This thesis presents a new theory of information modelling in natural language processing that attempts to resolve anaphoric references, while also addressing the problem of knowledge complexity. A modular model of semantic representation is introduced that addresses the deficiencies of existing representations, as well as the drawbacks associated with expanding these semantic representations. Rather than using a single semantic representation to model human knowledge and the knowledge within a sentence, the theory proposes a modular, multi-level model which is based around a semantic network. The behaviour of the model uses theories on the nature of working and long-term memory from cognitive psychology. Two methods of artificial neuron activation and decay were implemented - the ACT-R model and the Thompson model. Maximum success rates of 54.10% and 83.61% were achieved for The Three Brothers corpus, and maximum success rates of 56.00% and 86.67% were achieved for the Rumpelstiltskin corpus.
