The thesis is divided into two major parts. In the first part of the thesis background information on connectionist models in general will be provided followed by its implication and applicability on language processing. Two main streams of connectionist models, the simple recurrent network (SRN) and the recursive auto-associative memory (RAAM) will be reviewed along with their related research work in the language processing domain. Focus will be put on the RAAM model along with author's implementations and computer simulations of the network with the given language related tasks. The author's originality in modification of the model will be highlighted. Finally, by pointing out issues related to the problem faced by conventional artificial neural network training, the discussion will lead to the second part of the thesis. Through experimenting with computer simulations of the RAAM network with different artificial languages, shortcomings of conventional artificial neural network training methodology are revealed. In light of that, the second part of the thesis contributes to the innovation of new artificial neural network training schemata, summarised in one term - hybrid neural-genetic, which incorporates evolutionary computing into artificial neural network learning. Two neural network training schemata will be proposed with the ambition to further realise the potential of the RAAM paradigm. The second proposed training schema introduces the notion of stage-wise evolution of network's initial weights and highlights author's original contribution to new idea and methodology. Along with them, the presentation of simulation results showing strengths of the proposed schemata will also be given.