Research On Distantly Supervised Relation Extraction With Relation Paths

As the core technology of current artificial intelligence research and information intelligent services,knowledge graph often has the problem of incomplete information.Relation extraction based on deep learning extracts the relation of the target entity pair from the unlabeled text,which is a key method for the completion of the knowledge graph,but the labeled data required for training is usually difficult to obtain.Based on the distant supervision hypothesis,a large amount of labeled data can be automatically obtained by using the existing knowledge graph.However,distanly supervised relation extraction also has the problem of wrong label problem due to the defects of the distant supervision hypothesis itself.The current mainstream distantly supervised relation extraction model often adopts the multi-instance learning(MIL)method,which means dividing all sentences into bags according to the different entity pairs,regarding the relation of each entity pair as the label of each bag,and then starting the subsequent noise reduction and relation classification process.However,this method ignores the potential information dependence among bags,resulting in a poor performance when all sentences in the bag are incorrectly labeled,which in turn affects the robustness of the model and the overall relation extraction performance.To address this issue,this paper uses the potential relation paths among bags to model bag-level information dependence,and proposes the following two models according to the complexity of the relation paths:(1)A distantly supervised relation extraction model with two-hop relation paths.The model first adopts the multi-instance learning method to model the direct sentence information in each bag to obtain the probability distribution of the relation label,and models each two-hop relation path information that exists between the entity pairs of each bag,then only selects one relation path with the highest confidence to represent the inference information of the relation path module,obtains the probability distribution of the relation label through the relation path encoding module,and finally adopts the joint learning framework to synthesize the direct sentence information and the reasoning information of corresponding relation path module,then extracts the target relation of each bag.(2)A distantly supervised relation extraction model with multi-hop relation paths.Based on the two-hop path-based model,this model mainly makes the following two improvements for the relation path encoding module.One is to expand the scope of relation path mining,and introduce three-hop and four-hop relation paths between each entity pair,thereby modeling more complex label dependency information among bags,To enhance the performance of the relation path encoding module.The second is to improve the way of information fusion of multiple relation paths,use the attention mechanism to consider the contribution of all relation paths,and consider more positive relation paths,thereby improving the performance of relation path representation.In this paper,multiple sets of comparative experiments are performed on the Wikidata+NYT dataset to verify the effectiveness and necessity of introducing relation path information.At the same time,an ablation experiment is designed to verify the effectiveness of the attention mechanism for fusing relation path information.