Research On Entity Resolution Based On Markov Logic Networks And Implementation Of System

AI problems in real world always face two kinds of challenges. One is dealing with complexity, the other is handling uncertainty. How to solve these two kinds of problem simultaneous is always one of the most important attention in AI area. The proposal of Markov Logic Network meets these two aspects. Markov Logic Networks combines First Order Logic and Probability Graphic Model in order to obtain model of likelihood in relational databases. Markov Logic Networks has great value of research and application future which is recognized by international AI areas, and becomes a hotspot in Machine Learning and Data Mining area.This paper describes the entire theory of Markov Logic Networks, and introduces entity resolution based on Markov Logic Networks and its improvement. Finally, we demonstrate realization of system.Main work in this paper including:(1). Described the theory of Markov Logic Networks in detail, including definition of Markov Logic Networks, inference algorithms, weight learning algorithms, structure learning algorithm.(2) Demonstrates theory of entity resolution based on Markov Logic Networks. The solution imported an equivalence predicate in order to remove precondition of unique of names, and then based on equivalence predicate define the problem formation. And the following two section in this chapter, we introduced standard model of Fellegi-Sunter, and entity resolution system based on Markov Logic Networks.(3) Proposed the improved entity resolution algorithm based on Markov Logic Networks which is validated by experiments. First, we analysed the flaw of the original algorithm, and then we introduced a new rule into the original Markov Logic system. However, because the new rule is conflicted with the existing rules, and the discriminative training does not considered this confliction, we had to adjust learned weight of new rules. In this chapter, we introduced an extra coefficient k for the weight of new rule. When the new rule is representing one-to-one relationship, k is defined as 1, otherwise, k∈(0,1). Experiments shows that when k belongs to (0,1), the accuracy can be improved when k is less than 0.7. Experiments also shows that when k∈(0.3, 0.45), the accuracy truns up trumps.