Research On Event Extraction Based On Structured Learning

Event extraction refers to the extraction of valuable event information from unstructured text. With the increase of the data processing capacity, the automatic extraction of event information has very important significance.Traditional event extraction methods generally use the model of joint extraction architecture, and combine a lot of artificial design features to carry out event extraction; or use two-stage classifier cascade architecture,which features that still use artificial design Feature, or a parametric representation based on a word embedding. The contents and achievements of this paper are as follows:1. A new convolution neural network model, Skip-Window Convolutional Neural Networks, is proposed to solve the problem of generalization performance and data-sparsity which is usually existed when feature engineering is used to perform feature learning. Moreover,the traditional convolution neural network can only extract the sentence-level features, while the model can extract the word-level global structured features. First, the word vector of the specific word is concatenated with the word vector of all other words in the sentence together with its relative position information. Secondly, the result of the last step is regarded as the input of the convolution neural networks, after one round of convolution and pooling, we can get the global structured features of the specific word.2. A structural learning model based on recurrent neural networks is proposed, which can be structured using the parameterized features extracted from neural networks. The limitations of using one-hot feature in structured learning are solved. First, the output categories of each word are trained as label vectors. Secondly, the advantages of RNNs for variable-length sequence modeling are used to classify the whole label combinations of the sentence and use the method of beam search to find the final label combination of the input sentence.3. A new event extraction framework is designed, which can effectively combine the Skip-Window Convolutional Neural Networks and the Recurrent Neural Networks. The event extraction can be carried out jointly, and the event trigger and event arguments can be extracted simultaneously to solve the error propagation problem.In this paper, ACE2005 public corpus is used to experiment.Compared with several state-of-art event extraction models, the F1 score of event trigger recognition is 1.3% higher than the best reported event extraction system, and reaches comparable results on the other subtask.