Modeling early cross-situational name learning in real time

The problem of how children learn the meanings of their first words is an 
old one, going back even to the time of Augustine.  Now, with the tools 
of computational modeling available to us, we are in a better position to 
solve that problem.  Focusing only on learning names for objects, I 
propose a simple real-time probabilistic learning algorithm based on the 
Linear Reward-Penalty (LR-P) scheme (Bush and Mosteller 1951, see Yang 
2002 for an adaptation to linguistic learning) which makes use of certain 
filters on the hypothesis space, which I argue can be independently 
motivated.  While many current models (see Frank, Goodman, and 
Tenenbaum 2009 for one example) rely on complex post-hoc 
calculations, I propose a model that updates probabilities for word-to-
object mappings as new stimuli are perceived and constructs a lexicon 
from the ground up.

LR-P models with various enrichments are evaluated on hand-coded data 
from short videos of mother-infant interactions, taken from the CHILDES 
database (MacWhinney 2000).  Performance improves substantially with 
each enrichment.  The best model endows the learner with the 
assumption that words which bear some stress at the sentence level are 
more likely to refer to objects in the here-and-now than words which 
bear only lexical stress or no stress at all, and doubly filters the 
hypothesis space by 1) considering only objects that have been gestured 
to by the mother during an utterance as possible meanings for words in 
that utterance, and 2) making use of syntactic bootstrapping by 
considering nouns over verbs as possible names for objects.  I argue that 
these endowments can be independently justified and thus are 
uncontroversial additions to the model.  The satisfactory performance of 
such a simple and domain-general learning algorithm when given these 
enrichments lends support to the view that name learning is a pluralistic 
process.

References

Bush, Robert and Frederick Mosteller. 1951. A mathematical model for 
simple learning. Psychological Review 68, 313-23.

Frank, Michael, Noah Goodman, and Joshua Tenenbaum. 2009. Using 
speakers' referential intentions to model early cross-situational word 
learning. Psychological Science 20:5.

MacWhinney, Brian. 2000. The CHILDES Project:  Tools for Analyzing Talk. 
Vol. 2:  The Database (3rd ed.). Mahwah, NJ:  Erlbaum.

Yang, Charles. 2002. Knowledge and Learning in Natural Language. 
Oxford:  Oxford University Press.