Latent Topic Extraction from Relational Table for Record Matching

We propose a latent feature extraction method for record linkage. We first introduce a probabilistic model that generates records with their latent topics. The proposed generative model is designed to utilize the co-occurrence among the attributes of the record. Then, we derive a topic estimation algorithm using the Gibbs sampling technique. The estimated topics are used to identify records. The proposed algorithm works in an unsupervised way; i.e., we do not need to prepare labor-intensive training data. We evaluated the proposed model using bibliographic records and proved that the proposed method tended to perform better for records with more attributes by utilizing their co-occurrence.