Linguistics 431/631: Connectionist language modeling
September 26, 2006
Pinker and Prince
Pinker and Prince (1988) made two main critiques of the RM model
á They argue that it does not in fact behave like a child learning language
á Its failings are due to properties of connectionist systems, which means that any similar connectionist system will fail for the same reasons
Today weÕll talk about basic issues PP take with RMÕs model and their analysis of its performance.
They first consider the implicit assumptions about language made by RM.
á That the Wickelphone/Wickelfeature provides an adequate basis for phonological generalization, circumventing the need to deal with strings.
á That the past tense is formed by direct modification of the phonetics of the root, so that there is no need to recognize a more abstract level of morphological structure.
á That the formation of strong (irregular) pasts is determined by purely phonetic considerations, so that there is no need to recognize the notion 'lexical item' to serve as a locus of idiosyncrasy.
á That the regular system is qualitatively the same as the irregular, differing only in the number and uniformity of their populations of exemplars, so that it is appropriate to handle the whole stem/past relation in a single, indissoluble facility.
1. Problems with Wickelphones
á No way to represent certain distinctions
á DonÕt capture similarities
o Slit and silt have no Wickelphones in common
á Fails to exclude the impossible
o ABC -> CBA is a easy as ABC -> ABC
2. The variation of the past tense suffix is explainable as a general phonological principle
á Consider other English suffixes, like the 3rd person present indicative Ðs, or the possessive Ôs
á These show the same type of allomorphy, as well as a regular-irregular distinction
á Coda clusters must be matched for voicing Ð assimilation deals with this
á You canÕt have two identical sounds in a coda Ð schwa insertion takes care of this
á So PP critique RMÕs model for only treating the past tense, where more general phonological principles seem to apply
3. Because the relation between present and past tense is truly arbitrary, RM cannot explain the prevalence across languages of inflectional alternations that preserve of stem and affix identities.
4. Because the RM model has no representation for a lexical item, but rather is based entirely on the sounds of the present tense form, it cannot deal with homonymic forms with different past tenses
¤ He braked the car suddenly. *broke
¤ He flied out to center field. *flew
¤ He ringed the city with artillery. *rang
¤ He sleighed down the hill. *slew
¤ He spitted the pig. *spat
¤ He righted the boat. *rote
5. The RM model does not distinguish between strong and weak verb patterns. While the weak pattern is applied to any verb, the strong patterns have certain characteristics
¤ Family resemblances
¤ Failures of predictability
PP argue that the differences arise from the strong classes being memorized and the weak classes being the product of a rule (note dual-route models).
Problems with the results of the model
When RM tested the model with 72 new regular verbs, 24 produced incorrect outputs.
Some were simply perverse:
a. squat - squakt
b. mail - membled
c. tour - toureder
d. mate - maded
Along with some strange vowel changes:
a. shape - shipt
b. sip - sept
c. slip - slept
d. brown - brawned
e. mail Ð membled
Others overgeneralized the cut pattern, which should be restricted to verbs that end in t or d (as has been shown in children).
a. hug - hug
b. smoke - smoke
c. brown - brawned
Finally, some showed a strong or exclusive tendency to double marking with the regular past tense morpheme:
a. type - typeded
b. step - steppeded
c. snap - snappeded
d. map - mappeded
PP argue that this is not indicative of adult-like morphological knowledge.