Topic mining for call centers based on A-LDA and distributed computing

摘要

Early work on call centers focused on opinion collection, and service acceptance used manual entry tornrecord text information for incoming calls and to count various types of calls. With the development ofrntext mining technologies, it is possible for call centers to become information providers. Although traditionalrntext mining methods have been well studied, no algorithm has been designed specifically for callrncenter data, which contains structured information about the call center in addition to plain text. A telephonerncall usually has several properties such as caller number, call time, and so on. Aiming at the datarncharacteristics of call centers, in this paper, we propose an improved latent Dirichlet allocation (LDA)rnmodel called A-LDA, which includes information from external-related attributes and is suitable forrnbuilding topic mining models for call centers. We present a Gibbs sampling-distributed computingrnimplementation for our model’s inferences and use datasets from China Central Television telephonerncall center to perform an experiment on topic detection. The results show that A-LDA, which uses externalrncorrelation properties, has a lower perplexity value and a better generalization performance than therntraditional LDA method and can find topics associated with values contained by the external attributes.