Alec Marantz and Liina Pylkkänen (Massachusetts Institute of Technology)
Morphological Decomposition in Derivation: MEG Evidence and Some Theoretical Context
Background. A recent trend in both anti-linguistic connectionist theories (Seidenburg and Gonnerman 2000) and in Optimality-inspired morphophonological theories (Burzio 2000) is to suppose that all words are memorized in the mental Lexicon, with lexical-relatedness accounting for the appearance (illusion) of (constructive) derivational affixation and for the apparent internal syntax of words.
Seidenberg expresses the "all words are memorized" view in a particularly pointed way, by claiming that there is no "teach" in "teacher." Rather, such apparently morphologically related words involve independent representations, connected by strong phonological and semantic similarity. We can compare the radical connectionist (RC) view with a radical version of Distributed Morphology (Marantz 2000), which claims that syntax permeates every combination of morphemes. On this (RDM) view, "teacher" not only decomposes into "teach" and "er," the composition is syntactic in the same sense that composition of "the" and "cat" in "the cat" is.
New Neurolinguistic Evidence for Syntactic Decomposition of Derived Words. Our MEG experiment does not directly compare radical connectionism to radical DM but is limited to testing the claim of RC that "teach" and "teacher" are separate words connected only via phonology and semantics; any theory of morphology in which "teacher" contains "teach" would be compatible with our results. We employed a cross-modal priming paradigm with a lexical decision task, essentially replicating the behavioral results of Gonnerman (2000). Crucial contrasts in this experiment compare phonologically related (auditory)prime-(orthographic)target pairs such as "spinach-SPIN" and "teacher-REACH" with morphologically related pairs such as "teacher-TEACH." Phonological similarity between prime and target actually slows down lexical decision on the target in the first two cases, but by different mechanisms. In the case of "spinach-SPIN" the initial match between the first part of "spinach" and the word "spin" causes lateral inhibition of the representation of "spin" by "spinach" during auditory processing of "spinach" and thus a slow down of the MEG M350 response component to "spin," since the M350 indexes initial activation of lexical representations (Embick et al. 2001, Pylkkänen et al. 2001). In the case of rhyme matches like "teacher-REACH," phonological relatedness activates "reach" during auditory processing of "teacher" and thus speeds up the M350 response to "reach" - but competition between independent representations (of morphemes sharing the "each" sound) in the Lexicon slows down reaction time (Vitevich and Luce 1998). "Teacher-TEACH," which should, on the RC story, behave with respect to the M350 as does "spinach-SPIN" with an added effect of semantic priming, in fact shows priming of the M350 and also faster reaction time, strongly arguing against the hypothesis that "teacher" and "teach" have separate lexical representations. On the RC story we should expect lateral inhibition of "teach" during auditory processing of "teacher" (since the listener must process that s/he did not hear "teach"), plus competition between "teacher" and "teach" during the lexical decision on "teach." The morphological decomposition story correctly predicts the results since recognition of "teacher" is recognition of "teach" and another morpheme - syntactically combined, predicting priming of "teach" and no competition from the (non-existent) lexical representation of "teacher."
Theoretical Context: At What Cost Any Retreat from Radical Distributed Morphology? The brain evidence for "teach" in "teacher" should be set in the context of a theoretical discussion of radical DM in comparison to less radical alternatives. RDM claims that any feature structure that one may create via syntactic composition must always be created via syntactic composition both within a language and crosslinguistically. In contrast, RC claims that complex words must wear their syntactic features as properties of the words as a whole; thus the internal make-up of words is essentially irrelevant to the syntactic properties of words. In addition, if the feature-complex constructed in a syntactic phrase is equally well expressed as features of a memorized word, there can be no necessity to the particular syntactic realization of that feature structure. Although he doesn't endorse any radical anti-decompositionalism, Jackendoff's recent attempts to explicate a minimalist program that incorporates ideas from Construction Grammar leads to the same conclusion: since monomorphemes may map onto feature structures expressable in phrasal syntax, complex feature structures do not imply syntactic structure either above or below the level of the word and thus any constrained theory of syntax (any theory that claims that syntactic structure is motivated to create the types of feature structures one gets from combining morphemes and words) is on the wrong track. We will briefly discuss the following issue in light of the experimental results: does any theory allowing a monomorpheme to carry the same feature structure as a derived word necessarily lead to a Jackendoff-style theory in which there is no necessary relation between syntactic structure and feature structure? That is, what are the theoretical consequences of any departure from RDM?
© INSTITUT DE LINGÜÍSTICA APLICADA - UNIVERSITAT POMPEU FABRA