This paper presents a computational model of word learning that has roots in experimental literature and learns in real time with high precision from a small amount of data. In addition to incorporating external cues a la Yu & Ballard 2007, we give the learner the ability to test specific highly probable semantic hypotheses against new data. Performance is comparable to that of a more complex model (Frank et al. 2009) and better than that of a similar model (Fazly et al. 2010) that does not utilize hypothesis evaluation.
"Using salience and hypothesis evaluation to learn object names in real time,"
University of Pennsylvania Working Papers in Linguistics:
1, Article 23.
Available at: http://repository.upenn.edu/pwpl/vol18/iss1/23