Semantic knowledge representation and analysis

Abstract

Natural language is the means through which humans convey meaning to each other - each word or phrase is a label, or name, for an internal representation of a concept. This internal representation is built up from repeated exposure to particular examples, or instances, of a concept. The way in which we learn that a particular entity in our environment is a "bird" comes from seeing countless examples of different kinds of birds. and combining these experiences to form a menial representation of the concept. Consequently, each individual's understanding of a concept is slightly different, depending on their experiences. A person living in a place where the predominant types of birds are ostriches and emus will have a different representation birds than a person who predominantly sees penguins, even if the two people speak the same language. This thesis presents a semantic knowledge representation that incorporates this fuzziness and context-dependence of concepts. In particular, this thesis provides several algorithms for learning the meaning behind text by using a dataset of experiences to build up an internal representation of the underlying concepts. Furthermore, several methods are proposed for learning new concepts by discovering patterns in the dataset and using them to compile representations for unnamed ideas. Essentially, these methods learn new concepts without knowing the particular label - or word - used to refer to them. Words are not the only way in which experiences can be described - numbers can often communicate a situation more precisely than words. In fact, many qualitative concepts can be characterized using a set of numeric values. For instance, the qualitative concepts of "young" or "strong" can be characterized using a range of ages or strengths that are equally context-specific and fuzzy. A young adult corresponds to a different range of ages from a young child or a young puppy. By examining the sorts of numeric values that are associated with a particular word in a given context, a person can build up an understanding of the concept. This thesis presents algorithms that use a combination of qualitative and numeric data to learn the meanings of concepts. Ultimately, this thesis demonstrates that this combination of qualitative and quantitative data enables more accurate and precise learning of concepts.