This project will investigate the comprehension of morphologically complex words, words like "knowable" that can be broken down into pieces. Linguists and psycholinguists disagree on whether all words that consist of a possibly independent stem ("know" is a verb by itself) and an identifiable suffix ("-able" generally creates adjectives from verbs) are analyzed as complex by speakers and whether all or any such words are recognized via decomposition into their parts. The issue is particularly controversial for words like "tolerable" apparently built from roots that do not appear on their own ("toler-" is also seen in "tolerate" but not elsewhere). This project uses magnetoencephalographic (MEG) brain monitoring methods to test a theory about the interaction of linguistic representations and neural computations. The theory demands that every decomposition motivated by linguistic theory is a necessary computational step in recognizing a word and that each computation maps to neural activity of a set of brain regions at particular time latencies during word recognition. For this project, subjects will read or listen to individual words and non-words, judging their word status, while the electrical activity in their brains is monitored with MEG. A novel analysis technique will be developed that correlates the brain responses of each subject for each word with continuous stimulus variables such as word frequency, suffix frequency, and the probability of having a particular suffix following a particular stem. Given sufficient numbers of subjects and stimuli, this technique can provide meaningful data about individual words and individual subjects. For example, is a word like "vulnerable", whose root "vulner-" does not appear elsewhere in English, recognized in the same way as "knowable" or "tolerable"?
The project should elucidate the connections among linguistic theory, psycholinguistic models, and brain activity while testing hypotheses about the comprehension of morphologically complex words. Since for the theories of morphology being tested, the structure of words involves the same computations and representations as the structure of sentences, support for full decomposition to the root for words like "knowable" and "tolerable" will have implications for language processing at all levels of linguistic analysis. The single trial MEG analysis techniques developed by the project should aid greatly in the diagnosis of language-impaired populations and in the evaluation of remediation. Exploiting the results and techniques of this project, future research can ask, for example, how the brain responses of an individual dyslexic or autistic subject differ from the norm along various dimensions, and if any of these responses approach group norms after intervention. Progress in understanding and treating these deficits should come from an understanding of how neural systems perform linguistic computations and store and manipulate linguistic representations.
The smallest units of language that contribute to the meaning and structure of sentences are called "morphemes." Languages combine morphemes into words, as in English "sing-er" (verb "sing" and suffix "er," "one who sings"). Languages are structured such that pieces consisting of many parts, e.g., "[I saw] the very tall old man that works for the city," can take the place in a sentence of a piece consisting of only a single part, e.g., "[I saw] Peter." Such considerations have led psycholinguists to hypothesize that speakers of a language might treat apparently morphologically complex words (consisting of more than one morpheme), e.g., "marriage" (verb, ‘to marry,’ plus "-age"), as single units in processing, particularly if the words are common and if the meanings of the words are not predictable from the meaning and function of the constituent morphemes. On the other hand, linguists have reached the opposite conclusion: all words must be decomposed down to the smallest morpheme units, and accurate processing of language requires this full decomposition. If psycholinguists were correct that decomposition is not a necessary stage in word recognition, the consequence would be a growing divide between the account of language developed by linguists and the accounts of language processing developed in cognitive neuroscience. On the other hand, if we could demonstrate full decomposition in word recognition, linguistic theory could be fully integrated within the cognitive neuroscience of language. The work supported by the NSF has verified the full decomposition account of the recognition of morphologically complex words, while clarifying the neural bases of the cognitive processes underlying word recognition. The bulk of the research completed under the project involved participants reading individual words (visual word recognition). Here we developed novel techniques for analyzing data from MEG (magnetoencepholography), which measures the magnetic fields generated by neural activity in the cerebral cortex. By correlating MEG measures with properties of the words our participants read on a word-by-word basis, we were able to track brain activity computing different properties of the words millisecond by millisecond. Our results support the following stages in complex word recognition: first, recognition of the letters in visual cortex (~100ms after word onset), second, decomposition of the complex words into morphemes based on the visual forms of the morphemes (~150ms, inferior temporal lobe), third, retrieval of stored (lexical) information about the meaning and other properties of the morphemes (~300ms, medial and superior temporal lobes), and finally, recomposition of the morphemes into the whole word (starting around 300ms, involving various brain regions). The most striking results of our experiments were demonstrations (1) that words like "brother" are (wrongly) decomposed into "broth" and "er" in the second stage of processingå, as expected on decomposition theories where the analysis of words is first based on visual forms or morphemes; (2) that the relative frequencies of the meanings of homophones like "bank" (river or financial) are relevant to word recognition only at 300ms, during the lexical stage of processing, and not before, as predicted by account in which morpheme recognition by form precedes lexical recognition connected to meaning; and (3) that the frequency of the whole morphologically complex word ("singer" as opposed to "sing") only correlates with brain activity relatively late (~350-400ms) in word recognition, as predicted by the full decomposition theory. Toward the end of our grant, we were able to extend our finding to auditory word recognition. Here we demonstrated that in listening to words, speakers predict the upcoming sounds based on the likelihood of the continuations of what they have already heard to be words that they know. However, as predicted by the full decomposition model of word recognition, these predictions are modulated by predictions about morphemes that might follow the stems that they have heard (so, hearing "sing," "er" is a possible continuation as a suffix). These studies exploit our new analysis techniques to monitor the brain’s response to auditory words during the presentation of the words themselves, correlating properties of the sounds in individual trials with the brain responses from various brain regions millisecond by millisecond. As Steve Pinker explained in The Language Instinct, morphology involves all aspects of language in miniature – understanding the processing of morphologically complex words, then, already provides insight into language processing more generally. Developmental and acquired language deficits often involve the production of words missing the morphological pieces normally required by the language, as well as simplification of sentence structure. The basic science from this project on the neural basis of the comprehension of morphologically complex words should help lead to a better understanding of the breakdown that results from genetic deficits or brain injury.
