5.7 “Overgeneralization”
5.8. Rule, Regular and Default
DMT claims that mental rules are indispensable for knowledge o f language. This position has been supported and defended by rigorous arguments and ample evidence (Pinker and Prince 1988, 1994; Prasada and Pinker 1993; Marcus 1995; Marcus et al.
1992; Marcus et al. 1995; Clahsen et al. 1992; Kim et al. 1994, Xu and Pinker 1995; Pinker 1999). Rules are productive, predictable, symbol-concatenating, memory-independent mental operations essential to human language processing system. To take the English case, it operates with abstract categories (e.g. symbol “stem” or “verb”), hence equally and freely extends to novel verb stems, regardless o f their phonological properties (e.g. wug —►wugs, rick — ricked)^ it does not depend on frequency or similarity (about 15% o f verbs are regular in parental speech; Marcus et al.
1992:75), or laige class size (Clahsen and Rothweiler 1993). These properties o f rules and rule-generated derivatives and inflections, according to Pinker and Prince (1994), are widely seen across languages. As Marcus et al. (1992) put it: “ This hypothesis is about a kind o f mental machinery available to all language learners but put to use in different ways in different language systems” (p. 132).
Experimental results from a number o f PDP models have questioned this view (Rumelhart and McClelland 1986; MacWhinney and Leinbach 1991; Plunkett and Marchman 1991,1993; Seidenberg 1992; Daugherty and Seidenberg 1994; Hare, Elman and Daugherty 1995). Consistently present in their arguments is a suspicion that the dichotomy seems to stem from facts about the language rather than those about the language users. That is, the rule-like behaviour is nothing more than a description o f the
Chapter 5 - Psycholinguistic Issues
English inflection system, namely the proportional distribution o f regular and irregular past tenses. For example, Rumelhart and McClelland argue that a sudden influx o f regular verbs as a part o f the child’s vocabulary growth causes the emergence o f overregularization errors (see also Marchman and Bates 1994). However, Pinker (1999) confirmed that the proportion o f regular verbs in parental speech remains constantly low - about 25 to 30% - between the age o f two and five. The English plurals, for example, show dissimilar statistics, yet we observe the same U-shaped development with plurals as with past tense. Furthermore, a growth spurt in children’s vocabulary normally comes around a year or so earlier - mid-to-late ones - than the emergence o f errors in the mid-to-late twos. Pinker concludes that language acquisition is a robust process; it is achieved despite a wide range o f input mixtures across languages and cultures, and cannot be determined just by the nuances o f parental speech statistics.
This clearly illustrates the two widely different views o f the notion o f rules which are often easily confused. Perhaps the connectionist’s view represents one o f the traditional understandings o f so-called linguistic rules. For the proponents o f DMT, however, this is the biggest misunderstanding. According to DMT, a rule is to mean a
psychological ride', hence, it does not have to be statistically general. Equally, a regular rule is “freely generalizable” in the psychological sense, and must not be mistaken with “majority pattern” in the statistical sense (Marcus et al. 1995, p.216-7). This linguistic/psychological distinction o f the notion o f rules has been strongly maintained and emphasized repeatedly in DMT. As Pinker (1999) claims, it is crucial to distinguish two very different senses o f the notion rule and its related concept regular:
Chapter 5 - Psycholinguistic Issues
“Regular almost always is equated with the pattern followed by a majority o f words in a language, or with the pattern adopted by newly coined words. But I have been using it in a different sense, one that pertains to the mental processes o f speakers rather than to numbers and citations in a dictionary. “Regular” here refers to a rule that speakers treat as the default an inflectional pattern they can apply to any word in a category, even if the word has never been stored with that pattern, or with any pattern, in memory. According to this theory, a regular pattern could, in principle, apply to a minority o f words in a language, with the majority having to be learned one by one.
And the rule could fail to apply to a new word if the word is so similar to irregular words in memory that analogy is irresistible (as in to spling, which most people inflect as splang or splung. The only way to know whether an inflection is regular in the psychological sense is to see whether people apply it when their memory is blocked: when the word is new, rare, unusual, foreign, rootless, or headless.”
(Italics original; Pinker 1999:214)
So, the mere existence o f common patterns or productive generalization to novel forms does not distinguish regular inflections from irregulars (because irregulars do display both). The important claim here is that real rules do not require access to memorized forms or their sound patterns, but applies as the default whenever memory access is not possible. In other words, a rule is psychologically real and regular only when it manifests itself in a situation where analogies fail. It must be a kind o f generalization that is independent from analogies to stored patterns. Marcus et al. (1995) list 21 o f such “circumstances in which memory patterns are not accessed and regular inflection is applied” (p. 197). This is shown in Table 4.
Chapter 5 - Psycholinguistic Issues
Table 4.
‘Circumstances in which memory patterns are not accessed and regular inflection is applied’ (Marcus et al. 1995:197)
Circumstance Kind o f word Example
Lack o f entry or similar entries in memory
1. No root entiy Novel words snarfed, wugs
2. Weak entry Low-frequency words stinted, eked