingperceptual) attributesIf these attributes are disturbed, abstract nounswould be easier to access than concrete ones.

Natural/living objects vs artifactsSelective disturbances of- words for either natural objects (animal,fruits, vegetables) or- artifacts (tools, furniture, etc).Such patients may have very vaguesemantic information about one of thecategory types, but very detailed semanticinformation about the other type.Why?- mainly perceptual/sensory versus mainlyfunctional/motor basis (for natural objectsand artifacts, respectively)Support: naming of animals resulted inactivation in the left medial occipital lobe,while naming of tools activated the leftpremotor area.orwe really have more detailed category-specificity in the organization of ourmental lexicon.

• Semantic word substitutions seem to keep“within category” and resemble “speecherrors” made by normal speakers (e.g., bigbecomes small) (Buckingham, 1979). Others havestudied conceptual organization. Zurif, Caramazza,Foldi, and Gardner (1979) found that both Broca’sand Wernicke’s aphasics had difficulties instructuring lexical knowledge by conceptualfeatures. The subjects first had to sort wordsaccording to thematic relations and commonconceptual features. Then they were given a wordrecognition test in order to see whether theserelations were actually used. The words wererelated in the following ways:

NECK – HEAD – locative – CAP SAW – instrumental – WOOD – BRICK | | relation relation | |+body +body building +building part part material material

MethodologicalConsiderations

• Groups studies vs case studies• Central or access disorder

Mapping of Form and Meaning

• How words are processed - how form andmeaning are mapped onto each other inlanguage production and comprehension.in relation to process models of differenttypes and to the units studied,

• How much top-down versus bottom-upprocessing is assumed to take place?

• When we understand speech, do we startby mapping words or morphemes tomeaning units and construct the meaningof larger units from this mapping,or do we start with expectations andhypotheses about the meaning of a specificword or utterance based on backgroundknowledge, situation, and linguisticcontext?or do the two types of processing interactand, if so, how?

Models for wordprocessing

• Many of the earlier studies, andpresent-day studies as well, arebased on serial models, such as

• Garrett’s (1982) or• Levelt’s (1989) model of speech

production or• the PALPA model (Kay, Lesser &

Coltheart, 1992) of speechrecognition/comprehension,

• as well as the search and cohortmodels for visual and auditory wordrecognition presented below.

Interactive activationmodels

• Interactive activation models,simulated in artificial neuralnetworks (ANN) are, however,being used more and more,

• for example, the TRACE modelfor recognition/ comprehensionand

• the IAM model for productionused by Dell and coworkers

Word comprehensionmodels

• Forster’s active search model• Morton’s logogen model, which

posit a passively responsivelexicon where different units areactivated by incoming stimuli

• the cohort model proposed by• the TRACE model of

Active direct searchmodels

• In an active, direct search lexicon, itis assumed that one searches formatches of an incoming stimulusword in “bins” containing wordrepresentations (auditory or visual)which are ordered by someprinciple, for example, howfrequently they have beenencountered. The wordrepresentations in the “bins” are thenlinked to a “lexical master file,”where phonological and semanticcross-references exist. Similarly, aword can be accessed from asemantic representation inproduction.

Passive indirectresponsive models

• In a passive, indirect, responsivemodel, like the logogen model,auditory and visual incoming wordstimuli can activate a specific“logogen” (recognition unit) if thestimulus is strong enough (i.e.,contains enough activation and nottoo much inhibition in relation to thefeatures of the logogen to get overits activation threshold). Thelogogens are also connected to asemantic cognitive lexicon, whichcan increase or decrease theactivation and thus affect whichlogogen is activated.

Cohort model

A cohort model stresses theincremental build-up ofactivation as, for example, awritten word is encounteredletter by letter and a particularword is selected at a particulardecision point. An earlyperceptual analysis determinesthat a set of words are possiblecandidates and recognitionproceeds by eliminating thecandidates from left to right

The TRACE model

• The TRACE model is a highlyinteractive model for spokenword recognition, emphasizingtop-down processing, ordependence on the context, inword recognition. It claims thatlexical knowledge helpsacoustic perceptual processes.

Lexical DecisionExperiment

4040Infrequentbisyllabicnouns

4040Frequentbisyllabicnouns

Nonwords(made bychangingletters inrealwords)

Words

Semantic priming• A specific instance of the lexical decision

task is semantic priming, where therelationship between two stimuli isinvestigated.

• The first stimulus is presented, followed bythe second, which is the target word todecide on (as a real word or a nonword).Semantic priming tasks can establishpatterns of association.

• Semantically related word as prime wordfacilitates the recognition (lexical decision)of the target word, when compared to aunrelated prime word.

• For example, the word orange isrecognized faster as a real word if it hasbeen preceded by the prime word apple,than if it has been preceded by, forexample, the word chair.

Effects in word recognition that modelshave to explain

- The frequency effect: The more frequent aword is, the faster it is recognized.- The length effect: The shorter a word is, thefaster it is recognized.- The concreteness effect: The more concrete aword is, the faster it is recognized.- The word superiority effect: A letter isrecognized faster in a word than if it occurs inisolation.- The word/nonword effect: Words are identifiedas words faster than nonwords are identified asnonwords.- The context effect: words are identified fasterin context.- The degradation or stimulus quality effect: A wordthat is presented clearly is recognized faster than aword that is blurred in some way (e.g., covered by agrid), if it is presented visually.

Word production models

Garrett’s (1982a) productionmodel andLevelt’s (1989) model forspeaking are examples ofinfluential serial productionmodelsInteractive activation modelshave been presented by Dell andcoworkers, Harley, and others

Levelt’s model forspeech production

Counter evidence toserial production

• Blends, where two word forms arecombined in the “word” that is finallyproduced - lexical semantics andphonology can mutually influence eachother.

• E.g. plower, combining plant and flower.

• Substitutions where the target and thesubstitute have both a semantic and aphonological relation (e.g., cat for rat,thumb for tongue) are much more frequentthan one would expect if the two levels didnot interrelate in some way

• But the interaction is limited, so thatprocessing is only semantic at first, whilein the last stages phonology seems todominate.

Simulation of aphasicword substitutions

Simulations of aphasic word substitutionshave been done, using artificial neuralnetworks- modeling the interaction betweenphonology and semantics.Interactive activation models arenonmodular (i.e., each process is not self-contained).Activation spreads bidirectionally in acontinuous flow. In this way, earlier andlater “stages of processing” influence eachother.A compromise model which is interactivebut contains a lexical and a phonologicalstage was developed by Dell et al. (1997)and can be used as an example.

Dell’s ANN modelIt is a localist model, that is, nodescorrespond to psychologicalproperties (one-to-one).Three levels of nodes are included:semantic featuresknown wordsphonemesAll connections between nodes areexcitatory and bidirectional. Allconnections have the same (preset)weight and there is a preset decayfactor (which determines how fastthe activation of nodes decays) andnormally distributed noise. Externalupdating is used initially byactivating the semantic features ofthe target word.

Simulation - semantic vsphonological errors

This model allows simulation ofdissociations between patients with mostlysemantic and mostly phonological errors.Lesions affecting the decay factorproduced semantic, mixed, and formalerrors. If they were not severe, there was asemantic component in most errors.Lesions affecting connection weight, onthe other hand, produced mostlyphonological errors. Combinations of the two types of lesionsresulted in intermediate, mixed patterns.The model could also simulate longitudinaldata for the patients, with the same factorssimulating the later stage, only closer tonormal values.

Therapy

Lexical-semantic research at thesingle-word level has inspired anumber of therapy methods, whichare designed to train wordassociation patterns, for example,- Luria’s restoration therapy(Luria, 1963),- Deblocking therapy (Weigl &Bierwisch, 1970),- BOX therapy (Visch-Brink,Bajema, & Van de Sandt-Koenderman, 1997)- Semantic Association therapy(e.g., Martin & Laine, 2000).

Assignments

• Look at the following list ofsemantic word substitutions:

1. football -> balloon2. hose -> house3. poodle -> lion4. zebra -> trunk5. donkey -> horse6. doll -> cap7. boy -> man8. spade -> garden thing9. orange -> apple10. green -> purple

• Try to explain each substitution:(a) in terms of semantic fields, prototype

theory, or semantic distinctive features;(b) in terms of one or more diagnoses,

indicating in which patients such asubstitution might occur;

(c) in terms of what type of model – a serial(Levelt-type) model or an interactiveactivation (Dell-type) model – you thinkwould best explain the substitution.