Research On Incorporating Bidirection-interactive Information For Relational Facts Extraction

Relational facts extraction means to extract a pair of entities and their relation simultaneously from the sentence,which is an important task in natural language processing.Relational facts extraction is widely applied in many tasks,such as knowledge graph construction.However,relational facts extraction is very complicated,because the type of relational facts is uncertain.Overlapping relational triplets are very common in sentences.Relational facts extraction is a challenge task since overlapping relational triplets extraction is quite difficult.Relational facts extraction consists of two subtasks that are named entity recognition and relation classification.Pipelined methods which are traditional methods treat named entity recognition and relation classification as two separate subtasks.They first conduct entity recognition and then predict the relation between the extracted entity pairs.However,relation extraction is heavily affected by the errors caused by named entity recognition since relation extraction is conducted after named entity recognition.That is to say pipelined methods have the problem of error propagation.To solve the problems of error propagation,joint extraction methods which extract entities together with their relation using a joint model are proposed.The interaction between entity recognition and relation classification is essential for the task,because relation extraction can facilitate named entity recognition,in the meantime,named entity recognition is also helpful for relation extraction.Recently,most of works pay much attention to design a novel framework which is appropriate to extract overlapping relational facts,but they only consider unidirectional interaction between entity recognition and relation classification or even ignore the interaction between the two subtasks totally.That is to say the interaction between named entity recognition and relation classification is not fully modeled.The semantic of phrases between a pair of entities often reveal the relation of the entity pair,but most works ignore that.To address the problems above,we propose an end to end unified joint extraction model,which considers bidirection-interactive information between named entity recognition and relation classification.The model also incorporates semantic features of the phrases between a pair of entities.In sentence encoding module,we utilize Bi-LSTM and GCN to capture the sequence-semantic features and the structure-semantic features of a sentence.In interactive layer,we capture the bidirection-interactive information between named entity recognition and relation classification.Decoder layer generates all the relational triplets including overlapping relational triplets in the sentence.To improve the model's ability of relational facts extraction,we also design two optimization strategy.One strategy is to optimize the structure of the model,the other one is to utilize multi-task learning to optimize the model.We demonstrate the effectiveness of our method on two public datasets and achieve state-of-the-art results.We also conduct the experiments to verify the effectiveness of our proposed two optimization strategy.