From cbmvax!uunet!math.ucla.edu!jimc Mon Aug 12 03:57:54 1991 Return-Path: From: cbmvax!uunet!math.ucla.edu!jimc Return-Path: Message-Id: <9108120544.AA12089@luna.math.ucla.edu> To: lojban-list@snark.thyrsus.com Subject: Article about "Rules of Language" Date: Sun, 11 Aug 91 22:44:44 -0700 Status: RO A review of: Pinker, Steven, "Rules of Language", Science v253 p530 (2 Aug 1991). Info in [[ double brackets ]] is my commentary; the rest is freely paraphrased. The associationist theory of language is [[ simplified ]] that you learn language by experiencing the conjunction of words and referent events, and the brain's main job is to store and retrieve such associations. The rule theory alleges that the brain is programmed to do grammatical transformations on language input thereby to deliver useful meaning; in some versions the transformations are wholly learned while others suspect a genetic or hardwired component. Research shows that all three are true simultaneously. English past tense inflection has a specially simple form and has proven useful to probe the issue. A number of studies are reviewed in this article. The hypothesis is that regular pasts (like walk-walked) will be processed algorithmically while irregulars (like sing-sang) are associational. Therefore the two word classes should be affected differently in several classes of experiments -- and they are. Frequency affects association but not algorithms. Take an irregular verb common in the present but not the past (e.g. "Excuse me if I *forgo* the pleasure..." compared to "I *forwent* the pleasure..."); the past form sounds weird. Whereas a similar cliche with a regular verb doesn't sound weird in the past (e.g. "suffer fools gladly" vs. "suffered fools gladly"). When a person is shown a present form and is required to say the past tense as fast as possible, common irregular past forms come out faster than rare pasts (the present frequency being equal -- 16 to 29 msec difference) whereas when the verbs are regular the difference is less than 2 msec. Irregular verbs fall into similarity classes, e.g. {sing-sang, ring-rang, spring-sprang}. Imaginary words like spling which are similar to a big class are more often inflected with the class than words in a small class, suggesting that association is at work. Whereas, when a foreign word which fails to fit English verb stem morphology is inflected, the regular algorithm applies with no problem, as in "rhumba'ed". If rules and associations are handled by different brain areas, inflected words which are transformed a second time should react differently depending on whether they are regular or irregular. For example, verbs derived from nouns are always inflected regularly even if the root is irregular: "high-sticked" not "high-stuck" because the hockey stick (noun) is raised. When the verb-oid "line-drive" is presented as noun-based (hit a line drive) it is inflected regularly while if verb-based (drive a car along a line) it is irregular (line-drived vs. line-drove). A similar example [[ transitive diklujvo! ]] occurs with transitive compounds when the noun is plural; irregular plurals are accepted while regular ones are not. "Mice-infested" and "teethmarks" are OK while "rats-infested" and "clawsmarks" are not, because the irregulars are processed by association in a step upstream of compounding whereas regular plurals are processed in some downstream step and hence are not available when needed for the compound. Children show similar effects before they could possibly have learned such a pattern from listening to adults. If regular and irregular forms are handled by different brain areas, children should learn them at different times, and brain damage should affect them unequally. Children learn irregular pasts, then regular pasts afterward, and when they do they over-regularize (e.g. come-comed) when they fail to associate, a mistake they did not make earlier. Several syndromes of brain damage lead to high error rates in (among other things) regular past tenses, but irregular past tenses are affected much less, suggesting that algorithmic grammar is impaired but association is not. One such syndrome has a substantial inherited component. Another syndrome is associated with injury to Broca's area. "Focusing on a single rule of grammar, we find evidence for a system that is modular, independent of real-world meaning, non-associative (unaffected by frequency and similarity), sensitive to abstract formal distinctions (for example, root vs. derived, noun vs. verb), more sophisticated than the kinds of "rules" that are explicitly taught, developing on a schedule not timed by environmental input, organized by principles that could not have been learned, possibly with a distinct neural and genetic substrate." [[ ... coexisting with an associative word processor. ]] [[ Some of the examples given in the article suggest that there is hardware support for compound words, and at least some of their processing occurs before (or coincident with?) the algorithmic transformation of regular plurals and past tenses. A tempting conclusion (for me) is that large classes of compound words are processed algorithmically by hardware pre-provided for them. Possibly this question is decidable, but the data in the article don't give an answer to it. In any case, compound words are clearly closely related to algorithms supported by hardware. It must be understood that the hardware is capable of supporting a broad range of grammatical algorithms, but the human learns specific ones (within hardware capacity) from his environment. The visual cortex is another example in which hardware is present but has to learn from the experience of sight. Dubious is the relation of the hardware grammatical transformations to grammar rules inferred by the analytical mind. The distinction between compounds of regular vs. irregular plurals is an example of a rule which is important to the hardware but of negative interest to today's style of grammarian -- "Not irregular crap again?!" I am particularly struck that morphology, grammar and semantics are hopelessly commingled, whereas a foundation design goal of -gua!spi was to segregate them to the maximum possible degree. A second striking lesson is that both theories coexist; it is not true that there is one right answer. The brain is capable of supporting a jungle jumble grammar which drives an analyst up the wall. But that doesn't mean the Lojban (and -gua!spi) people ought to push the brain to its limit in that particular direction. Our quest for simplicity and regularity brings us other benefits that were not relevant to our randomly evolved ancestors. ]] -- jimc