Research On Joint Entity And Relation Extraction Method Based On Multi-Feature Fusion

With the rapid development of the Internet,the number of users has gradually shown explosive growth,the volume of data is exponentially expanding,how to deal with massive data and obtain valuable information from it has become a hot issue.Entity relation extraction came into being,which aims to extract entities and the relation between entity pairs from unstructured or semi-structured text,and transform them into structured information in the form of triple,so as to provide strong data support for downstream intelligent applications such as knowledge graph construction and question answering systems.Aiming at the problem that the current entity relation extraction model does not fully utilize the entity hierarchy features,which affects the model extraction effect,a multi-feature joint entity and relation extraction model MF＿Joint based on entity type is proposed.For relation extraction,the entity feature information in the text is particularly important,and the entity type feature can enhance the binding force of entity information on relation extraction and help the correct prediction of the relation between entities.MF＿Joint model evaluates the probability that the current head entity belongs to each type by feeding the head entity feature vector into the neural network,then sums the probability values to obtain the head entity type feature vector.The text feature vector 、the head entity feature vector and the head entity type feature vector are used as the input of the relation extraction sub-model,and finally output the complete triple.In this paper,the triple prediction of the MF＿Joint model is performed on two public datasets NYT and Web NLG.The experimental results show that the MF＿Joint model performs better than the baseline model,but the enhancement effect is limited in complex text scenes.In order to improve the performance of the model in complex text,this paper explores the feature fusion method of the model from the perspective of layer normalization and attention mechanism,finally a joint entity and relation extraction model MF＿Joint+CLN based on CLN and a joint entity and relation extraction model MF＿Joint+AFF based on attention mechanism are proposed.Firstly,the feature fusion based on CLN is proposed,regards the features to be fused as the weight and bias in layer normalization for feature fusion,adaptively adjusts the proportion of features during the model training process.Secondly,the feature fusion method AFF based on the attention mechanism is proposed,and the attention value is obtained by combining the global information and the local information to perform attention calculation on the fusion features.The above two methods make the model focus on the text information related to the current processing head entity during relation extraction,which is conducive to the correct extraction of triple.In this paper,the MF＿Joint+CLN model and MF＿Joint+AFF model are used for triple prediction on the NYT and Web NLG datasets.The experimental results show that the MF＿Joint+CLN model and MF＿Joint+AFF model can effectively improve the feature expression ability of the model and enhance the interaction between the entity recognition sub-model and the relation extraction sub-model in the model,which helps to further improve the performance of the model.Among them,the MF＿Joint+AFF model performs better,especially in dataset NYT with more complex data,the improvement effect is more obvious.