Research On Entity Coreference Resolution Based On Deep Learning

With the continuous development of information technology,various industries will generate massive amount of text and data.At the same time,it is increasingly important and challenging to use natural language processing related techniques to mine the connections between data and the underlying semantic relationships.The task of coreference resolution is an important research area in the field of natural language processing,which is widely used in automatic question answer,text summarization,reading comprehension,knowledge graph,and so on.That has important academic research value and commercial use prospects.At present,co-reference resolution research in the framework of deep learning is becoming a mainstream research approach,focusing on how to effectively identify coreference relationships between entities and entities using word embedding information based on sentences.Solving the coreference resolution problem has the limitation of considering only word level information,also needs to consider the information of the context information in the document and the logical relationships between sentences.Therefore,in this paper,in order to improve the accuracy of the coreference resolution model,the coreference resolution is studied from the following two aspects.(1)Entity coreference resolution with fusion neural networks and global inference.In order to explore a more effective method for coreference resolution,we propose an entity coreference resolution algorithm by fusing neural network and global inference to address the problems of complex entity information and unclear referent information in text and poor consideration of global features in the context of the document.Firstly,the neural network model is used to extract the entities and their antecedents in the document,and secondly,the global inference is combined with the contextual information of the sentence,and the results of this inference are added to the neural network model to improve the accuracy of entity coreference resolution.The global inference is a document-level global optimization and inference of the coreference chain,which can be combined with the intrinsic connection between sentences,while mining the contextual semantic information of the entities.The experimental results of entity coreference resolution performed by the model on the Onto Notes5.0 dataset demonstrate the effectiveness of the method.Subsequent experiments are conducted by replacing the joint word vector in the end-to-end coreference resolution model using a Bertbase pre-trained model,which again validates the effectiveness of the method on the coreference resolution task.The method is effective in improving the coreference disambiguation performance and better understanding the semantic information of the text,and the final model performance achieves an F1 value of 74.76% under the CONLL evaluation standard.Comparing the experimental results of this model with other coreference resolution models in recent years,the effectiveness of this method is verified.(2)A joint model of named entity recognition and coreference resolution incorporating knowledge base information.Both named entity recognition and coreference resolution depend on the learning of adjacent textual information of entities,and both currently achieve state-ofthe-art results in terms of individual independent tasks,and in fact they maintain a high degree of connection.To improve the accuracy of entity coreference resolution models,we consider a joint knowledge-base based model for named entity recognition and coreference resolution.The joint model is based on end-to-end network architecture with a global inference-optimized coreference resolution algorithm model,using wiki data as an external knowledge base and fusing entity representations from the background knowledge base to combine the named entity recognition task with the coreference resolution task,and experiments demonstrate that the method effectively improves the task accuracy in the joint model.And setting up ablation experiments to explore the effects of global reasoning and knowledge base modules and attention mechanism on the model performance.The F1 value is used as an evaluation index,and the experimental results show that the method in this paper has a better coreference resolution result.