Matching Reviews to Database Objects Based on Labeled Latent Dirichlet Allocation Model

We develop a method for matching unstructured reviews to database objects in data integration, where each object has a set of attributes. To this end, we propose a Labeled Latent Dirichlet Allocation model. We model reviews as if they were generated by a two-stage stochastic process. Each review is represented by a probability distribution over attributes, and each attribute is represented as a probability distribution over words for that attribute. We introduce the label for each attribute, and then the model integrates object information. We use an unsupervised manner to estimate the model parameters, and use this model to find, given a review, the most likely object to be the topic of the review. Experiments in multiple domains show that our method is superior to the TFIDF method as well as a recent RLM method for the review matching problem.