Research On Related Technology Of End-to-end Neural Coreference Resolution

Coreference resolution is a key task in the field of natural language processing.Find-ing accurate and unambiguous coreference chains can promote the overall understanding of text semantics.Coreference resolution plays an very important role in supporting natural language applications such as information extraction,automatic summarization,question answering system and machine translation.Recently,with the rise and development of deep learning,more and more researcher-s begin to use deep learning for coreference resolution.Deep learning models have great advantages over traditional machine learning models because of its powerful fitting abil-ity.However,most of neural coreference resolution models only focus on the sequential information of text,and a large number of studies have shown that structural information is crucial to coreference resolution.Therefore,we carry out a series of research work on structural information.The research contents mainly include the following three aspects:(1)We reimplement the state-of-the-art end-to-end neural coreference resolution model and improve the time-consuming operations in the model.Compared with the origxnal model with the same experimental settings,the improved model has more advantages in training time and resource utilization.Based on this model,we propose the following two structural embedding methods.(2)We propose the compact constituency parse tree to add structural information.The original trees have many nodes and complex structures and exist a lot of redundant infor-mation unrelated to the coreference resolution task.To solve this problem,we propose a compression algorithm for constituency parse trees.This algorithm preserves the structural information of key nodes,reducing the redundant nodes and compressing the structure of the constituency parse trees dramatically.Then two key features,depth and siblings,are extracted from the compact constituency parse tree as the representation of structural infor-mation.The experimental results on the CoNLL 2012 Shared Task corpus show that the structural embedding based on compact constituency parse tree improves the performance of the baseline significantly more than the original tree.(3)We propose the node representation approach to add structural information.This approach models the structure by encoding the node sequence representation.And we can use feature supplement and node update mechanism to further improve the sequence rep-resentation.This approach avoids the problem that existing models do not directly support batch computation.In addition,we also use node enumeration to replace the original span extraction strategy.All kinds of experiments on the CoNLL 2012 Shared Task corpus show the effectiveness of our proposed approach.