A new entity relation tuples filtration method for weakly supervised relation extraction

The phenomenon of invalid entity relation tuples generating more invalid tuples in the next cycle is called as “circular dependencies” in relation extraction. In order to prevent the existence of the phenomenon of “circular dependencies”, we present a entity relation tuples filtration method that filters the invalid entity relation tuples for Weakly Supervised Relation Extraction. In our article, tuples are filtered with Expectation-Maximization algorithm according the confidence of templates based on the assumption that the more seed tuples generated are, the more reliable the templates generating the seed tuples are. If the confidence of the tuples below a certain threshold, it will be discarded as an invalid relation tuple. The experiment shows that the accuracy have improve to varying degrees compared with some methods for the relation of couple, the relation of the hometown and the relation of minister of finance. The improvement is up to about 1 percent, 13 percents, 20 percents for the relation of minister of finance, the relation of the hometown and the relation of couple. It implies that our method can filter the invalid entity relation tuples effectively.