English Past Tense by peisty474

VIEWS: 897 PAGES: 18

									Psycholinguistics                         Past Tense                  1
   '                                                                  $


                          English Past Tense

      Traditional Characterization: A Rule Plus Exceptions
       1. General Rule: verb + /-ed/
       2. Memorized Exceptions: sing → sang, bring → brought, . . .
       3. Blocking Principle: more specific (2) blocks (1).
      Acquisition
       1. Children initially memorize forms.
       2. Children learn and overapply rule.
       3. Children learn exceptions that block the rule.


   &                                                                  %
Psycholinguistics                         Past Tense                     2
   '                                                                     $



                    Nativism Vs. Empiricism 1

      Implicit in Traditional Characterization:
       1. Children are innately predisposed to posit rules.
       2. Children innately know the blocking principle.
       3. Fundamental division between regular and irregular forms.
      Poverty of the Stimulus Argument:
       1. Children go from finite input to infinite generalization.
       2. How could the blocking principle be inferred from the input?



   &                                                                     %
Psycholinguistics                           Past Tense                     3
   '                                                                       $

                     Nativism Vs. Empiricism 2

      Alternative:
       1. Only general learning mechanisms.
       2. Learning can be done on the basis of the input alone.
       3. No division: children learn general rule and exceptions in the
          same manner.
      Responding to the Poverty of the Stimulus Argument:
       1. Generalization can emerge from efficient representation.
       2. System configures itself to minimize error—no explicit
          generalization (rule).
       3. Large role for statistical frequencies.
   &                                                                       %
Psycholinguistics                         Past Tense                        4
   '                                                                        $




                        Connectionist Models

        • Network of connected units: input units linked to output units.
        • Influence of input unit on output unit depends on the strength
          of the connection between them.
        • Strength of connections between units altered by training.
        • Attempts to show learning can take place just on the basis of
          the input.




   &                                                                        %
Psycholinguistics                         Past Tense               5
   '                                                               $



                    Rumelhart and McClelland 1986

        • Input: 460 units representing sounds
        • Every unit connected to every other unit
        • Links are weighted, weight modifiable by training
        • Output: 460 units representing sounds
      Learning Phase:
        • 420 verbs presented 200 times (84,000 trials)
        • Past tense form computed by the network is compared to
          correct form, weights adjusted to reduce the difference


   &                                                               %
Psycholinguistics                        Past Tense                       6
   '                                                                      $




                                 Findings

       1. Acquired hundreds of regular and irregular verbs.
       2. Generalized properly to new verbs.
       3. Appeared to go through U-shaped development.
       4. Produced blends sometimes produced by children (give-gaved ).




   &                                                                      %
Psycholinguistics                         Past Tense                        7
   '                                                                        $


                                 Criticism

       1. U-shaped development results from first learning
          high-frequency irregulars, then being swamped by influx of
          hundreds of regulars. BUT:
          (a) Proportion of regular verbs in parental speech constant
              throughout relevant period (30%).
          (b) Spurt in vocabulary growth occurs a year earlier than onset
              of overregularization.
       2. There is no consistent correlation between overregularization
          rates of different verbs and the sums of the frequencies of
          rhyming verbs.


   &                                                                        %
Psycholinguistics                          Past Tense                      8
   '                                                                       $



                            Criticism (Cont.)

       3. Errors are not based on sound:
          (a) Homophonous verbs can have different past tense forms
              (ring-rang, wring-wrung).
          (b) Do, have, be never overregularized as auxiliaries, but are
              overregularized as main verbs.
          (c) Denominal/deadjectival verbs are always regular, even when
              based on irregular verbs (grandstanded, high-sticked ).
       4. The network produced errors not found in children
          (mail →membled ).


   &                                                                       %
Psycholinguistics                          Past Tense                          9
   '                                                                           $


                    Rule + Blocking + Memory Failure

      Pinker’s theory:
       1. Children’s errors are a result of memory failure; children fail to
          retrieve the irregular forms fast enough (hold >held ).
       2. English requires tense to be marked; so they apply regular rule
          (hold →holded ).
       3. Prior to having learned the rule, when memory fails they do
          nothing (hold ).
       4. After enough exposure memory improves, and they achieve
          adult performance (hold →held ).


   &                                                                           %
Psycholinguistics                         Past Tense                      10
   '                                                                      $


                                 Evidence

      Marcus et al. 1992
       1. Errors are “sporadic malfunctions”: mean rate across children
          only 4.2 percent, holding steady ages 2–5.
       2. Error rates correlate with frequency of a given irregular in
          parental speech.
       3. Onset of overregularization coincides with onset of adding
          suffix to regular verbs (acquiring the rule).
       4. Overregularization errors replace error of doing nothing, not
          correct forms.
       5. Children judge overregularization as errors (Kuczaj 1978).

   &                                                                      %
Psycholinguistics                        Past Tense                        11
   '                                                                       $



                           Evidence (Cont.)

       6. Anecdotal: children get mad when adults use
          overregularization.
       7. Adults occasionally make errors too (getting worse as they get
          older).
       8. Adults unsure about past tense of some verbs (dived ? dove?).
       9. Over time, infrequent irregulars have been regularized.
      10. Irregular verbs that do survive are among the most common
          verbs in the language.


   &                                                                       %
Psycholinguistics                         Past Tense                 12
   '                                                                 $


                    Later Connectionist Models 1

      Plunkett and Marchman 1991
        • Multilayered network.
        • Training set of 500 verbs, no discontinuities.
        • Distinction between type and token frequencies.
      Findings:
       1. Type and token frequencies that worked best are those of
          English: low type frequency but high token frequency for
          irregulars.
       2. Regular verbs go through U-shaped development.

   &                                                                 %
Psycholinguistics                         Past Tense                   13
   '                                                                   $




                    Later Connectionist Models 1

      Findings (Cont.):
       3. Irregular verbs go through multiple micro U-shaped curves.
       4. High token frequency irregulars learned well, few errors.
       5. No initial period of correct past tense formation.




   &                                                                   %
Psycholinguistics                          Past Tense                      14
   '                                                                       $



                    Later Connectionist Models 2

      Plunkett and Marchman 1993
        • Initial training period of 10 regular and 10 irregular verbs.
        • Then one verb at a time, each 80% likely to be regular.
        • Tested regularly on learned verbs and novel verbs.
      Findings:
       1. Initial period without overregularization (extending well into
          growth stage).
       2. Low rate of overregularization (<10%).


   &                                                                       %
Psycholinguistics                        Past Tense                       15
   '                                                                      $




                    Later Connectionist Models 2

      Findings (Cont.):
       3. During earliest phases no suffix is added to novel verbs (rote
          learning only).
       4. Overregularization onset claimed to correlate with “critical
          mass” of regular verbs in children’s vocabulary (Marchman and
          Bates 1994).




   &                                                                      %
Psycholinguistics                        Past Tense                        16
   '                                                                       $




                                Criticism

       1. Even these models cannot account for grammatical factors
          (auxiliaries, denominal verbs) without building in grammatical
          knowledge.
       2. These models also cannot handle homophonous verbs with
          different past tenses (ring-rang, wring-wrung).




   &                                                                       %
Psycholinguistics                         Past Tense                          17
   '                                                                          $



                          A Role for Analogy?

      Irregulars are analogized to other irregulars that sound like them
      (sink-sank, drink-drank, shrink-shrank ):
       1. Children overregularize less often irregulars that are similar to
          other irregulars.
       2. Children sometimes over-irregularize: wipe-wope.
       3. Adults create new irregulars on the basis of analogy:
          sneak-snuck.
      Pinker: Both rules and analogy-based networks might be necessary
      to characterize linguistic knowledge.


   &                                                                          %
Psycholinguistics                          Past Tense                      18
   '                                                                       $




                                Final Points

        • Purely emergent systems operating without constraints do not
          accurately model acquisition of the past tense in Englsh. But:
        • Connectionist models have been proposed that incorporate
          innate knowledge/constraints.
        • Assumption of innate knowledge does not entail symbolic
          computation/rules.
        • It is possible for symbolic rules and connectionist-style
          representations to co-exist.



   &                                                                       %

								
To top