Research Of Type Detection On English Event Extraction

Event is defined as an objective fact involving participants.Event extraction task in Automatic Content Extraction(ACE)involves the identification of certain instances of event types,along with the corresponding arguments in unstructured source text.Event extraction tasks can be further divided into event type classification and argument role classification,we focus on the previous one.Current event extraction methods are generally limited by data sparseness problems,which make it difficult to extract the undiscovered or uncommon constitutes of events.The research principally include the following three aspects:Optimizing Event Type detection with relevance propagationWe propose a method based on relevance propagation to optimize event extraction.This method aims at mining relevant event types of the test document by utilizing external documents,then propagates them to the “missed” samples of system,which can modify the results of the base extraction system to improve the recall rate of extraction.Moreover,in order to alleviate “over-propagation”,we also introduce restrictions based on “classification confidence scores” and “ theme distribution entropy” to further improve the accuracy under the premise of ensuring the recall rate.Experiments show that the relevance propagationbased method can assist event type detection in its performance improvement.Using text paraphrases to improve Event Type detectionMultiple events of the same type tend to be included in a document,while sentences that carry these same types of events tend to have different textual representations.The base extraction system often “misses” some of the results for the same type of event limited by the sparse and unbalanced nature of event extraction corpus.Therefore,we propose text paraphrases to identify sentences with the same semantics to improve the performance of the base system.In this paper,we propagate the type of “seed event” under the different selection strategies,which improves the final performance of event type detection.Combining Deep Learning and Active Learning for Event Type detectionTraditional supervised learning methods often suffer from small scale,imbalanced distribution and uneven quality of training corpus.In addition,traditional event extraction methods based on feature engineering are complicated and will always cause error propagation.To address these issues,we present a method to combine deep learning and active learning by the confidence of the query function based on Jordan-RNN's trigger classification,in order to improve the quality and efficiency of corpus annotation as well as the ultimate performance.