Learning Words and Their Meanings from Unsegmented Child-directed Speech

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 501–509, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Learning Words and Their Meanings from Unsegmented Child-directed Speech Bevan K. Jones & Mark Johnson Dept of Cognitive and Linguistic Sciences Brown University Providence, RI 02912, USA {Bevan Jones,Mark Johnson}@Brown.edu Michael C. Frank Dept of Brain and Cognitive Science Massachusetts Institute of Technology Cambridge, MA 02139, USA mcfrank@mit.edu Abstract Most work on language acquisition treats word segmentation—the identification of lin- guistic segments from continuous speech— and word learning—the mapping of those seg- ments to meanings—as separate problems. These two abilities develop in parallel, how- ever, raising the question of whether they might interact. To explore the question, we present a new Bayesian segmentation model that incorporates aspects of word learning and compare it to a model that ignores word mean- ings. The model that learns word meanings proposes more adult-like segmentations for the meaning-bearing words. This result sug- gests that the non-linguistic context may sup- ply important information for learning word segmentations as well as word meanings. 1 Introduction Acquiring a language entails mastering many learn- ing tasks simultaneously, including identifying where words begin and end in continuous speech and learning meanings for those words. It is com- mon to treat these tasks as separate, sequential pro- cesses, where segmentation is a prerequisite to word learning but otherwise there are few if any depen- dencies. The earliest evidence of segmentation, however, is for words bordering a child’s own name (Bortfeld et al., 2005). In addition, infants begin learning their first words before they achieve adult- level competence in segmentation. These two pieces of evidence raise the question of whether the tasks of meaning learning and segmentation might mutually inform one another. To explore this question we present a joint model that simultaneously identifies word boundaries and attempts to associate meanings with words. In do- ing so we make two contributions. First, by model- ing the two levels of structure in parallel we simu- late a more realistic situation. Second, a joint model allows us to explore possible synergies and interac- tions. We find evidence that our joint model per- forms better on a segmentation task than an alterna- tive model that does not learn word meanings. The picture in Figure 1 depicts a language learn- ing situation from our corpus (originally from Fer- nald and Morikawa, 1993; recoded in Frank et al., 2009) where a mother talks while playing with var- ious toys. Setting down the dog and picking up the hand puppet of a pig, she asks, “Is that the pig?” Starting out, a young learner not only does not know that the word “pig” refers to the puppet but does not even know that “pig” is a word at all. Our model simulates the learning task, taking as input the un- segmented phonemic representation of the speech along with the set of objects in the non-linguistic context as shown in Figure …

Content truncated to 500 words