Research On Neural Network Method Of Event Coreference

Event coreference resolution is used to link the instances of the same real-world event,which is one of the main sub-tasks of information extraction.Because of the flexible expres-sion of event and the complex relationship between events,the task of event coreference resolution is more challenging than that of entity coreference resolution.In the past,a lot of manual extracted features have been used in event coreference resolution,which is not only time-consuming and labor-consuming,but also poor portability.With the wide application of neural network in the field of natural language processing,this dissertation focuses on neural network method of document-level event coreference resolution.The main contents of this disseration include the following three aspects:(1)To solve the problem of single event representation in the existing event coreference resolution methods,a neural network model based on attention mechanism for English event coreference resolution is proposed.Firstly,we use a variety of linguistic knowledge to enrich the event representation.Secondly,we use Bi-directional Long Short-term Memory Network,Convolution Network and Attention Mechanism to extract the event global and local features,respectively,and then filter out redundant information.On this basis,the similarity model is used to generate the coreferencial event chain.Finally,the global optimization is used to further optimize the coreferencial event chain.Experiments on the KBP and ACE English Corpus show that the proposed method outperforms the baselines(2)To solve the problem of polysemy and multiple references in Chinese,a neural net-work model for Chinese event coreference resolution based on multi-similarities is proposed.Firstly,while the recurrent neural network is used to encode various linguistic knowledge,the gated attention mechanism is introduced to control the information flow and screen out the relatively important information in the event.Then,three similarities and the matching features are introduced to distinguish whether the event pairs are coreferencial.Experiments on ACE Chinese corpus show that the performance of this method is better than the baselines.(3)Aiming at the problem of insufficient learning of neural network caused by the small scale of the Chinese and English corpus,a neural network method of event coreference res-olution based on Chinese and English cross-lingual learning is proposed.First of all,the machine translation tool is used to translate the source language corpus into the target lan-guage partially aligned corpus.And the single events are extracted according to the depend-ency words of trigger words and their arguments.Secondly,cross-lingual learning is carried out by sharing parameters on the basis of English corpus.Finally,the linear similarity and non-linear similarity between the two single events are calculated,and the similarity neural network model is used to judge whether the event pairs are coreferencial.Experiments on the ACE English corpus show that this method is superior to the baselines in many aspects.Aiming at the problems existing in the event coreference resolution,this disseration proposes three effective event coreference neural network methods.These methods improve the performance of event coreference resolution and contribute to its further development.