Language Processing - Morph Processing

Language Processing - Morph Processing

1. WORDS AND RULES DEBATE (PINKER): We process regulars and irregulars differently

2. Mopheme = smallest linguistic unit of meaning

   • walk-ed
   • rat-s
   • use-able, re-use-able
   • govern-ment
   • Either a word that can stand on its own or an affix which cannot
     • Suffix
     • prefix
     • STEM
   • INFLECTIONAL MORPHOLOGY = different word forms but do not change stem (walks, walk, walking)
     • Don't change syntactic category of the word
     • Regular / irregular - (go/went, mouse/mice, walk/walks)
     • Strict line between storage and computing may not be accurate
       • Perhaps we can both store and compute regular inflections
         • Studies on FREQUENCY EFFECTS: Frequent words are accessed faster - so F.E. of regulars and irregulars can be diagnostic of storage
         • Frequency FX well attested for irregulars (e.g. Prado & Ullman, 2009)
           • "went" vs "bled"
           • IF REGULAR PAST TENSE IS COMPUTED ONLINE FREQ SHOULDN'T AFFECT SPEED OF ACCESS
             • (Has to be tested with words of equal frequency so as not to confound exp.)
             • ALEGRE & GORDON (1999): Lexical decision task w Eng. regular infl
               • Reaction times for series of regular infl. forms ranging in freq
               • FOUND: For low freq words, there were no freq FX (0-6 / million) but across a broader freq range there WERE FREQ FX 0 – 44 / million
                 • This suggests that very high frequency words can be stored, not just computed
                   • WHICH IS INCOMPATIBLE with an extreme DUAL MECH MODEL!!!
                     • Ullman’s (2001) declarative/procedural is one dual-mechanism model that takes this into account
                       • Regular inflections (e.g. walked) can be computed online via procedural memory or stored in declarative memory
                       • Irregulars (e.g. went) are stored in declarative memory
     • Infl. morph inv in lang imp
       • SLI: Genetically based impairment to normal acq. of lang.
         • Diffs interpreting certain types of sentecnes, e.g. passives
         • Diffs with inflectional suffixes: "The boy walk home"
         • Ullman says children w SLI have deficit w PROCEDURAL mem, so instead rely on DECLARATIVE
         • ULLMAN & GOPNIK (1999): Investigated Inflectional morph in members of KE family
           • FOUND: Children in this family did no produce over-regularisaion errors
             • Diffs producing novel regular inflections, e.g. Wugs
               • However they were MORE ACCURATE w HIGH FREQ than LOW one
             • Maybe over relying on declarative mem.
           • (A british family of Pakistani origin half of whom have language diffs)
       • VAN DER LELY & CHRISTIAN (2000): Examined compound information in SLI
         • 10-18 years old
           • Lang matched controls & age matched controls
         • Used a version of Gordon’s (1985) exp.
         • FOUND: children with SLI included similar numbers of regular plurals inside compounds as irregular plurals
           • The control groups almost never included regular plurals inside compounds
             • Support the notion that individuals with SLI have deficit in GRAMMAR COMPUTATION
               • Individuals with SLI store both regular and irregular plurals in the mental lexicon - which is why they are appearing inside compounds
   • DERIVATIONAL MORPHOLOGY = create new leximes/words (happy/happiness)
     • Suffixes (but NOT prefixes) usually change word category (govern - government)
     • Semantically transparent / opaque (department)

3. Semantic priming suggests words related in meaning are connected in the mental lexicon

   • MORPH also provides means of studying mental lex.
     • BERKO 1958
       • Used non word to test children's production of regular plural
         • Found they could apply the rule 'add s' even to words they'd never heard before
           • This, again, refutes the old school imitation theory
           • Dual mechanism: "This is productive application of a regular rule."
           • "This is because of frequency and analogy with regular forms."
   • GENERALLY AGREED THAT IRREGULARS ARE STORED AND REGULARS COMPUTED
     • Irregulars idiosyncratic so can't be computed
     • PINKER (1999) DUAL MECHANISM MODEL
       • Storage
       • Computation
       • "Inflection is the most basic example of a grammatical rule being applied."
     • SINGLE MECHANISM MODELS, e.g. Rumelhart & McClelland, (1986)
       • Everything is stored
       • Words are stored in a HIGHLY CONNECTED NETWORK
       • Differences in retrival of irregulars and regulars argued to result from strength of different connections
       • "There is no explicit level of mophology."

4. OVER REGULARISATION ERRORS in children

   • The U SHAPED CURVE: Children go through a stage of producing over-inflected endings
     • "My teacher holded the baby rabbits" - correct tense, wrong form - should be irregular
     • Goes against early theories that children just imitate what they hear
     • Single mechanism accounts explain this in terms of frequency effects
       • E.g. analogy with other words: ANALOGY - call – called, crawl – crawled, fall – falled
   • Pinker takes this as evidence of acquisition of the ed rule
     • Children notice a pattern of past tense -ed rule
       • They start computing it
         • They over apply the rule
           • As vocab increases, they learn there are exceptions to the rule
             • "fall - fell"
           • "fall - falled"

5. GORDON (1985): Children aged 3-5, elicited novel compounds

   • Regular: "What would you call something that eats rats?"
   • Irregular: "What would you call something that eats mice?"
   • FOUND: Children did not produce regular plurals inside compounds
     • "A rat eater."
     • But they did often produce irregular plurals inside compounds
       • "A mice eater."
   • Compounding involves combining words stored in the mental lexicon
     • Pinker: "These findings support dual mech. model - as irregulars are stored, they can appear inside compounds."
     • N.B. majority of compounds in Eng. contain singular nouns
     • Ramscar & Dye (2011) - single mechanism: "These results exist because of question priming. Irregulars included more often because of overlap in plur. and sing. form ((mice – mouse vs. rats – rat))
       • Could this be a confound of the experimental design???

6. STANNERS et al. (1979): Primed lexical decision exp. inv reg and irreg past tense

   • Regular: Repeated prime - walk - walk
     • Past tense prime - walked - walk
   • Irregular: Repeated prime - shake - shake
     • Past tense prime - shook - shake
   • Reaction times measured
   • FOUND: Regular past tense primed the target to the same extend as repetition priming
     • But irregular past tense produced slower responses and less priming than repetition condition
     • These results compatible with supposition that regular past tense forms are decomposed: stem + affix
       • Suggests irregulars have separate lexical entry
       • Pinker: "Ahhmaygaad!!! Evidence for Dual Mechanism!!!"
       • Single Mechanism: "It's because of frequency! Regular forms are more frequent!"

7. Studies of other languages necessary to tease apart the effects of overlap in form and frequency

   • IN GERMAN REGULAR PAST LESS FREQUENT
     • REGULAR AND IRREGULAR can share the same degreee of FORMAL OVERLAP
   • SONNENSTUHL et al. (1999): Cross modal priming study (ears to eyes)
     • REGULAR
       • Repeated: P -öffne T-öffne
       • Past tense: P - geöffnet T - T-öffne
       • Control: P-wünche T- T-öffne
     • IRREGULAR
       • Repeated: P - schlafe T- schlafe
       • Past tense: P-geschlafen T - schlafe
       • Control: P-beuge T- schlaf
     • Expect quick response w repetition prime, slow resp. w control
     • FOUND: Regular past primed the stem as much as repetition priming
       • Irregulars primed the stem more than the control word, but NOT AS MUCH as repetition priming
     • Sonnenstuhl et al: "This supports the Dual Mechanism!"
       • Full priming regulars = decompsed
       • Single Mech: Can't be explained in terms of freq, or formal overlap btwn prime and target
     • These guys did supporting work on german plurals too...

Images and files in the map