Semantic Mining Dynamics for Games Language Processing
David Al-Dabass, Manling Ren · 2007
This paper attempts to determine conditions for `recogniseability' with application to games language processing. In its broadest sense, a biological reader of a string of characters has a `trial' internal model of the semantics of the lexical sequence being read. This internal model generates its own lexical string which is compared with the observed string. Errors between the two are fed back to the internal `semantic generator' to guide it to modify its lexical output closer to the observed string. The process continues dynamically until convergence, indicated by the observer `recognising' the meaning of the seen string. The theoretical foundations for this process are put forward and the conditions for successful `observation' using hybrid recurrent nets are reviewed. Semantic mining architectures are formulated and consist of a recurrent hybrid net hierarchy of multi-agents, extended such that the composite behavior of agents at any one level is determined by those of the level immediately above