Topic mining for call centers based on LDA

Latent Dirichlet Allocation, which is a non-supervised learning method, can be used for topic detection, automatic text categorization, keyword extraction and so on. It only focuses on the text itself, not considering other external correlation properties. External association property refers to some structured attributes that correspondence with the text data, for example, a paper usually has several properties like authors, publishing time etc. A telephone call usually has several properties like caller number, call time etc. To iron out flaws; we propose an improved model A-LDA based LDA. We use data sets from telephone call centers (a kind of data centers in rapid growth) to experiment on topic detection. The topic results show that A-LDA with introduce of external correlation properties, compared with the traditional LDA, is decreased in perplexity value and has better generalization performance. At the same time, we can obtain the topic that external attributes contained.