Doctoral Thesis Defense: Incremental and Memory-Limited Probabilistic Generative Models of Early Word Learning

Abstract

Word learning in ambiguous contexts is a challenging task which is intertwined with the process of understanding the referential intentions of the speaker. It has been suggested that infant word learning benefits from the aggregation of co-occurrence statistics of words and their referents across learning situations and that word learning can be bootstrapped by the knowledge of syntactic regularities. Previous work studied early word learning in the context of ideal learners, ignoring the incremental nature of input data as well as memory and computational resource limitations faced by learners (e.g., an infant or a robot). In my dissertation, I focus on local approaches to learning (1) which account for real-world memory constraints faced by learners and depart from computationally expensive ideal learners, (2) without a huge loss in functional performance, (3) while replicating human performance patterns. This dissertation consists of three main parts. The first part studies the problem of early word learning in isolation from the acquisition of other linguistic capabilities and presents an incremental and memory-limited computational model which describes word learning as the joint inference about the speakers’ referential intentions and the meanings of words. The second part studies the joint acquisition of language word order and words' meanings and presents a probabilistic framework for early syntactic bootstrapping to acquire the notion of "language word order" in the absence of any prior syntactic knowledge (e.g., knowledge of "subjecthood" or "objecthood"). The third part investigates the application of the proposed models in real-time word learning in robots.