Eated Their Own Words: Symbolic Accounts of Overregularization Errors
Biopsychology News linked to an article by Hartshorne & Ullman about sex differences in over-regularization error rates - i.e., the formation of words like "holded." Such over-regularization errors have long been a topic of interest for at least two reasons: first, they seem to reveal the overuse of a mental "rule," providing support for theorists who consider cognitive processing to be rule-based or symbolic; second, these errors have a curious developmental trajectory, such that overregularization is essentially absent among both younger speaking children and also among older children.
Although connectionist models have been successful in accounting for this U-shaped performance curve without the use of "mental rules," Hartshorne & Ullman note that such models do not predict any sex differences in overregularization errors - in fact, the authors argue that sex differences have been largely ignored in the past-tense domain.
In contrast, Ullman's "Declarative/Procedural" model of language processing does make predictions about sex differences. This theory posits that two distinct processes mediate past tense formation: first, a mental lexicon stores various instances of past-tense mappings and relies mostly on declarative memory, and second, a "mental grammar" of rules is subserved by the procedural memory system. Because females are generally better than males at verbal declarative memory tasks, Hartshorne & Ullman predicted that females may be more likely to "remember" regular past-tense forms (e.g., "walked") whereas males may be more likely to use the mental grammar mechanism to form these past-tense mappings anew each time. In other words, females may tend to use the mental lexicon, whereas males may tend to use the mental grammar mechanism.
According to this account, over-regularization errors happen when irregular past-tense forms - stored in the mental lexicon - are not sucessfully retrieved, and the irregular verb stem is thus mistakenly passed through the mental grammar rule system, where it is given a regular past-tense ("-ed"). Hartshorne & Ullman predicted that females should therefore make fewer overregularization errors than males, given their superior mental lexicon. To test this hypothesis, the authors analyzed mor ethan 100,000 utterances from 15 boys and 10 girls in the CHILDES transcript database.
Contrary to their hypothesis, Hartshorne & Ullman discovered that girls actually made more than three times more overregularization errors than boys! There were no significant sex differences in age, the size or contents of transcript samples, social class, nor the types of verb stems used by adult conversation partners in the CHILDES transcripts.
To account for this, Hartshorne & Ullman reconsidered their original hypothesis. Perhaps over-regularization errors are more prevalent among females because they are more likely to store regular past-tense forms in the mental lexicon, and thus more likely to inappropriately utilize these regular past-tense mappings in the case of irregular verb stems not through the mental grammar system, but through an associative system in the lexicon. According to this new hypothesis, female overregularization errors should occur primarily for irregular verb stems that sound similar to a variety of regular verb stems.
Hartshorne & Ullman confirmed this prediction with a sophisticated analysis of overregularization errors made by boys and girls, demonstrating that the tendency to overregularize verb stems with a high number of similar-sounding regular verb stems was significantly stronger among girls than boys. In fact, boys showed no such tendency. The authors also attempted to rule out alternative explanations for this result, including that boys' conversation partners in the CHILDES transcripts tended to use fewer similar-sounding regular verb stems, or that floor effects in boys' overregularization rates would "wash out" the correlation seen among girls (although this latter explanation still remains convincing to me, particularly given the small sample sizes here).
The authors note several limitations to their work, including that they did not demonstrate that boys' overregularization errors occur as a result of the rule-based system, that the connection between female over-regularization and associative computation is merely correlational rather than causal, and finally that the typical superiority of females in verbal memory tasks was not confirmed in this particular sample.
Hartshorne & Ullman conclude by suggesting that connectionist models of past-tense formation - which rely entirely on associative links between verb stems and past-tense forms, without recourse to symbolic rule structures - may actually explain more about past-tense formation than the current authors had given them credit for.
One shortcoming of symbolic accounts - such as Ullman's Declarative/Procedural model, or Pinker's "Words and Rules" theory - is the lack of plausible mechanisms by which "grammar rules" could be extracted from memorized verb stem/past-tense pairs (as pointed out by Tomasello). In contrast, cognitive computational mechanisms are well-specified by connectionist and PDP accounts of language learning, which reliably demonstrate the kind of "regularity extraction" (by way of hebbian learning) that would be required for the formation of rule-like representations.
Related Posts:
Disentangling Two Debates: Domain-Specificity and Nativism
Word Learning in Feature Space
Although connectionist models have been successful in accounting for this U-shaped performance curve without the use of "mental rules," Hartshorne & Ullman note that such models do not predict any sex differences in overregularization errors - in fact, the authors argue that sex differences have been largely ignored in the past-tense domain.
In contrast, Ullman's "Declarative/Procedural" model of language processing does make predictions about sex differences. This theory posits that two distinct processes mediate past tense formation: first, a mental lexicon stores various instances of past-tense mappings and relies mostly on declarative memory, and second, a "mental grammar" of rules is subserved by the procedural memory system. Because females are generally better than males at verbal declarative memory tasks, Hartshorne & Ullman predicted that females may be more likely to "remember" regular past-tense forms (e.g., "walked") whereas males may be more likely to use the mental grammar mechanism to form these past-tense mappings anew each time. In other words, females may tend to use the mental lexicon, whereas males may tend to use the mental grammar mechanism.
According to this account, over-regularization errors happen when irregular past-tense forms - stored in the mental lexicon - are not sucessfully retrieved, and the irregular verb stem is thus mistakenly passed through the mental grammar rule system, where it is given a regular past-tense ("-ed"). Hartshorne & Ullman predicted that females should therefore make fewer overregularization errors than males, given their superior mental lexicon. To test this hypothesis, the authors analyzed mor ethan 100,000 utterances from 15 boys and 10 girls in the CHILDES transcript database.
Contrary to their hypothesis, Hartshorne & Ullman discovered that girls actually made more than three times more overregularization errors than boys! There were no significant sex differences in age, the size or contents of transcript samples, social class, nor the types of verb stems used by adult conversation partners in the CHILDES transcripts.
To account for this, Hartshorne & Ullman reconsidered their original hypothesis. Perhaps over-regularization errors are more prevalent among females because they are more likely to store regular past-tense forms in the mental lexicon, and thus more likely to inappropriately utilize these regular past-tense mappings in the case of irregular verb stems not through the mental grammar system, but through an associative system in the lexicon. According to this new hypothesis, female overregularization errors should occur primarily for irregular verb stems that sound similar to a variety of regular verb stems.
Hartshorne & Ullman confirmed this prediction with a sophisticated analysis of overregularization errors made by boys and girls, demonstrating that the tendency to overregularize verb stems with a high number of similar-sounding regular verb stems was significantly stronger among girls than boys. In fact, boys showed no such tendency. The authors also attempted to rule out alternative explanations for this result, including that boys' conversation partners in the CHILDES transcripts tended to use fewer similar-sounding regular verb stems, or that floor effects in boys' overregularization rates would "wash out" the correlation seen among girls (although this latter explanation still remains convincing to me, particularly given the small sample sizes here).
The authors note several limitations to their work, including that they did not demonstrate that boys' overregularization errors occur as a result of the rule-based system, that the connection between female over-regularization and associative computation is merely correlational rather than causal, and finally that the typical superiority of females in verbal memory tasks was not confirmed in this particular sample.
Hartshorne & Ullman conclude by suggesting that connectionist models of past-tense formation - which rely entirely on associative links between verb stems and past-tense forms, without recourse to symbolic rule structures - may actually explain more about past-tense formation than the current authors had given them credit for.
One shortcoming of symbolic accounts - such as Ullman's Declarative/Procedural model, or Pinker's "Words and Rules" theory - is the lack of plausible mechanisms by which "grammar rules" could be extracted from memorized verb stem/past-tense pairs (as pointed out by Tomasello). In contrast, cognitive computational mechanisms are well-specified by connectionist and PDP accounts of language learning, which reliably demonstrate the kind of "regularity extraction" (by way of hebbian learning) that would be required for the formation of rule-like representations.
Related Posts:
Disentangling Two Debates: Domain-Specificity and Nativism
Word Learning in Feature Space
posted by Chris Chatham at 9:19:00 AM. Now go to the current Developing Intelligence blog!
0 Comments:
Post a Comment
<< Home