Natural Language Acquisition: State Inferring and Thinking

摘要

Natural language understanding plays an important role in our daily life. It is very significant to study how to make the computer understand the human language and produce the corresponding action or response. Most of the prior language acquisition models adopt handcrafted internal representation, and they are not sufficiently brain-based and not sufficiently comprehensive to account for all branches in psychology and cognitive science. An emergent developmental network( DN) is used to learn, infer and think a knowledge base represented as a finite automaton, from sensory and motor experience grounded in this operational environments. This work is different in the sense that we emphasize on the mechanism that enable a system to develop its emergent representations from its operational experience. By emergent, we mean a pattern of responses of multiple elements that corresponds to an event outside the closed skull but each element (e.g. pixel, muscle, neuron) of the pattern typically does not have a meaning. In this work, internal unsupervised neurons of the DN are used to represent short contexts, and the competitions among internal neurons enable them to represent different short contexts. By internal, we mean that all the neurons inside a brain are not directly supervised by the external environment - outside the brain skull. In this work, we analyze how internal neurons represent temporal contexts and how the feature neurons of the DN represent earlier contexts. Accuracy of Z state inferring and X thinking of a relative complex training sequence( denoted as DN-2 in this work) can reach 100% and 75%, respectively. Comparative experiment results between this emergent method and the symbolic method, their corresponding Z state inferring and X thinking accuracy are 100% and 82.1%, 85.7% and 75%, respectively( taking DN-6 in this work as the example), demonstrate the efficiency of the DN on natural language inferring and thinking. Complexity of the finite automaton is low and so is the temporal contexts, but the same principle is potentially applicable to more complex cases.