Research On Event Detection Method Combining Feature Optimization And Negative Sample Sampling

Event detection aims at structured representation of event expression language extract-ed from free text.(e.g.triggers are extracted and labeled).In the process of event detection,the same word in different contexts may trigger different event types,while the methods based on neural networks only rely on a single form of word vector,which can not represent different context information.Besides,in the training process,fake features in the contexts will lead to bias in event detection.Meanwhile,it is found that the distribution of event types in the corpus is extremely unbalanced by observing the ACE2005 corpus,and the problem above causes that the triggers with fewer event types in the training set can not be assigned to real event types when identifying triggers in the test set.In order to solve the problems above,the following methods are proposed:(1)Using linguistic features and convolutional neural network to improve event detectionThis paper proposes an event detection method based on linguistic features and convo-lutional neural network model.Analogy word embedding,the feature information is trans-formed into real vector form which called "feature embedding" in this paper.The feature embedding is used as the input of the neural network.To some extent,the method alleviates the "semantic gap" caused by discrete representation of linguistic features in natural lan-guage processing tasks.In addition,the method can weaken the influence of fake features in the training process according to the objective function,thus reducing the negative effect of error propagation.Experimental results show that the method can achieve comparable performance to the state-of-the-art systems.(2)Using generative adversarial network to improve event detectionWe propose a self-regulated learning method by utilizing a generative adversarial net-work to generate spurious features.Due to the ability of encoding and mapping semantic information into a high-dimensional latent feature space,neural networks have been suc-cessfully used for detecting events to a certain extent.However,such a feature space can be easily contaminated by spurious features inherent in event detection.In this paper,we propose a self-regulated learning approach by utilizing a generative adversarial network to generate spurious features.On the basis,we employ a recurrent network to eliminate the fakes.Detailed experiments on the ACE 2005 and TAC-KBP 2015 corpora show that our proposed method is highly effective and adaptable.(3)Employing negative sample sampling to improve event detectionWe propose an approach to optimize event detection by negative sample sampling.Sentence-level event detection model regards sentences without triggers as negative samples,and negative samples play an important role in the model training process.Therefore,this paper proposes a negative sample sampling method based on reinforcement learning method to optimize the training of the model.The experimental results show that the proposed method can effectively improve the performance of event detection.By using the methods above,we can optimizes the use of not only feature information but also training set samples in the neural network-based models training to a certain extent.All in all,Our methods can improve the performance of event detection and the F1 score in trigger identification and classification can reach 77.0%and 74.8%separately.