Inferring Definite-Clause Grammars to Express Multivariate Time Series

In application domains such as medicine, where a large amount of data is gathered, a medical diagnosis and a better understanding of the underlying generating process is an aim. Recordings of temporal data often afford an interpretation of the underlying pattens. This means that for diagnosis purposes a symbolic, i.e. understandable and interpretable representation of the results for physicians, is needed. This paper proposes the use of definitive-clause grammars for the induction of temporal expressions, thereby providing a more powerful framework than context-free grammars. An implementation in Prolog of these grammars is then straightforward. The main idea lies in introducing several abstraction levels, and in using unsupervised neural networks for the pattern discovery process. The results at each level are then used to induce temporal grammatical rules. The approach uses an adaptation of temporal ontological primitives often used in Al-systems.