Research Of Semantic Relation Classification Based On Deep Learning

Semantic relation classification between entities is a significant task in the natural language processing research and it has a wide range of applications in information retrieval,summarization,machine translation,knowledge base construction,and semantic network.Moreover,semantic relation classification also plays an important auxiliary role in the research of word sense disambiguation,language modeling,text interpretation and text entailment.In recent years,with the development of deep learning research,deep learning based methods in NLP tasks has also received wide attention and application.In the field of semantic relation classification,the deep learning based methods have achieved better relation classification results than the traditional methods that based on features and statistics while saving the cost of manual processing.The existing deep learning semantic classification methods mostly rely on shallow neural network models,which do not take full advantage of the depth of neural networks to extract deeper semantic features.This paper designs an end-to-end deeper network model on semantic relation classification referring to the work of gated convolutional neural networks in language modeling and the work of deep residual network,which achieved great success in the field of computer vision.This model relies on the convolutional neural network and avoids gradient explosion and gradient vanishing problem based on the concept of gate linear unit and residual learning when extracting high-level abstract features from the shallow textual features layer by layer.The end-to-end gated convolutional neural network model proposed in this paper achieves 85.6%F1 score on the SemEval-2010 semantic relation classification dataset without integrating external linguistic features,and achieves state-of-the-art result among the previous end-to-end methods.The outperforming result proves the effectiveness of the theory proposed in this paper that using network depth improving the relation classification result.The dependency relation between words in the sentence,especially the shortest dependency path between entity nouns contains important information for judging the semantic relationship between entity nouns.The predicate-argument sequence on the shortest dependency path has a strong correlation on the judgment of semantic relations.Based on this concept,this paper integrates the dependency syntax features to explore the semantic relation classification task further based on the end-to-end model.In order to make the information expressed in the dependency relation more accurate,this paper also integrates the part-of-speech tagging feature,grammar relation feature,and WordNet hypernym feature to help extracting the dependency features.The dependency features was extracted by using two independent LSTM recurrent neural networks along the direction of dependency on the two subtrees of the shortest dependence path.The dependency features with entire sentences features extracted from the end-to-end model are integrated features.The hybrid model proposed in this paper can effectively use the integrated features obtained by two different modeling methods.The experimental results are improved to 87.0%comparing to the end-to-end model,and achieves state-of-the-art result among all of the previous methods.