Effective Retrieval of Audio Information from Annotated Text Using Ontologies

To improve the accuracy in terms of precision and recall of an audio information retrieval system we have created a domain-specific ontology (a collection of key concepts and their interrelationships), as well as a novel, pruning algorithm. Taking into account the shortcomings of keyword-based techniques, we have opted to employ a concept-based technique utilizing this ontology. The key problem in the retrieval of audio information is to achieve high precision and high recall. Typically, in traditional approaches, high recall is achieved at the expense of low precision, and vice versa. Through the use of a domain-specific ontology appropriate concepts can be identified during metadata generation (description of audio) or query generation, thus improving precision. In case of the association of irrelevant concepts to queries or documents there is a loss of precision. On the other hand, if relevant concepts are discarded, a loss of recall will ensue. Therefore, in conjunction with the use of a domain specific ontology we have proposed a novel, automatic pruning algorithm which prunes as many irrelevant concepts as possible during any case of query generation. By associating concepts in the ontology through techniques of correlation, this algorithm presents a method for the selection of concepts in the query generation. To improve recall, controlled and correct query expansion mechanism is proposed. This guarantees that precision will not be lost. Moreover, we present a way for the query generation in which domain-specific ontology can be used to generate information selection requests in terms of database queries in SQL. In trial implementations we have demonstrated that our ontology-based model outperforms keyword-based technique (vector space model) in terms of precision and recall.