From dolphin biosonar to document retrieval: from spectra to semantic profiles

The problems of recognizing object characteristics from dolphin biosonar signals and recognizing the meaning of words can both be characterized as pattern recognition problems. The configuration of features of the dolphin echo and the configuration of the context in which words are used provide the meaning by which these events can be interpreted. The paper describes a neural network that captures the meaning of words relative to the context in which they appear. Each text object (e.g., a document, paragraph, or word) is translated into a vector for input. A modified Hebbian learning algorithm is used to learn the relationships among these words that carry the contextualized meaning of the text. The resulting semantic profiles allow one to implement a fuzzy semantic comparison among text objects. They allow one to retrieve documents that correspond in a fuzzy sense to the query terms, even if the specific terms do not appear in the relevant documents. Rather than structuring the storage of the documents (e.g., by assigning them keywords), the technology structures the retrieval of the documents by allowing each user to formulate ad hoc categories reflecting his or her information needs.