Logic relation refinement using unlabeled data

The difference in distributions between datasets from different domains, such as different information sources, hinders the direct application of a learned model from one domain to another. We have developed a framework for the adaptation of relational logic models, in particular, Markov Logic Network (MLN), from a source domain to a target domain solving the same task using only unlabeled data in the target domain. In our proposed framework, we modify the model from two aspects, the dependency information across the two domains and within the target domain. First, the relational logic models of the two domains should share certain amount of similarities due to the same goal and similar nature of the data. Hence, we perform model adaptation by penalizing the difference in the two domains and jointly maximizing the likelihood of the target domain and minimizing the difference between the source and the target domain MLNs. Second, closely related information appeared within the target domain is used as additional clues in resolving ambiguous decision making. Potential ambiguity of the model is identified and is refined through incorporating such closely related information. As a result, the adapted model is tailored to the target domain. Our experimental results demonstrate that our adaptation framework is able to improve the performance on the target domain.