Models of word production and reaction-time evidence

Last week 3 stages to production: Conceptualisation Formulation Articulation Formulation involves lexical retrieval: Semantic/syntactic content (lemma) Phonological content (word-form) Tip of tongue state when lemma is retrieved without word-form being retrieved

This week Levelt et al.’s theory of word production Modularity in language production Experimental evidence for this account Dell’s interactive account Recent experimental evidence that supports interaction Can the modular approach explain these results?

Levelt, Roelofs and Meyer’s (1999) model of word production See two figures: theory outline and description of network Theory outline: multiple levels of representation lack of feedback except via the monitor Network has three strata conceptual stratum lemma stratum word-form stratum

Conceptual stratum Conceptual stratum is not decomposed thus, one lexical concept node for “escort” not decomposed into “walk”, “be with”, “safeguard” instead, conceptual links from “escort” to “safeguard”, etc.

Activation of concepts “pragmatically”, via the intention to communicate something (e.g., describing an object) together with “perspective” (e.g., using “dog” vs. “animal”) but also via spreading activation from other concepts or via direct activation of that concept (e.g., the word “dog”) or perhaps via some random, spontaneous activation

Lexical selection First, lemma activation occurs This involves activating a lemma or lemmas corresponding to the concept thus, concept DOG activates lemma “dog” but also involves activating other lemmas DOG also activates CAT (etc.) to some extent and CAT activates lemma “cat”

Lemma selection Distinguished from activation Only one lemma is selected (in this model) probability of selecting the target lemma (“dog”) is the ratio of that lemma’s activation to the total activation of all lemmas (“dog”, “cat”, etc.) hence competition between semantically related lemmas competitors can be activated in other ways (see below)

Determining grammatical properties Fixing “diacritic parameters” correspond to grammatical properties of the word include grammatical category (noun, verb, etc.) include number, person, tense, mood for verbs include number, person, gender, count/mass status for nouns DOGS leads to selection of “dog” lemma plus plural number and noun category parameters Note: function words can be selected on purely syntactic grounds e.g., that in John saw that is activated by saw

Morpho-phonological encoding (and beyond) (a very rough sketch) The lemma is now converted into a phonological representation called “word-form” (or “lexeme”) If “dog” lemma plus plural (and noun) are activated Leads to activation of morphemes dog and s Morpheme = smallest meaningful unit of the language

Other late stages (in speaking) Accessing metrical shape basically, syllable structure and stress pattern escort has two syllables and is stress-final Accessing segmental make-up basically the phonemes that make up the word e.g., /d/, /o/, /g/ Then a phonetic representation is constructed This specifies the articulatory task that will produce the word Finally, the word is articulated

A modular account claim that processes are encapsulated, so that the output to a module depends only on its inputs cf. The Modularity of Mind (Fodor, 1983) but Fodor didn’t really consider language production in production, modularity amounts to the claim that later processes cannot affect earlier processes

Levelt et al.’s model allows no feedback during encoding critically, no feedback from word-form stratum to lemma stratum contrasts with Dell’s account (see below)

Also, only one lemma activates a word form if “dog” and “cat” lemmas are activated, they compete to produce a winner at the lemma stratum Only the “winner” activates a word form the word forms for the “losers” aren’t accessed hence, it doesn’t allow “cascading” activation

A rift between lemma and word form Levelt et al.’s model assumes a clear divide only one word (lemma) can “cross the divide” between lemma and wordform strata and no feedback possible Accords well with TOT and anomia data sometimes the divide cannot be crossed

Self-monitoring listening to yourself Correcting yourself Critical role in Levelt et al.’s model “external loop” monitoring what is said internal loop” monitoring the phonological representation (probably) used to explain results that appear to be due to feedback or cascading activation

Picture-word interference Participants name basic objects as quickly as possible e.g., picture of a dog Distractor words are embedded in the object e.g., the word cat participants are instructed to ignore these words

A stroop-like effect Basic finding that semantically related words can interfere with naming e.g., the word cat in a picture of a dog However, form-related words can speed up processing e.g., the word dot in a picture of a dog

Experiments manipulate timing: picture and word can be presented simultaneously or one can slightly precede the other We draw inferences about time-course of processing and test word-production models

Schriefers, Meyer, and Levelt (1990) Auditory presentation of distractors hence, definitely phonological (not orthographic) effect Conditions unrelated word SHIP phonologically relatedDOT alliterative (i.e., same beginning) semantically related CAT TARGET: DOG

SOA (Stimulus onset asynchrony) manipulation -150 ms (word …150 ms … picture) 0 ms (i.e., synchronous presentation) +150 ms (picture …150ms …word)

Results Semantic effect: -150 ms Word … 150 ms … Picture  inhibition (related slower than control) 0 ms, +150 ms Word = Picture; Picture …150 ms … Word  no effect

Results Phonological effect: ms: Word … 150 ms … Picture  no effect 0 ms, +150 ms Word = Picture; Picture …150 ms … Word  facilitation (related faster than control)

Separate semantic and phonological stages? Early semantic inhibition Late phonological facilitation Fits with the assumption that semantic processing precedes phonological processing No overlap suggests two discrete stages in production an interactive account might find semantic and phonological effects at the same time

Other studies support and extend these results other picture-word experiments (e.g., Meyer & Schriefers, 1991, JEP:LMC) other methods (e.g., Levelt et al., 1991; Van Turrenout et al., 1998) but many other experiments criticise these claims (see below)

Dell’s interactive account Dell (1986) presented the best-known interactive account other similar accounts exist 3 levels of representation semantics (decomposed into features) words phonemes (sounds)

Interactive because information flows “upwards” as well as “downwards” e.g., the semantic features mammal, barks, four- legs activate the word “dog” this activates the sounds /d/, /o/, /g/ these send activation back to the word level, activating words containing these sounds (e.g., “log”, “dot”) to some extent this activation is upwards (phonology to syntax) and wouldn’t occur in Levelt’s account

Evidence: “Mixed” errors Both semantic and phonological relationship to target word Target = “cat” semantic error = “dog” phonological error = “hat” mixed error = “rat” Occur more often than predicted by modular models if you can go wrong at either stage, it would only be by chance that an error would be mixed

Dell’s explanation the semantic features of dog activate lemma “cat” some features (e.g., animate, mammalian) activate “rat” as well “cat” then activates the sounds /k/, /ae/, /t/ /ae/ and /t/ activate “rat” by feedback this confluence of activation leads to increased tendency for “rat” to be uttered Also explains the tendency for phonological errors to be real words Sounds can only feed back to words (non-words not represented) so only words can feedback to sound level

Why might interaction occur? Can’t exist just to produce errors! Perhaps because the same network is used in comprehension So feedback would be the normal comprehension route Dell argues against this because many aphasics have good auditory word recognition yet disturbed phonological encoding

Alternatively, it simply serves to increase fluency in lemma selection advantageous to select a lemma whose phonological form is easy to find

Evidence for interactivity A number of recent experimental findings appear to support interaction under some circumstances (or at least cascading models) Peterson & Savoy (JEP:LMC, 1998) Cutting & Ferreira (JEP:LMC, 1999) Griffin & Bock (JML, 1998) Damian & Martin (JEP:LMC, 1999)

Peterson & Savoy found evidence for phonological activation of near synonyms: Participants slower to say distractor soda than unrelated distractor when naming couch Soda is related to non-selected sofa remember that Levelt et al. assume that only one lemma can be selected and hence activate a phonological form Levelt et al’s explanation: Could be erroneous selection of two lemmas?

Damian and Martin (1999) Extension of Schriefers et al.’s picture- word interference task remember that semantic inhibition occurred early, phonological facilitation occurred late (with no overlap) various methodological changes and developments focus on Experiment 3

The critical difference from Schriefers et al. is the addition of a “semantic and phonological” condition Picture of Apple peach (semantically related) apathy (phonologically related) apricot (sem & phono related) couch (unrelated) (also no-word control, always fast)

Results Relatedness-150 ms0 ms+ 150 ms Unrelated Semantic Phono S & P

Summary of findings early semantic inhibition (- 150 and 0 ms) late phonological facilitation (0 and ms) shows overlap, unlike Schriefers et al. but S & P condition didn’t show early semantic inhibition

This last finding demonstrates that semantic interference is reduced in the simultaneous presence of a phonological relationship (which should facilitate) Thus the finding appears to contradict the “discrete two-step” account of Levelt et al.

Can the two-stage account be saved? Evidence for interaction is hard to reconcile with the Levelt account however, most attempts are likely to revolve around the monitor basically, people sometimes notice a problem and screen it out Levelt argues that evidence for interaction really involves “special cases”, not directly related to normal processing

Summary Levelt et al.’s theory of word production: Strictly modular lexical access Syntactic processing precedes phonological processing Dell’s interactive account: Interaction between syntactic and phonological processing Experimental evidence is equivocal, but increasing evidence that more than one lemma may activate associated wordform

Caramazza’s alternative Caramazza and colleagues argue against the existence of the lemma node instead they propose a direct link between semantic level and lexeme syntactic information is associated with the lexeme Also assumes separate lexemes for written and spoken production This is really a different issue

Much evidence comes from patient data But also evidence from the independence of syntactic and phonological information in TOT states see discussion of Vigliocco et al. also Caramazza and Miozzo (Cognition, 1997; see also replies by Roelofs et al.)