Rule learning by seven-month-old infants

Science

Volumn 283, Issue 5398, 1999, Pages 77-80

  Marcus, G F ^a   Vijayan, S ^a   Bandi Rao, S ^a   Vishton, P M ^b  

^a NEW YORK UNIVERSITY   (United States)

^b AMHERST COLLEGE   (United States)

Author keywords

[No Author keywords available]

Indexed keywords

ARTICLE; AUDITORY DISCRIMINATION; CHILD DEVELOPMENT; CONTROLLED STUDY; HUMAN; INFANT; LANGUAGE DEVELOPMENT; LEARNING; NERVE CELL NETWORK; NORMAL HUMAN; PRIORITY JOURNAL; THINKING;

BRAIN; COGNITION; HUMANS; INFANT; LANGUAGE DEVELOPMENT; LEARNING; MATHEMATICS; NEURAL NETWORKS (COMPUTER); NEURONS;

EID: 0032935801     PISSN: 00368075     EISSN: None     Source Type: Journal    
DOI: 10.1126/science.283.5398.77     Document Type: Article

References (32)

1. J. Saffran, R. Aslin, E. Newport, Science 274, 1926 (1996).
2. J. L. Elman, Cogn. Sci. 14, 179 (1990).
3. E. Bates and J. Elman, Science 274, 1849 (1996); M. S. Seidenberg, ibid. 275, 1599 (1997).
4. E. Bates and J. Elman, Science 274, 1849 (1996); M. S. Seidenberg, ibid. 275, 1599 (1997).
5. N. A. Chomsky, Rules and Representations (Columbia Univ. Press, New York, 1980); S. Pinker and A. Prince, Cognition 28, 73 (1988); S. Pinker, Science 253, 530 (1991); G. F. Marcus, U. Brinkmann, H. Clahsen, R. Wiese, S. Pinker, Cogn. Psychol. 29, 186 (1995). G. F. Marcus, The Algebraic Mind (MIT Press, Cambridge, MA, 1999), in press.
6. N. A. Chomsky, Rules and Representations (Columbia Univ. Press, New York, 1980); S. Pinker and A. Prince, Cognition 28, 73 (1988); S. Pinker, Science 253, 530 (1991); G. F. Marcus, U. Brinkmann, H. Clahsen, R. Wiese, S. Pinker, Cogn. Psychol. 29, 186 (1995). G. F. Marcus, The Algebraic Mind (MIT Press, Cambridge, MA, 1999), in press.
7. N. A. Chomsky, Rules and Representations (Columbia Univ. Press, New York, 1980); S. Pinker and A. Prince, Cognition 28, 73 (1988); S. Pinker, Science 253, 530 (1991); G. F. Marcus, U. Brinkmann, H. Clahsen, R. Wiese, S. Pinker, Cogn. Psychol. 29, 186 (1995). G. F. Marcus, The Algebraic Mind (MIT Press, Cambridge, MA, 1999), in press.
8. N. A. Chomsky, Rules and Representations (Columbia Univ. Press, New York, 1980); S. Pinker and A. Prince, Cognition 28, 73 (1988); S. Pinker, Science 253, 530 (1991); G. F. Marcus, U. Brinkmann, H. Clahsen, R. Wiese, S. Pinker, Cogn. Psychol. 29, 186 (1995). G. F. Marcus, The Algebraic Mind (MIT Press, Cambridge, MA, 1999), in press.
9. N. A. Chomsky, Rules and Representations (Columbia Univ. Press, New York, 1980); S. Pinker and A. Prince, Cognition 28, 73 (1988); S. Pinker, Science 253, 530 (1991); G. F. Marcus, U. Brinkmann, H. Clahsen, R. Wiese, S. Pinker, Cogn. Psychol. 29, 186 (1995). G. F. Marcus, The Algebraic Mind (MIT Press, Cambridge, MA, 1999), in press.
10. R. Bijeljac-Babic, J. Bertoncini, J. Mehler, Dev. Psychol. 29, 711 (1993).
11. R. L. Gomez and L.-A. Gerken, in Boston University Conference on Language Development 21, E. Hughes, M. Hughes, A. Greenhill, Eds. (Cascadilla Press, Somerville, MA, 1997), p. 194.
12. We leave open the question of whether infants interpreted our materials as genuinely linguistic and thus also leave open the question of whether the mechanisms that acquire abstract rules are specific to language learning or are more generally used in many domains.
13. P. Jusczyk and R. N. Aslin, Cogn. Psychol. 29, 1 (1995).
14. Infants sat in a three-sided booth on the laps of their parents (parents wore headphones playing classical music so that they could not hear the stimulus materials) and listened to sounds generated off-line by a speech synthesizer. The booth had a yellow bulb on the center panel; each side panel had a red bulb. A speaker was behind each of the red bulbs. The speakers were connected to a G3 Power Macintosh computer that presented the stimuli and controlled the lights. During the familiarization phase, the yellow light flashed to draw the infant's attention to the center panel of the testing booth while the familiarization speech segment played from both speakers. After the familiarization ended, the infant was presented with test trials. At the beginning of each test trial, the central light was flashed. Once an observer (who also wore headphones playing music to mask the stimuli) indicated that the infant had fixated on the flashing light, the central light was turned off and one of the two side lights began flashing. When the observer indicated that the infant had turned toward the side light, the computer played a three-word test sentence from the speaker that was hidden behind the light, which repeated the test sentence over and over (with a 1.2- to 1.5-s pause between presentations of the test sentence) until either the infant had turned away for two continuous seconds or until 15 s had elapsed. The dependent measure was the total time that the infant spent looking at the light associated with the speaker. Infants who became fussy prior to completion of at least four test trials were not included in the statistical analyses.
15. The first six subjects (three in each condition) were familiarized with 3-min speech samples.
16. The 16 sentences that followed an ABA pattern were "ga ti ga," "ga na ga," "ga gi ga," "ga la ga," "li na li," "li ti li," "li gi li," "li la li," "ni gi ni," "ni ti ni," "ni na ni," "ni la ni," "ta la ta," "ta ti ta," "ta na ta," and "ta gi ta." The 16 sentences that followed the pattern ABB were "ga ti ti," "ga na na," "ga gi gi," "ga la la," "li na na," "li ti ti," "li gi gi," "li la la," "ni gi gi," "ni ti ti," "ni na na," "ni la la," "ta la la," "ta ti ti," "ta na na," and "ta gi gi". Vocalizations of the words used in the above sentences were created with a speech synthesizer, which is available at www.bell-labs.com/ project/tts/voices-java.html. The vocalizations were then combined to form the sentences listed above by using a sound editor. A 250-ms pause was placed between consecutive words in each sentence. The sentences were presented in random order and separated by pauses of 1 s.
17. The 12 test trials, which were randomly ordered, included three repetitions of each of four test sentences, two following the ABB pattern ("wo fe fe" and "de ko ko") and two following the ABA pattern ("wo fe wo" and "de ko de").
18. Similar stimuli were used in a study of children's memory and attention [J. V. Goodsitt, P. A. Morse, J. N. Ver Hoeve, Child Dev. 55, 903 (1984)]. That study does not, however, answer our question about rules, because it tested only how well an infant could remember target B in the context of sequences ABA versus AAB versus ABC and not whether infants familiarized with one of those sequences could distinguish it from another.
19. Results for the ABA and ABB conditions were combined, because there was no significant interaction between them, F(1,14) = 0.15.
20. Similar results involving transfer from one finite state grammar to another with the same structure but different words have been reported for adult subjects [A. Reber, J. Exp. Psychol. 81, 115 (1969)] and for 11-month-old infants (R. L. Gomez and L.-A. Gerken, paper presented at the Annual Meeting of the Psychonomics Society, Philadelphia, PA, November 1997). These researchers, whose focus was not on rule learning, did not include the phonetic control we introduce in experiments 2 and 3.
21. Similar results involving transfer from one finite state grammar to another with the same structure but different words have been reported for adult subjects [A. Reber, J. Exp. Psychol. 81, 115 (1969)] and for 11-month-old infants (R. L. Gomez and L.-A. Gerken, paper presented at the Annual Meeting of the Psychonomics Society, Philadelphia, PA, November 1997). These researchers, whose focus was not on rule learning, did not include the phonetic control we introduce in experiments 2 and 3.
22. The 16 habituation sentences that followed the ABA pattern were "le di le," "le je le," "le li le," "le we le," "wi di wi," "wi je wi," "wi li wi," "wi we wi," "ji di ji," "ji je ji," "ji li ji," "ji we ji," "de di de," "de je de," "de li de," "de we de"; ABB items were constructed with the same vocabulary. The test trials were "ba po ba," "ko ga ko" (consistent with ABA), "ba po po," and "ko ga ga" (consistent with ABB).
23. Results for the ABA and ABB conditions were combined, because there was no significant interaction between them, F(1,14) = 1.95.
24. In principle, an infant who paid attention only to the final two syllables of each sentence could distinguish the AAB grammar from the ABB grammar purely on the basis of reduplication, but they could not have succeeded in the experiment of Saffran et al. (1).
25. We thank an anonymous reviewer for suggesting this comparison. The vocabulary used to construct the test and familiarization items was the same as in experiment 2; hence, as in experiment 2, the phonetic features that distinguished the test words from each other did not vary in the habituation items.
26. The ability to extend reduplication to novel words appears to depend on an algebraic rule. To recognize that an item is reduplicated, a system must have the ability to store the first element and compare the second element to the first; the storage, retrieval, and inferential mechanisms that are involved may appear simple but are outside the scope of most neural network models of language and cognition. Conversely, adults are strongly sensitive to the presence of reduplication and its location in phonological constituents [I. Berent and J. Shimron, Cognition 64, 39 (1997)]. For further discussion, see references cited in (22).
27. Results for the AAB and ABB conditions were combined, because there was no significant interaction between them, F(1,14) = 0.002.
28. The sort of generalizations that such models can draw are dictated by the choice of input representations. If input nodes correspond to words, the model cannot generalize the abstract pattern to new words; if the input nodes correspond to phonetic features, the model cannot generalize to words containing new phonetic features [G. F. Marcus, Cognition 66, 153 (1998); Cogn. Psychol., in press]. An appropriately configured SRN that represented each word by a set of nodes for phonetic features, if it were trained that a voiced consonant followed by an unvoiced consonant was always followed by a voiced consonant, could use memorized sequences of features as a basis to distinguish the test items in experiment 1. However, such a model could not account for the results of experiments 2 and 3, because in those experiments the feature sequences that the network learned about in the familiarization phase would not distinguish the test items.
29. The sort of generalizations that such models can draw are dictated by the choice of input representations. If input nodes correspond to words, the model cannot generalize the abstract pattern to new words; if the input nodes correspond to phonetic features, the model cannot generalize to words containing new phonetic features [G. F. Marcus, Cognition 66, 153 (1998); Cogn. Psychol., in press]. An appropriately configured SRN that represented each word by a set of nodes for phonetic features, if it were trained that a voiced consonant followed by an unvoiced consonant was always followed by a voiced consonant, could use memorized sequences of features as a basis to distinguish the test items in experiment 1. However, such a model could not account for the results of experiments 2 and 3, because in those experiments the feature sequences that the network learned about in the familiarization phase would not distinguish the test items.
30. An enhanced version of the SRN [Z. Dienes, G. T. M. Altmann, S. J. Gao, in Neural Computation and Psychology, L. S. Smith and P. J. B. Hancock, Eds. (Springer-Verlag, New York, 1995)] aims to model how speakers who are trained on one artificial language are able to learn a second artificial language that has the same structure more rapidly than a second artificial language that has a different structure. This model would not be able to account for our data, however, because the model relies on being supplied with attested examples of sentences that are acceptable in the second artificial language, whereas our infants succeeded in the absence of such information.
31. The problem is not with neural networks per se but with the kinds of network architectures that are currently popular. These networks eschew explicit representations of variables and relationships between variables; in contrast, some less widely discussed neural networks with a very different architecture do incorporate such machinery and thus might form the basis for learning mechanisms that could account for our data [J. E. Hummel and K. J. Holyoak, Psychol. Rev. 104, 427 (1997)]. Our goal is not to deny the importance of neural networks but rather to try to characterize what properties the right sort of neural network architecture must have.
32. Supported in part by an Amherst College Faculty Research Grant to P.M.V. We thank L. Bonatti, M. Brent, S. Carey, J. Dalalakis, P. Gordon, B. Partee, V. Valian, and Z. Zvolenszky for helpful discussion, and P. Marcus and F. Scherer for their assistance in construction of the test apparatus. We also thank Bell Labs for making available to the public the speech synthesizer that we used to create our stimuli. Some subjects in experiment 1 were tested at Amherst College all other subjects were tested at New York University. The parents of all participants gave informed consent.