Structured Named Entity Retrieval in Audio Broadcast News

This paper focuses on the role of structures in named entity retrieval inside audio transcription. We consider the transcription documents structures that guide the parsing process, and from which we deduce an optimal hierarchical structure of the space of concepts. Therefore, a concept (named entity) is represented by a node or any sub-path in this hierarchy. We show the interest of such structure in the recognition of the named entities using the conditional random fields (CRFs). The comparison of our approach to the hidden Markov model (HMM) method shows an important improvement of recognition using combining CRFs. We also show the impact of time axis in the prediction process.