Efficient music information retrieval (MIR) require to have meta information about music along with content based information in the knowledge base. Discussion forums on music are rich sources of information gathered from a wider audience. Taking into consideration the nature of text in these web resources, the yield of relation extraction is quite dependent on resolving the entity references in the document. Among the few music forums dealing with Indian classical music, rasikas.org (rasikas, 2015) having rich information about artistes, raga and other music concepts is taken for our study. The forum posts generally contain anaphoric references to the main topic of the thread or any other entity in the discourse. In this paper we focus on coreference resolution for short discourse noisy text like that of forum posts. Since grammatical roles capture relation between mentions in a discourse, those features extracted from dependency parsing are widely explored along with semantic compatibility feature. On investigation of issues, the need for integrating known dependencies between features emerged. A Bayesian network with predefined network structure is evaluated, since a Bayesian belief network enacts a probabilistic rule based system. To the extent possible the superior behaviour of Bayesian network over SVM is analysed.
展开▼
