Relation Classification Combined With Knowledge Graph

With the data of Internet expands sharply,how to extract and use information from these data becomes a hot research field.Relation Classification is a crucial subtask of Information Extraction.Its purpose is to extract relations between entity pairs from free text in order to turn the non-structure data into semi-structured or structured data which could be used in knowledge base construction and completion,question answering and other related field.Traditional relation classification task requires large labeled data for training which is labor intensive,to alleviate this problem,some researchers take knowledge base as a background knowledge for data labeling,if two entities have a relation in KB,then all sentences that contain these two entities will express this relation.In this way,there are sufficient data for training,but wrong labeling problem of this method are disturbing for sentences with same entities does not necessarily express this relation in KB.To alleviate problems above,we propose an attention based model to reduce the influence of wrong labeling data.To further improve the model performance,we propose a position-enhanced embedding model to extract sentence feature more effectively.The contribution of this paper can be summarized as follows:1.To alleviate the wrong labeling problem,we proposed a series of attention based models,including position-based attention model,sum-of-attention model and multi-attention model.Experimental results shows that our proposed model can achieve better result by recognizing important information of the input text.2.By analyzing the data of relation classification task,we proposed a position-enhanced embedding model.In this model,we split the sentence representation into three parts based on the entity pairs in the sentence,and use three independent convolutional networks to learn features,at last,we concatenated the output from different branches and employ a softmax layer to compute the probability for each relation.Furthermore,we combine the multi-instance learning with this position-enhanced embedding model and apply it into relation classification integrated with KB task.Experimental results on wildly used datasets achieve considerable improvements which shows that our proposed model can make full use of sentence information.