Learning Inductive Constraints: The Acquisition of Verb Argument Constructions
Amy Perfors, Charles Kemp, Josh Tenenbaum, Elizabeth Wonnacott · eScholarship (California Digital Library) · 2007
Learning Inductive Constraints: The Acquisition of Verb Argument Constructions Amy Perfors ([email protected]) Charles Kemp ([email protected]) Josh Tenenbaum ([email protected]) Elizabeth Wonnacott ([email protected]) Keywords: inductive constraints; verb argument construc- tions; computational approaches; Bayesian inference Introduction Children seeking to learn about the world are faced with the need for inductive constraints: without them, every datum may have an infinite number of possible interpretations. We suggest that hierarchical Bayesian models (HBMs) can help to explain the computational principles that might allow these constraints to be learned, and demonstrate this by exploring how a simple HBM – a Dirichlet-multinomial model – can explain important aspects of the acquisition of verb argument constructions. HBMs learn on multiple levels simultaneously, inferring which hypothesis best accounts for some data as well as which is the best higher-level hypothesis (overhypoth- esis) about the nature of the lower-level hypotheses. For instance, if the data is a set of sentences, the hypothe- ses might pertain to which verbs are grammatical in which sentence constructions (verb “sub-categorization”), and the overhypotheses might be information about what those hy- potheses could look like: do verbs tend to occur in multiple constructions, or in just one? Which constructions are most common in the language as a whole? Our model has been pre- viously shown to capture overhypotheses about feature vari- ability and the learning of the shape bias (Kemp, Perfors, & Tenenbaum, 2007). The acquisition of verb argument con- structions is additionally interesting for several reasons. Learning verb argument constructions One issue in learning verb argument constructions is whether the child can identify grammatical combinations of verbs and constructions simply from distributional information ab- stracted from exposure to the input (Pinker, 1989; Braine & Brooks, 1995). Another issue is how to capture the role of verb-specific and more abstract biases in sentence process- ing (Juliano & Tanenhaus, 1994). A recent Artificial Lan- guage Learning demonstrated that (adult) learners do acquire verb-specific statistics, although the usage of these statistics depends upon the overall distributional properties of the lan- guage, which suggests the necessity of over-hypotheses. Our model accounts for performance in these tasks. It can ac- quire both specific and general knowledge about the distrib- ution of verb argument constructions on the basis of realistic child-directed syntactic input alone (from CHILDES) and it qualitatively matches the pattern of acquisition in children, in which there is little generalization in the early stages, but later, generalization occurs, particularly with low-frequency verbs (Theakston, 2004). It can also explain why in some circumstances novel constructions are easier to learn when anchored by a highly frequent verb (Goldberg, Casenhiser, & Sethuraman, 2006). Other computational models of verb construction learning perform inference on multiple levels: for instance, the models of both Alishahi and Stevenson (2005) and Dowman (2000) can learn information about both verb constructions and spe- cific verbs. However, neither model learns about variability on both levels, which is necessary to explain the “anchor- ing” effect of highly frequent verbs (Goldberg et al., 2006) as well as people’s performance in the artificial language learn- ing tasks of Wonnacott, Newport, and Tanenhaus (in press). Furthermore, the fact that our model can also apply to non- linguistic phenomena like the shape bias suggests that it may be valuable to view aspects of domain-specific verb construc- tion knowledge as a solution to the deeper, domain-general problem of learning inductive constraints. In sum, we present a simple domain-general model of learning overhypotheses that can address a number of issues surrounding the acquisition of verb argument constructions. By learning about variability on multiple levels simultane- ously, the model captures a range of phenomena, and also suggests the utility of viewing the learning verb argument constructions as a specific application of the broader problem of learning inductive constraints in general. References Alishahi, A., & Stevenson, S. (2005). A probabilistic model of early argument structure acquisition. Proceedings of the 27th annual meeting of the Cognitive Science Society. Braine, M., & Brooks, P. (1995). Verb argument structure and the problem of avoiding an overgeneral grammar. In M. Tomasello & W. Merriman (Eds.), Beyond names of things: Young children’s acquisition of verbs (pp. 353–376). Hillsdale, NJ: Erlbaum. Dowman, M. (2000). Addressing the learnability of verb subcat- egorizations with Bayesian inference. Proceedings of the 22nd annual conference of the Cognitive Science Society. Goldberg, A., Casenhiser, D., & Sethuraman, N. (2006). Learning argument structure generalizations. In E. Clark & B. Kelly (Eds.), Constructions in acquisition. Stanford: CSLI Publications. Juliano, C., & Tanenhaus, M. (1994). A constraint-based lexicalist account of the subject/object attachment preference. Journal of Psycholinguistic Research, 23, 439–472. Kemp, C., Perfors, A., & Tenenbaum, J. B. (2007). Learning over- hypotheses with hierarchical Bayesian models. Developmental Science, 10(3), 307–321. Pinker, S. (1989). Learnability and cognition: The acquisition of argument structure. Cambridge, MA: MIT Press. Theakston, A. (2004). The role of entrenchment in children’s and adults’ performance on grammaticality judgment tasks. Cognitive Development, 19, 15-34. Wonnacott, E., Newport, E., & Tanenhaus, M. (in press). Acquiring and processing verb argument structure: Distributional learning in a miniature language.