Research On Relation Extraction Via Incorporating Knowledge Representation

With the technological innovation brought by the information revolution,the scale of information generated by human society is experiencing explosive growth and dissemination.However,in the face of the influx of massive information,the high-quality information in the Internet has been submerged in the information torrent,and it is increasingly difficult for people to exchange information effectively with the real world.Therefore,how to efficiently mine and retrieve target information from massive Internet resources has become a key issue that needs to be resolved urgently.As a key step in the information retrieval process,information extraction plays an important role and has attracted widespread attention from the current academic and industrial communities.As a key task in the field of information extraction,relation extraction has always been a hot research topic.Existing relation extraction methods are limited by their dependence on large-scale human annotated data,and are often difficult to be applied.In order to solve this problem,distantly supervised relation extraction came into being.Distant supervision has alleviated the problem of large-scale data annotation through automated method,thus it is also an important research topic in the field of relation extraction.Most current distantly supervised relation extraction methods focus on the denoise processing of noisy data,and ignore the mining and utilization of external knowledge information.Therefore,in this thesis the knowledge representation and fusion methods of different perspectives such as entities and texts are explored,and additional knowledge information is introduced to improve the model’s ability of understanding texts.Relation extraction models fused with knowledge representation are designed to achieve better relation extraction performance.Firstly,an entity knowledge aware relation extraction model(Entity Knowledge Enhanced Neural Network,EKNN)is proposed.Considering the importance of entity knowledge for linking entities and texts,a knowledge-aware word embeddings method is designed.The semantic knowledge and structure knowledge of entities are dynamically injected into the model,which improves the semantic understanding ability of the model.Then,based on the selective attention mechanism,the feature representation at the bag level is completed,which improves the extraction of relations between entity pairs.Finally,extensive experiments and result analysis are carried out on a widely used distantly supervised public dataset by comparing our models with some current mainstream distantly supervised relation extraction models.The EKNN model deeply explores and utilizes the rich information by fusing entity knowledge and word embedding,thus effectively improves the model’s ability of understanding text and achieves better results than the baseline methods.Secondly,we further propose a global contextual knowledge fused relation extraction model(Global Context with External Knowledge,GCEK).Aiming to address the problem of lacking textual semantic information in the existing distantly supervised methods,global contextual knowledge is constructed from the text-level to combine with entity knowledge to achieve the integration of different levels of knowledge.Besides,a global context enhanced selective attention mechanism is proposed to effectively improves the feature representation ability at the sentence bag level,and further improves the relation extraction ability.The experimental analysis on the public dataset shows that the GCEK model introduces global contextual knowledge based on entity knowledge,which can effectively carry out semantic modeling of different granular levels of text,so it enhances the noise reduction ability of the model and achieves the best performance.