Claudius Gläser, "Making Sense of Words through the Eyes of a Child: A Computational Framework for the Acquisition of Word Meanings", 2011.

Equipping machines with the ability to understand and use natural language is a difficult task. One aspect underlying this task is the acquisition of language semantics or – to narrow the problem even more – the learning of individual word meanings. A machine, which copes with this problem, has to learn the meaning of a word based on observations of the word in different contexts. Children perform marvelously well in this task. Even though it still remains an open question how children acquire word meanings so efficiently, research in the fields of developmental psychology and neurobiology start to shed light onto some aspects of the underlying learning principles. Designing an artificial system based on such findings may consequently lead a way to overcome the restrictions of existing approaches, thereby striving towards child-like learning abilities. In this thesis, a computational framework for the acquisition of word meanings is presented. The framework is largely inspired by findings on child development and learning. It hence establishes a link between the individual disciplines of developmental psychology, neurobiology, and computer science. Therefore, the thesis is structured around three central issues: Firstly, based on the abundant literature on word learning by children the different ways how children acquire word meanings are discussed. Secondly, the thesis not only discusses the respective learning processes from a phenomenological point of view, but also aims at identifying commonalities with more detailed theories on neuronal learning. More precisely, I will argue that specific neurobiological learning principles can explain the developmental patterns observed in children and, hence, may constitute the biological underpinnings of the learning processes. Lastly, based on this unified viewpoint, biologically inspired computational models for the acquisition of word meanings are presented and applied in selected word learning scenarios. In summary, this thesis investigates a multitude of aspects that contribute to the word learning capabilities of children. It thereby promotes the view that different learning processes and biases have to be taken into account when trying to construct artificial systems with child-like learning skills. For the individual aspects, it is shown that the development of biologically inspired computational models indeed constitutes a viable approach as compared to other methods. The tight integration of the different models into a coherent overall system for word meaning acquisition, however, is necessary to finally build robots that exhibit the desired capabilities. This integration may constitute the biggest challenge future research has to overcome.